UPDATE: We managed to get Blender to run on the CMP card after posting this video. The results are not very good due to limited memory and PCIe bandwidth. Here are render time results for the BMW blender demo. Lower is better. CMP 170HX - CUDA: 36.82s - Optix: 39.33s RTX 3090 TUF - CUDA: 18.40s - Optix: 10.56s
Well reason you get crap performance is a VM can only see 4 of the GPU clusters per VM. To expose more to a VM you must buy NVidia VGPU licenses. Run into it often with GRID and ESRI.
Don't lie to us claiming it's "not worth it" to take it apart. Spending $4000 on a card is nothing to you. You'll make it back and then some with this video alone.
@@Psyopcyclops Jake used to be pretty bad in front of the camera back when Linus was constantly getting irritated with him. Looks like Linus mellowed out with him though.
from an engineering standpoint. the black fins on the ends are probably "Flow straighteners" they reduce the mixing of air in the channel to decrease differential pressure through the cooling section. allowing for a larger heat transfer from the fins to the cooling medium.
I read everything you wrote Nathan, and I realized I agreed for one reason: I didn't know what the actual fuck you were talking about, but it sounded so good, I had to believe it. 💪
Hey indy, feel lucky to say nvidia made this prod cause chips have some bad things if used other things? That crypto-thing eats crazy amount energy, but hey old bank system eat more ;)
Linus is putting so much care into this card specifically. If I remember correctly, he shook a $4000 Titan out of its anti-static bag like it was a $5 Christmas gift.
Ok, ask yourself this, why would they buy a card that takes longer to pay itself off? It literally doesn't improve anything for anybody, not even miners as there's no resale value there.
Just to let you know, when you're doing your very first render on the Eeves (or Cycles i don't remember) engine, the engine has to compile itself before actually doing rendering. After waiting like 10-15 minutes you will be able to do rendering after rendering. You can see at 19:22 on the line where you can see your frame count, the message "Loading render kernel (may take a few minutes for the first time" indicating you that you will have to wait a bit. Hope you will test this out later or read this comment 😀
Could be, but Blender not responding is a hint it probably crashed and would never render. In my experience the BMW test also does not take more than 1-2mins of compilation. Edit: Maybe im wrong, my experience is without OptiX and only with OpenCL / older CUDA
@@somestarwarsnerd9068 ..Blender not responding is as much a hint as it simply not responding for several other reasons outside the specific GPU... How many of us* have had to revert to an auto-save after a lock up or crash & it wasn't due to hardware issues to start with? Lol.. No matter how 'simple' or complex a given scene is, there's always the unexpected. Furthermore, if due to instability- you do the same thing again and again and prove it IS the card; or something related to it (to help give your comment a sense of merit). But that wasn't the case now, was it?.. /shrug Edit: "*" typo
The OptiX render kernel isnt precompiled in blender and always takes time in new scenes. CUDA on the other hand comes precompiled with blender so there isnt any kernel compilation happening anymore. This is why alot of times CUDA will be faster on big single frame scenes that uses nearly all Cycles features, its already done when OptiX is still compiling. When it comes to rendering out big animations OptiX will get the throne. But yeah this is why in my opinion these OptiX vs CUDA tests in blender are crap because they dont show the complete picture
@@somestarwarsnerd9068 No not at all. Blender is mostly single threaded and is poorly coded in python. Everything runs on one thread including the user interface. Most programs use at least 2. One should be reserved solely for the user interface. That way when you run tasks on the other thread it doesn’t freeze the interface. Even when blender uses more than one thread it still likes to run stuff on the user interface thread. So when doing any operation it freezes. This is because when operating on a single thread code most run in order. So The user interface cannot update until the other code runs. Windows considers programs with unresponsive interfaces to be unresponsive even if it’s just because they are running long code in the UI thread. When modeling in blender applying modifiers or moving objects or editing them in anyway way really can cause blender to become unresponsive but it almost always finishes whatever code it is running and comes back. The exception to this would be if it runs into an error and gets stuck in an infinite loop. Then it stays frozen forever and there is no way to cancel it. This is a common issue with modifiers because they use algorithms which do not always work on the object but blender isn’t smart enough to figure that out so it just keeps retrying. It has a lot of work left to do and certainly is not an ideal program to be using for these kinds of test. Blender is by far the best 3D modeling software and for the best software for physics, animation, and many other things too. It is free to yet still significantly better than stuff that coats you to pay several thousand dollars a year. It shameful for it to cost that much in the first place but subscription based model on top of that really is bullshit. People should own the stuff they buy especially when it’s that expensive. It doesn’t even cost them anything it’s 100 percent profit. They spend a small amount developing the software after the first few customers all the expenses are paid for but the price never drops. And the super cheap support that you may need to use once every few years for a few seconds doesn’t justify it either. The initial cost can easily cover that. Blender while open source still has a foundation with paid employees which has much more expensive and educated employees and much more of them. The research and programming that they put into it goes above and beyond anything any of the companies do yet they do it with a very small fraction of the money that they only get from donations. Things like autodesk and maya may have 50 developers at the most. Autodesk doesn’t even make most of its own software. They buy successful software. Then they make it free to bankrupt any competition out of business. Once the competition is gone they take away the free version and sell it at an extreme cost or at least limit the free version to the point where it is unusable. Almost none of their software is original and not purchased from another company. These companies also have a very small market share. If you take the publicly available revenue for these companies and only look at the small portion of it that they get from their software than divide that by the insane cost you find out that the number of copies they are selling is actually incredibly small. They just want you to think everyone is using it. The only way they stay a float is by making partnerships with universities and collages to teach people to use their software for free so that once they go out into the work force that is what they use. They also makes exclusivity deals with big companies and get them using their software. And once a company is already using a software and the people coming out of school are too it means it would take a huge up front investment of time and money to switch to blender or another alternative even if they would more than make the money back in the long term they don’t want to make the risk and may not have the cash flow to do so. These companies and their software really are unethical predatory crap that isn’t even as good as free alternatives that they didn’t even make or contribute much to. They just steal the ideas of others. Blender however is only better when it comes to capabilities and features. When it comes to performance with large scenes it is still behind. Plus because it has so many features they are often buried in menus and with complicated keyboard shortcuts and they like to change where everything is each update making it a real pain. You can do a lot more with blender but expect it to be super slow and unresponsive while doing it. It doesn’t even matter if you have a super computer because it can easily saturate that too. The only real fix is to recode blender from the ground up because it’s issues are so deeply rooted. They have almost completely redone it a few times and have made big improvements but it’s got a ways to go. They plan on fixing a lot though in the next major update then maybe it will finally make its way into the big companies and snuff out the other crap.
Glad we aren't wasting aluminum (aluminum alloy) on this. Fun fact we are facing a magnesium shortage which will affect aluminum alloy production... Which gets used in Many products including cars. So yes...
The outer fins kinda seem like they guide the air a bit more since running copper fins that far would be expensive. Also they cool the card body and having all those cards nearby in slots would probably be pretty good at keeping them all cooler to the touch.
Watching them trying to understand this project really makes me think that it would be a really good idea to invite current/former industry engineers onto the show to explain things to them and the audience from the perspective of an actual expert.
A part of the entertainment is watching enthusiasts working out what this stuff is. If I wanted to watch a video of someone explaining a product, I would just subscribe to companies
I work with A100's on trays of 8 in servers for a major cloud provider. I can attest that we've had nothing but problems with the 40GB A100 trays and individual GPUs causing a seemingly never ending RMA process. Nvidia has been having big QC issues with their production.
This card was clearly made with low volumes in mind. I mean a bunch of the stuff you guys pointed out and were confused by is likely influenced by the relatively low volume. 6:12 Believe it or not, that's a very common ways to connect a wire to a board at low volumes. Or high volumes with some removable custom fixtures. A lot of LED boards in the small red & yellow lights on semi tractor-trailers are constructed this way so you know it's reliable if done right. In this case, it also allows the connector to be screwed to the enclosure independent of the card, which helps reduce strains on the solder joint due to vibration, etc. 7:12 Milling away 90% of the material is fine - you just dump the scrap in a vat to burn a away the cutting fluid and melt the slurry, clean it up a bit, and cast another block. Great recovery rates. As for why, it's a low investment way to get a box with 3 sides. The cost per piece is going to be on the order of $10, and you don't need to spend a few hundred thousand dollars on a forming die, to then still spend about $10 per piece on top of that because of your low volumes. On a related note, 5:46 I wonder if those heat sinks are for cooling the enclosure. Aluminum has got to pick up lots of heat from adjacent cards and this helps carry some of that away.
Regarding the soldered connector: soldering those couple of wires is likely much cheaper than using a larger PCB just to bring the connector out further. It is also less likely to fail than a directly soldered connector and it does not bend the PCB. All around, not a bad design.
I know next to nothing about manufacturing or tech and was thinking many of the same things. Why incur "conventional" costs for an unconventional product...
The heat sink is clearly designed to have some thermal mass to stabilize long term temperature fluctuations and cooling requirements when running these at scale at 100% uptime.
it was probably cheaper to use the aluminum back plate from another card rather than create and order a new design of backplate for that card, they probably had to order them in bulk from another company for the other card so they used the access plates or defective units for this card since visuals don't matter. might be a strange area where being cost effective looks nice, feels great.
It's simply a matter of binning. Whoever said binning is not a thing anymore is an idiot. These card's silicone wouldn't be suitable for a gaming environment, but for more focused one, like crypto mining, they are more than enough. It's a great way to use everything and also finally stop the gaming cards shortage.
I know this is an old comment but I always wondered this. Even if they just did a 5 hour validation test on each card they manufacture just to make sure it will preform to it specs... that is a lot of hashpower... lol
I believe the milled back plate would indicate their expected volume for this product. If they were to cast it and then mill it that would be expensive molding costs. On low volume milling is sometimes more cost effective even if you’re milling 90% away. Just hog it out with a big tool to rough in and then finish it.
I used to work at an hp assembly factory putting together servers, testing them and boxing them up to ship off to whatever country. There was one line boxing up these cards, or an earlier version of them about 6 years ago. Same style minimal casing, no output ports and I could never find them talked about on an Nvidia site but could find other people selling them for 4k+
Yes, that looks identical to an A100 PCIE card, and the IHS is on the PCIE version. The A100 does have 3 NVLINK spots as well. (Ps) on the $12k A100 the power connection is soldered just like the mining card is.
A100 has floating computations + INT computations processors Maybe this mining card has only INT computations and zero floating point computations processors
@@Haskellerz I doubt it. I do some work on deep learning. It would be interesting to test. There’s some simple ways using Tensorflow to test for speed up. It’s down to if the tensor cores are still active.
@@jfolz not gonna argue with you there, there are some CUDA tricks around that. If the card sold for a lot less it might be worth it. Especially with the PCIE 1 link, it would be really slow. It would be fun to do a deep dive and see what functions are still left alive on the chip.
those black fins at the ends might be intended as baffles to direct air into the enclosure, as you will have some amount of turbulence with narrow ducts like that, and turbulence can prevent the air from pushing through it
19:21 thats a blender thing. it does it every once in a while at the start of a render, just let it do its thing. sometimes it takes 20 seconds other times it takes upwards of 5 minutes
Worth pointing out that the A100 doesn't have NVENC, only NVDEC. It also lacks most of a graphics pipeline, so not really usable for much except pure CUDA compute
The real reason for the slow PCI is that this would go into a big, cheap chassis with a huge PCI fanout - think 8 cards in parallel using PCI 1.1 risers so that it can fit into a 4u server case. The rest of the design seems to be optimized for energy and cost efficiency.
I worked for one software company that had AMD make a custom 'graphics' card. It was super duper odd and impressive since it did a damn good job. Really made me think how much faster I could build and test my stuff if the entire computer was designed for it.
yes, that is the whole point why RISC architectures like ARM exist in the first place. the problem is that you lose all other functionality. sometimes it is better to have 100% at one thing, other times it is better to have 60% at all the things. it's a trade-off
The efficiency benefit is probably related to memory type as well. HBM is more efficient than GDDR. If this card has the same tensor core capabilities as the A100, it will be a pretty good deep learning card, though the 8GB memory is too small for that.
Yeah 8GB. Makes no sense. You cant train anything for that except for very small networks. I mean this is why you can get the clusters of A100. I personally had problems with the AWS EC2-P4d instances on big models. And they have up to 8 A100.
Its possible the black fins are to ensure that the air passes in a straight line through the card rather than whirling around. They probably found extending the copper didnt help with temps as thermal transfer isnt affective that far from the die (or ihs) in this case
After stripping it out of some of the parts that makes it an unit for processing graphics, should we still call it a GPU? I mean, I guess if we strip a car out of its wheels and hang it up in a fixed place for using its engine for powering something else... It's still an automobile... which is not being used for automatic mobilization.
i can see why they might have done what they did with the chassis around the card. the big thing is that its serving as a shroud to channel air both through the heat-sink and then passing it all the way to the end of the card. the dummy heat-sinks on either end, those server fans can move a lot of air and those are going to act as grills to help protect against things being pushed/pulled into the card by the air flow regardless of which direction the server is using. i can also see the aluminum box being useful in those server racks with protection against things pushed against the card. i think the high end milled aluminum surface is a bit out there, but if it remains a long term thing they can probably either l simplify it or switch to a casting. but if they use it for other designs or it remains in low rate production, its probably as cheap as any other option.
That power connector extension has been normal on Quadro for years now. Since all Quadro/Tesla boards are standard length and have rear-facing power connectors, they make a cable extension whenever the PCB is shorter than the shroud.
Well, there is one problem though. The Manufacturing for the CMP Card has to be splitted even if it is not on the same process. This uses TSMC 7nm, like the Ryzen 5000 Series Processors or AMDs Graphics Cards for that same matter. So this new Card may not cut on NVidias own productions, yet. But it may hurt other ones productions.
Actually it sounds like this card runs from the discarded datacenter production line of their products So the main purpose of that line is to produce the card that end up in the true compute datacenters and these cards are the ones that fail QC for that branch and they can't create a tier for. So instead they squeeze out money from the miners that are willing to drop 5k on it.
@Bat Barfalamule huh, I don't have too many issues running my desktop PC with a 5700XT and tesla K80 at the same time. Maybe i'm just lucky. (Dual booting, using *nix more often)
@Bat Barfalamule Stop talking like that you scaring the kids lol. Btw Nvidia users like me experience the same pain you saw on this video but with REGULAR GPUs on Linux! 😂 That's what the OP of this comment just meant. Nvidia is the REAL PAIN for any other platform than Windows.
well im imagining how a product that came to the market sold out instantly.. thats like dropping something from your hand towards the floor but it never hits the floor, it just disappeared :p
@@emanu1674 yeah dont expect kids to understand its a mining card so they stop complaining about shortages and over pricing from us terrible miners hogging
19:22 Blender especially with OPTIX takes time to build render kernels during the initial render (can be 10 minutes depending on scene complexity). A little more patience might have yielded a different result especially considering that blender depends on CUDA compute.
10:40 Being in the process of scaling up for manufacture, it's development vs. volume costs. If it'll take your team a week to adjust things so that the connector is in the right place, and another 2 for verification and testing, and you're only making a few thousand of the things, it's cheaper to just pay the factory workers to fiddle with wires, rather than the engineers to develop a solution that doesn't need them. I'm starting to see how that function evaluates now, being the person who both develops and manufactures the parts, which things to spend time developing versus simply producing as they are is one of the harder decisions.
I recently went through the process of getting a electromagnetics simulation server for my job and we got an A100 PCIe 40GB. Really had to hold back and not take it apart. It absolutely slams in CST Studio Suite 2021 though. No complaints about the power and sim time reduction. Really such a bizarre card and such a specialized product that it's completely underwhelming for most people
Mining cards are great for mining, and thats it. Its a professional mining card, as as Linus mentions even for most miners it wouldn't be worth it because you cannot repurpose a card so it leads to landfill waste when they are no longer making sufficient profit. I get that they don't want the card to be repurposed but its not a good practice ethically or logically.
@@SlawcioD I understand it just fine. Billions upon billions of dollars in wasted equipment, colossal amounts of wasted power and needless emissions, all to make a small group of people who have the millions of dollars to set up a mining farm, obscenely rich at the expense of everyone else.
Currently in a mechanical engineering major at a top university in canada, I think the fins linus was confused about was intended to oriente the air into a lamar flow which would increase air flow and air speed through the actual heatsink mounted to the chip, with all the air moving in the same direction (as the other air and the fins) less energy is lost between particle interactions compared to if the air had to oriente in the heatsink. This assumes that the fans output is turbulent which I dont know for sure.
I have actually heard that closely packed GPUs can cause mining to slow due to electromagnetic interference. I wonder if that was one of the reasons to encase the whole card in metal?
@@tradingnichols2255 it would be fairly expensive if you don’t already have the cards but it should be just a matter of getting two of those rigid connecting bus card things filling them with mining cards and putting one upside down over the others or just getting the flexible PCIe cables and sticking the cards as close to tether as you can. And then fiddling with the shielding on the cards and the card positions.
holy shit the blender part drove me insane, "LOADING RENDER KERNEL, THIS CAN TAKE A FEW MINUTES" You just had to wait two minutes and maybe READ the text on the screen
What you guys should do is a watt for watt comparison of 30 series vs 40 series cards and include some odd ones like workstation cards. Cards like the rtx a5000 and that a100, if you still have it. Target certain break points, like no aux power, 1 6-pin, 1 8-pin, 2 8-pin, etc.
Not sure where to ask this? Would LOVE LOVE to see a complete how to vid 😉 for properly setting gaming settings? Especially for mid to low range computer set ups??? This would be soooo helpful as issues frequently comes up.
As an aeronautical engineer i kinda cringed when i saw that 3d printed part. It will perform wat better if you decrease the angle of the slope and give it more space after passing it through the fan. If your any good at cad you could make it elliptical but at these airspeeds thats just for minor gains
I am waiting for the result of that meeting linus got set up with UA-cam management. I believe it is going to be a 15 minute "$hit management says" episode XD
I imagine Nvidia got a REALLY good deal on those copper heats, then realized it didn't match length of the enclosure of the card which were already in production. So got those "fake" fins to fill out the card more and give it a professional look
Another comment I read said they were probably flow straighteners which help even out the distribution of air across the internal channels for better cooling efficiency.
@@1000_Gibibit innit I legit was going to build my pc last year for around 1.8k AUD with a 2070 super OC, now I have to build it for 2.6k this year with a 2060 OC, I hate miners
You're correct that is not gaming silicon, it would be compute anyway, but that is only so because nvidia choose for it to work that way! It's one thing to save on the output (they could have put only one like some old firepros but whatever) but to completely lock it down when people have been asking for years for server features like sriov just shows once again that they simply do not care
But does the output matter when the way it’s configured already makes it a horrible gpu for gaming. (I’m rather new to the hobby so correct me if I’m wrong)
I'd say those black aluminium fins are to clear out air vortexes before the actual copper cooler to get better air flow and there for efficiency thru the chassis. They look good, but could also have a function too.
The problem with this efficiency is it's not extreme enough. Two 3080's will mine $13-15/day and use $1.5 in energy. This card will use $0.6 and mine $11-12.5. Thus the two 3080's will net hundreds more profit but cost hundreds less. In less words, efficiency only matters when it limits hash due to heat, you've maxed out your electrical service, or the extra energy costs exceed hash returns. Even $1 more in energy for $1.01 extra mined is profit, and across 2000 cards in a month it's not a trivial amount.
Those miners gotta be some giant meatheads to even consider this card. I mean, get 2 RTX 3090s and you get much more than one of these, for a considerable less chunk of cash.
@@tridiots3681 this just shows how little you understand mining. Double the upfront cost for double the efficiency is far better financially before you even take into account that it only takes up one slot, one power cable and you can stack them close and not worry about cooling. For the serious miner these cards a god send.
In this video i learned about HBM. That its mounted closer to GPU, gives way higher bandwidth and its expensive. I even learned a little about crypto mining. I watch You Linus for years, for humor and education - keep it Up. If its possible, put even more knowledge in the videos. I think Your viewers brains can take it ;)
HBM unfortunately breaks very quickly. It's also why Vega 56/64 and Radeon VII are insane for mining. Could technically mine on an R9 Fury as well, but they only have 4GB of HBM and consume tons of power.
@@Akkbar21 uses less power than gaming lmao. Making a living is more important than mindless recreation anyway. I have an effectively 'free' 3080Ti + 10850k gaming PC now because of mining, even with the crazy market prices today. Only an idiot with no financial sense wouldn't mine on any GPUs they have when they're not being used for gaming. You really thing an astute businessman like Linus doesn't have a mining rig set up with all the spare GPUs they have lying around? Also I don't see the point of CMP when four 2060 supers or 5600XTs have the same hashrate of this card for half the price. Sure it's more efficient but the difference in power costs will take years to make up for the difference in GPU costs.
@@The_Bird_Bird_Harder actually, the cards are basically free after 6-8 months of mining, even in this inflated market. Do you not understand how the financial system works kid? It's called return on investment and it's why I haven't paid a cent of my own money for my gaming PC. Don't be stupid.
I would love to see more videos on weird crypto cards. This kind of video fascinates me because it shows how often preexisting technology can be used in a different way, usually for applications different to those of a consumer. I really enjoy seeing technology that we don't usually hear about because it is not designed for us to buy.
These aren't weird crypto cards (by the way crypto stands for cryptography not cryptocurrency), they are cards for general purpose computing useful for accelerating many tasks of which mining cryptocurrencies is the least important.
@@tokiomutex4148 Thank you for your comment. I assumed their main use was for cryptocurrency mining, based on the content of the video. Now I know better. Now I would enjoy a video from LTT outlining the other uses of this card and the general technology behind it.
@@christopherparsons7125 They could show some video transcoding, 3D rendering and physics simulations maybe accelerated compression, I can't imagine Linus training AI in an LTT video
@@tokiomutex4148 No, I can't see Linus training AI. I would be interested if they did a video on the technology that makes this card so special in some workloads (I know this is touched on in this video).
When you launch Blender in Optix with a new graphics card, it will spend 5+ minutes sorting itself out the first time. Not sure how long they waited, but yeah, it just stares at you for a while the first time out.
The front black fins are there to line up the airflow to go through the copper heatsink more efficiently. The back looks like it ties into the housing to help dissipate some of the heat it will collect out the exhaust instead of allowing it all to radiate into the case.
That milled alu chassis and hand soldering are consistent with low volume production methods. Perhaps you were given one of the pre-production test mules or and early production model that's been around the block once or twice, and they've improved the manufacturing efficiency since then?
Tbh i think that efficiency is what is so cool about these cards. Imagine a mining farm running of renewable energy using efficient cards like this. Now imagine a hack or new driver for these cards in 2 years time when they are otherwise e waiste allowing them to be used for compute in other applications at similar efficiency.
alternatively, ditch crypto bc the computations themselves are pointless, make at least the firmware and drivers open (if not the hardware itself) and use them to run useful things off renewable energy, like protein folding and deep learning applications the efficiency would be impressive if applied elsewhere, that's the big thing
@@ashlyy1341 Crypto is actually really useful. It's used in everything security related. So I imagine these cards could get reused for security purposes.
@@teddysmith8725 true, for vpns or stuff like this they might be useful but I'm entirely not sure if it would work well for real time because of the pcie 1.1 x4
@@leeroyjenkins0 Yes, he was referring to crypto mining, which is of practically zero real world use. But these cards are not ASICs and their compute cores can be used for a lot more than mining. The lack of memory bandwidth will be a problem for a lot of applications, but it will be useful for applications that do a lot of computing and return relatively small tensors/matrix arrays/values. ML could be a decent application for this, depending on the model as well as VPNs/server related setups. (Edit: Also, ASICs can be used for cryptography research)
@Linus Tech Tips Those cards are 30 million USD minimum order at 3000 USD a piece. So a minimum order of 10k pieces. I was looking for those mining cards earlier this year as we had a client that possibly wanted 600 gpu's. When we contacted an Nvidia rep (this was months ago) they basically laughed at us when we said we needed 600 and would not put us in contact with a reseller.
I'd love to see more information on the 3D load (supposedly) caused by passing graphics though the card running the display. Seeing a similar issue on an XPS 15 9510 with both integrated and dedicated graphics, and it leaves the dedicated card almost useless - as the passthrough causes 100% load on the integrated card, bottlenecking the whole process.
UPDATE: We managed to get Blender to run on the CMP card after posting this video. The results are not very good due to limited memory and PCIe bandwidth. Here are render time results for the BMW blender demo. Lower is better.
CMP 170HX
- CUDA: 36.82s
- Optix: 39.33s
RTX 3090 TUF
- CUDA: 18.40s
- Optix: 10.56s
Well reason you get crap performance is a VM can only see 4 of the GPU clusters per VM. To expose more to a VM you must buy NVidia VGPU licenses. Run into it often with GRID and ESRI.
@@thepoliticalstartrek Did I miss something in the video stating that they were using a VM?
Interesting
Don't lie to us claiming it's "not worth it" to take it apart. Spending $4000 on a card is nothing to you. You'll make it back and then some with this video alone.
Can you do AI related work on it?
With the state of the GPU market "$5000 graphics card" could mean anything from a 3090 to a GT 1030
true KEK
stonks
ah yes a 3090 at that price must be broken or a scam. GT 1030 for the win
Lmao?
i wouldnt even sell my prized GT710 from my "Fast Fortnite Gaming Pc" from ebay
"its his five grand, not yours"
best thing an employee has ever said
@SuperWhisk while annoying the shit out of Linus for the viewers. They're the fan-insert characters.
@@Psyopcyclops Jake used to be pretty bad in front of the camera back when Linus was constantly getting irritated with him. Looks like Linus mellowed out with him though.
@@Psyopcyclops I don't really remember linus ever having a problem with him but the comments *hated* him, I think he was like 19 when he joined
from an engineering standpoint. the black fins on the ends are probably "Flow straighteners" they reduce the mixing of air in the channel to decrease differential pressure through the cooling section. allowing for a larger heat transfer from the fins to the cooling medium.
Ah this makes a lot of sense. Essentially creating a uniform flow distribution for the heated air?
Aka creates something closer to laminar flow
I read everything you wrote Nathan, and I realized I agreed for one reason: I didn't know what the actual fuck you were talking about, but it sounded so good, I had to believe it. 💪
This is how politicians work
@@em4392 dumb
“The clock goes up but it doesn’t seem to be doing anything” it’s doing nothing, but faster
That's what happens when you're waiting for the bus and you get a call telling you to do it "faster"
This reminded me of: ua-cam.com/video/oZRKozAzZmY/v-deo.html
@@frecio231 **waits for the bus faster**
OoOool
Its funny because if you know how computer works this can be true
Modern day Indianna Jones goes on a quest to acquire the forbidden GPU during a worldwide chip shortage.
That's how I feel whenever browsing neweggs or ebay
Ah yes, the """"""""shortage"""""""""
@@romanlinnik7441 what do you mean "shortage"
Worldwide everything shortage tbh
Hey indy, feel lucky to say nvidia made this prod cause chips have some bad things if used other things? That crypto-thing eats crazy amount energy, but hey old bank system eat more ;)
"Started with a slab of aluminum, milled away 90+% of it...." That's aerospace manufacturing in a nutshell.
Guess that makes the AR15 make a whole lot more sense, since ArmaLite was apart of Fairchild Aerospace!
thats just any sort of machine work really haha, lathes, mills etc
@@MEMUNDOLOL I think his point is that they could have used cast or stamped aluminum instead of milled.
@@gregorycosteas3539 Milled is always better than cast or stamped.
edit: From a quality build stand point.
@@noahbuzelli4245 why?, the piece of aluminium you're milling itself is cast
Linus is putting so much care into this card specifically. If I remember correctly, he shook a $4000 Titan out of its anti-static bag like it was a $5 Christmas gift.
Probably because it is out of stock
@@AirNeat stonkx
LooOool
Probably because he's actually going to use it for mining afterwards
because he didnt pay out of pocket for it XD
I love when Jake and Linus do a video together. The unscripted troubleshooting bits are great.
a bot stole ur comment
@@oldacc_ihatebrickplanet12 your mom stole you
@@jesuschrist4185 at least his mom wanted him
@@catalyst429 from the orphanage?
@@jesuschrist4185 no, just dodged an abortion and have a decent dad 🤙
I hope they start releasing more of these "mining" cards so the rest of the market can be freed up and we can actually get 1080's again
😭
it's not helping. they just have more gpus to buy now
Ok, ask yourself this, why would they buy a card that takes longer to pay itself off? It literally doesn't improve anything for anybody, not even miners as there's no resale value there.
I just want crypto to fail already so we can get 1080s
These cards have shit resale value
Just to let you know, when you're doing your very first render on the Eeves (or Cycles i don't remember) engine, the engine has to compile itself before actually doing rendering. After waiting like 10-15 minutes you will be able to do rendering after rendering.
You can see at 19:22 on the line where you can see your frame count, the message "Loading render kernel (may take a few minutes for the first time" indicating you that you will have to wait a bit.
Hope you will test this out later or read this comment 😀
Could be, but Blender not responding is a hint it probably crashed and would never render. In my experience the BMW test also does not take more than 1-2mins of compilation. Edit: Maybe im wrong, my experience is without OptiX and only with OpenCL / older CUDA
@@somestarwarsnerd9068 ..Blender not responding is as much a hint as it simply not responding for several other reasons outside the specific GPU... How many of us* have had to revert to an auto-save after a lock up or crash & it wasn't due to hardware issues to start with? Lol.. No matter how 'simple' or complex a given scene is, there's always the unexpected.
Furthermore, if due to instability- you do the same thing again and again and prove it IS the card; or something related to it (to help give your comment a sense of merit). But that wasn't the case now, was it?.. /shrug
Edit: "*" typo
They did! Check the pinned comment for benchmarks
The OptiX render kernel isnt precompiled in blender and always takes time in new scenes. CUDA on the other hand comes precompiled with blender so there isnt any kernel compilation happening anymore.
This is why alot of times CUDA will be faster on big single frame scenes that uses nearly all Cycles features, its already done when OptiX is still compiling. When it comes to rendering out big animations OptiX will get the throne.
But yeah this is why in my opinion these OptiX vs CUDA tests in blender are crap because they dont show the complete picture
@@somestarwarsnerd9068 No not at all. Blender is mostly single threaded and is poorly coded in python. Everything runs on one thread including the user interface. Most programs use at least 2. One should be reserved solely for the user interface. That way when you run tasks on the other thread it doesn’t freeze the interface. Even when blender uses more than one thread it still likes to run stuff on the user interface thread. So when doing any operation it freezes. This is because when operating on a single thread code most run in order. So The user interface cannot update until the other code runs. Windows considers programs with unresponsive interfaces to be unresponsive even if it’s just because they are running long code in the UI thread. When modeling in blender applying modifiers or moving objects or editing them in anyway way really can cause blender to become unresponsive but it almost always finishes whatever code it is running and comes back. The exception to this would be if it runs into an error and gets stuck in an infinite loop. Then it stays frozen forever and there is no way to cancel it. This is a common issue with modifiers because they use algorithms which do not always work on the object but blender isn’t smart enough to figure that out so it just keeps retrying. It has a lot of work left to do and certainly is not an ideal program to be using for these kinds of test. Blender is by far the best 3D modeling software and for the best software for physics, animation, and many other things too. It is free to yet still significantly better than stuff that coats you to pay several thousand dollars a year. It shameful for it to cost that much in the first place but subscription based model on top of that really is bullshit. People should own the stuff they buy especially when it’s that expensive. It doesn’t even cost them anything it’s 100 percent profit. They spend a small amount developing the software after the first few customers all the expenses are paid for but the price never drops. And the super cheap support that you may need to use once every few years for a few seconds doesn’t justify it either. The initial cost can easily cover that. Blender while open source still has a foundation with paid employees which has much more expensive and educated employees and much more of them. The research and programming that they put into it goes above and beyond anything any of the companies do yet they do it with a very small fraction of the money that they only get from donations. Things like autodesk and maya may have 50 developers at the most. Autodesk doesn’t even make most of its own software. They buy successful software. Then they make it free to bankrupt any competition out of business. Once the competition is gone they take away the free version and sell it at an extreme cost or at least limit the free version to the point where it is unusable. Almost none of their software is original and not purchased from another company. These companies also have a very small market share. If you take the publicly available revenue for these companies and only look at the small portion of it that they get from their software than divide that by the insane cost you find out that the number of copies they are selling is actually incredibly small. They just want you to think everyone is using it. The only way they stay a float is by making partnerships with universities and collages to teach people to use their software for free so that once they go out into the work force that is what they use. They also makes exclusivity deals with big companies and get them using their software. And once a company is already using a software and the people coming out of school are too it means it would take a huge up front investment of time and money to switch to blender or another alternative even if they would more than make the money back in the long term they don’t want to make the risk and may not have the cash flow to do so. These companies and their software really are unethical predatory crap that isn’t even as good as free alternatives that they didn’t even make or contribute much to. They just steal the ideas of others. Blender however is only better when it comes to capabilities and features. When it comes to performance with large scenes it is still behind. Plus because it has so many features they are often buried in menus and with complicated keyboard shortcuts and they like to change where everything is each update making it a real pain. You can do a lot more with blender but expect it to be super slow and unresponsive while doing it. It doesn’t even matter if you have a super computer because it can easily saturate that too. The only real fix is to recode blender from the ground up because it’s issues are so deeply rooted. They have almost completely redone it a few times and have made big improvements but it’s got a ways to go. They plan on fixing a lot though in the next major update then maybe it will finally make its way into the big companies and snuff out the other crap.
The biggest flex is getting a gpu case made entirely out of aluminium even though no one can tell
Glad we aren't wasting aluminum (aluminum alloy) on this.
Fun fact we are facing a magnesium shortage which will affect aluminum alloy production... Which gets used in Many products including cars. So yes...
@@robertt9342 magnesium shortage? lol just strip mine using a fortune III and eff V diamond pick, mining isnt that hard
It's a big heat sink - maybe not very effective due to placement
Linus is right too. Looks like it was a solid thick piece of aluminum and they just milled 90% of it away. Oof.
@@robertt9342 everything is undershortage right now
The outer fins kinda seem like they guide the air a bit more since running copper fins that far would be expensive. Also they cool the card body and having all those cards nearby in slots would probably be pretty good at keeping them all cooler to the touch.
Watching them trying to understand this project really makes me think that it would be a really good idea to invite current/former industry engineers onto the show to explain things to them and the audience from the perspective of an actual expert.
I would agree. What are we missing?
@@southwestaudiovisual5069 I think he means so they have a more clear and precise explanation vs endless theorization
A part of the entertainment is watching enthusiasts working out what this stuff is. If I wanted to watch a video of someone explaining a product, I would just subscribe to companies
@@JoshuaValentine26 This echoes my sentiments.
Experts probably have NDAs signed if they are in a position where they don't have to speculate.
I work with A100's on trays of 8 in servers for a major cloud provider. I can attest that we've had nothing but problems with the 40GB A100 trays and individual GPUs causing a seemingly never ending RMA process. Nvidia has been having big QC issues with their production.
This card was clearly made with low volumes in mind. I mean a bunch of the stuff you guys pointed out and were confused by is likely influenced by the relatively low volume.
6:12 Believe it or not, that's a very common ways to connect a wire to a board at low volumes. Or high volumes with some removable custom fixtures. A lot of LED boards in the small red & yellow lights on semi tractor-trailers are constructed this way so you know it's reliable if done right. In this case, it also allows the connector to be screwed to the enclosure independent of the card, which helps reduce strains on the solder joint due to vibration, etc.
7:12 Milling away 90% of the material is fine - you just dump the scrap in a vat to burn a away the cutting fluid and melt the slurry, clean it up a bit, and cast another block. Great recovery rates.
As for why, it's a low investment way to get a box with 3 sides. The cost per piece is going to be on the order of $10, and you don't need to spend a few hundred thousand dollars on a forming die, to then still spend about $10 per piece on top of that because of your low volumes.
On a related note, 5:46 I wonder if those heat sinks are for cooling the enclosure. Aluminum has got to pick up lots of heat from adjacent cards and this helps carry some of that away.
^ this needs more upvotes
Regarding the soldered connector: soldering those couple of wires is likely much cheaper than using a larger PCB just to bring the connector out further. It is also less likely to fail than a directly soldered connector and it does not bend the PCB. All around, not a bad design.
I know next to nothing about manufacturing or tech and was thinking many of the same things. Why incur "conventional" costs for an unconventional product...
100% for you
The heat sink is clearly designed to have some thermal mass to stabilize long term temperature fluctuations and cooling requirements when running these at scale at 100% uptime.
it was probably cheaper to use the aluminum back plate from another card rather than create and order a new design of backplate for that card, they probably had to order them in bulk from another company for the other card so they used the access plates or defective units for this card since visuals don't matter. might be a strange area where being cost effective looks nice, feels great.
It's simply a matter of binning. Whoever said binning is not a thing anymore is an idiot. These card's silicone wouldn't be suitable for a gaming environment, but for more focused one, like crypto mining, they are more than enough. It's a great way to use everything and also finally stop the gaming cards shortage.
The question that remains: How much has Nvidia already earned by mining for itself?
I know this is an old comment but I always wondered this. Even if they just did a 5 hour validation test on each card they manufacture just to make sure it will preform to it specs... that is a lot of hashpower... lol
its probably more profitable to design and sell the tools to the rubes than waste time and pr to mine themselves.
@@archieames1968 wrong
@@archieames1968 in a gold rush don't mine gold, just sell the shovels
@@Filipex13 exactly, the only surefire way of making a profit in a gold rush is to sell the shovels
Can't wait to see how much these babies go for on eBay in a few months :'D
*days
@@justiceplethora2682 hours*
Shit, before it was released you’d probably see it listed for $10,000
You ever try to take out a mortgage on a graphics card?
Well the aluminium block is quite expensive. :D
I believe the milled back plate would indicate their expected volume for this product. If they were to cast it and then mill it that would be expensive molding costs. On low volume milling is sometimes more cost effective even if you’re milling 90% away. Just hog it out with a big tool to rough in and then finish it.
But it doesn't need a milled backplate.
Just a bend sheetmetal or plastic part would function the same.
@@Jehty_ plastic injection molding requires tooling, sheet metal stamps also would require tooling.
yeah i thought the same thing when i saw the hand soldered supplemental power connector. obviously arent being sold in high volume
My guess is that it is an extruded profile that they mill down for clearance. It's kind of hard to tell from the pictures though.
@@Finder245 my thoughts as well
I used to work at an hp assembly factory putting together servers, testing them and boxing them up to ship off to whatever country. There was one line boxing up these cards, or an earlier version of them about 6 years ago. Same style minimal casing, no output ports and I could never find them talked about on an Nvidia site but could find other people selling them for 4k+
Vipera Tech's website: "Error establishing a database connection."
Guess they've been Linus'd.
technical term is "hug of death"
2 hours later still down.
no no. the database DROPed
ill see myself out
Yes, that looks identical to an A100 PCIE card, and the IHS is on the PCIE version. The A100 does have 3 NVLINK spots as well.
(Ps) on the $12k A100 the power connection is soldered just like the mining card is.
A100 has floating computations + INT computations processors
Maybe this mining card has only INT computations and zero floating point computations processors
@@Haskellerz I doubt it. I do some work on deep learning. It would be interesting to test. There’s some simple ways using Tensorflow to test for speed up. It’s down to if the tensor cores are still active.
@@davidedwards7172 wouldn't bother. 8G memory is too big of a limit for it to be useful.
@@jfolz not gonna argue with you there, there are some CUDA tricks around that. If the card sold for a lot less it might be worth it. Especially with the PCIE 1 link, it would be really slow. It would be fun to do a deep dive and see what functions are still left alive on the chip.
Would nvlink work?
And it's now completely worthless
$400 😂
Frfr
those black fins at the ends might be intended as baffles to direct air into the enclosure, as you will have some amount of turbulence with narrow ducts like that, and turbulence can prevent the air from pushing through it
what I was thinking is the airflow straighteners you see on some fans, but that'd be the same idea
19:21 thats a blender thing. it does it every once in a while at the start of a render, just let it do its thing. sometimes it takes 20 seconds other times it takes upwards of 5 minutes
Yeah
Yeah
Just saying, I've been loving seeing some of the more enterprise stuff you guys cover. Really cool to see the back end of things.
Worth pointing out that the A100 doesn't have NVENC, only NVDEC. It also lacks most of a graphics pipeline, so not really usable for much except pure CUDA compute
Some GitLab projects then i guess that uses CUDA or Tensor cores for calculations and stuff.. ?
but the mining cards only having 8GB of HBM2 really limits their usability in compute workloads. Pcie 1.1x16 does not help either :P
"This $5000 Graphics Card can't game"
Me: Ohh Linus is going Old school by talking about the GT730
Lol that’s what I use
The real reason for the slow PCI is that this would go into a big, cheap chassis with a huge PCI fanout - think 8 cards in parallel using PCI 1.1 risers so that it can fit into a 4u server case. The rest of the design seems to be optimized for energy and cost efficiency.
I worked for one software company that had AMD make a custom 'graphics' card. It was super duper odd and impressive since it did a damn good job. Really made me think how much faster I could build and test my stuff if the entire computer was designed for it.
I think apple is trying to do the same with the new MacBook Pros.
yes, that is the whole point why RISC architectures like ARM exist in the first place.
the problem is that you lose all other functionality. sometimes it is better to have 100% at one thing, other times it is better to have 60% at all the things.
it's a trade-off
@@robertlinke2666 yeup. Can only use two cards
FPGA adaptive accelerators are the future, the hardware can be “re-programmed” to adapt to your workload and improve performance tremendously.
The efficiency benefit is probably related to memory type as well. HBM is more efficient than GDDR. If this card has the same tensor core capabilities as the A100, it will be a pretty good deep learning card, though the 8GB memory is too small for that.
so it won’t
I almost thought that I was going to be able to score a cheep A100 until I heard about the 8GB memory :(
Yeah 8GB. Makes no sense. You cant train anything for that except for very small networks. I mean this is why you can get the clusters of A100. I personally had problems with the AWS EC2-P4d instances on big models. And they have up to 8 A100.
That's the point. They also don't want people buying 5k mining cards for deep learning instead of the 10k+ ones...
Nah, memory chips use like 1w of power. It's all core when it comes to power usage.
Its possible the black fins are to ensure that the air passes in a straight line through the card rather than whirling around. They probably found extending the copper didnt help with temps as thermal transfer isnt affective that far from the die (or ihs) in this case
After stripping it out of some of the parts that makes it an unit for processing graphics, should we still call it a GPU? I mean, I guess if we strip a car out of its wheels and hang it up in a fixed place for using its engine for powering something else... It's still an automobile... which is not being used for automatic mobilization.
I guess it's still a car engine, even if it's not being used to power horseless carriages.
@@technicalfool Generator. Stationary engine. Marine engine. :P
"movilization" and "automovile" bruh
Marbelous conclusion.
@@bigboi9856 in some languages, "mobile" variations have a V, in example "automobile" in portuguese would be "automóvel"
i can see why they might have done what they did with the chassis around the card. the big thing is that its serving as a shroud to channel air both through the heat-sink and then passing it all the way to the end of the card.
the dummy heat-sinks on either end, those server fans can move a lot of air and those are going to act as grills to help protect against things being pushed/pulled into the card by the air flow regardless of which direction the server is using.
i can also see the aluminum box being useful in those server racks with protection against things pushed against the card. i think the high end milled aluminum surface is a bit out there, but if it remains a long term thing they can probably either l simplify it or switch to a casting. but if they use it for other designs or it remains in low rate production, its probably as cheap as any other option.
That power connector extension has been normal on Quadro for years now. Since all Quadro/Tesla boards are standard length and have rear-facing power connectors, they make a cable extension whenever the PCB is shorter than the shroud.
Cool. I was thinking maybe it was for added durability (for some reason) so there would be no chance of breaking the solder connections to the board.
"Jam it in and twist it?"
"I tried, he wouldn't let me."
Some quality out of context quotes right there.
LMFAO!
@@sinuslebastian6366 just here to see this comment blow up
Lol I read this as they said it 😂
"What? You're stripping already?!"
"Is it gonna come out?"
"I just touched...all over it."
"Oh wow! Oh gosh! Oh gosh! Oh...gentle! Oh gosh!"
here before it blows up
When linus said "in true LTT fashion..." I was expecting him to drop it.
lmao
@@H.A.L9000 I taught you were replying to your own comment for a moment there
19:43 "All its base are belong to us"
LMAO NICE ZERO WING REFERENCE
I think being curious about obscure hardware is a good reason to explore this
Well, there is one problem though. The Manufacturing for the CMP Card has to be splitted even if it is not on the same process. This uses TSMC 7nm, like the Ryzen 5000 Series Processors or AMDs Graphics Cards for that same matter. So this new Card may not cut on NVidias own productions, yet. But it may hurt other ones productions.
Probably a happy little accident for Nvidia then.
@@TheRealAtello But is it an accident though? ;)
Actually it sounds like this card runs from the discarded datacenter production line of their products So the main purpose of that line is to produce the card that end up in the true compute datacenters and these cards are the ones that fail QC for that branch and they can't create a tier for. So instead they squeeze out money from the miners that are willing to drop 5k on it.
Its funny how many people think that people that mine in bulk actually use these cards to mine XD
8:09 Linus: don’t touch the copper
Also Linus: touches the copper
"It's crazy that this thing has a Windows driver, because who in their right mind would be running this on Windows"
Well, it is an Nvidia card
@Bat Barfalamule ah, I see you don't use linux or macos.
AMD are the best option on any platform that isn't windows.
@Bat Barfalamule huh, I don't have too many issues running my desktop PC with a 5700XT and tesla K80 at the same time.
Maybe i'm just lucky.
(Dual booting, using *nix more often)
@Bat Barfalamule clueless lol
@Bat Barfalamule Stop talking like that you scaring the kids lol. Btw Nvidia users like me experience the same pain you saw on this video but with REGULAR GPUs on Linux! 😂 That's what the OP of this comment just meant. Nvidia is the REAL PAIN for any other platform than Windows.
@Bat Barfalamule Literally plug and play and modern driver interface unlike nvidia.
Imagine a graphics card being sold out that does not even do anything or exists in itself…
Seems very probable these days.
Alex, what are diamonds?
well im imagining how a product that came to the market sold out instantly..
thats like dropping something from your hand towards the floor but it never hits the floor,
it just disappeared :p
Cryptomining
@@emanu1674 yeah dont expect kids to understand its a mining card so they stop complaining about shortages and over pricing from us terrible miners hogging
7:26
More likely they cast it and milled away any small amount of excess.
19:22
Blender especially with OPTIX takes time to build render kernels during the initial render (can be 10 minutes depending on scene complexity).
A little more patience might have yielded a different result especially considering that blender depends on CUDA compute.
10:40
Being in the process of scaling up for manufacture, it's development vs. volume costs. If it'll take your team a week to adjust things so that the connector is in the right place, and another 2 for verification and testing, and you're only making a few thousand of the things, it's cheaper to just pay the factory workers to fiddle with wires, rather than the engineers to develop a solution that doesn't need them.
I'm starting to see how that function evaluates now, being the person who both develops and manufactures the parts, which things to spend time developing versus simply producing as they are is one of the harder decisions.
Current LHR unlocks are at 74% not 71%. Really only effects performance on one crypto currency too.
I recently went through the process of getting a electromagnetics simulation server for my job and we got an A100 PCIe 40GB. Really had to hold back and not take it apart. It absolutely slams in CST Studio Suite 2021 though. No complaints about the power and sim time reduction. Really such a bizarre card and such a specialized product that it's completely underwhelming for most people
In the event the need for that card evaporates, or gets replaced, it is your moral duty to the computer community to strip that thing to the traces.
Mining cards are great for mining, and thats it. Its a professional mining card, as as Linus mentions even for most miners it wouldn't be worth it because you cannot repurpose a card so it leads to landfill waste when they are no longer making sufficient profit. I get that they don't want the card to be repurposed but its not a good practice ethically or logically.
The ROI on that card would be insane too.. dont see hows its practical for anyone
Well mining isn't a good practice, ethically or logically either, so that tracks at least...
@@Arclite02 good to know that you don't understand purpose of mining proces.
@@SlawcioD I understand it just fine. Billions upon billions of dollars in wasted equipment, colossal amounts of wasted power and needless emissions, all to make a small group of people who have the millions of dollars to set up a mining farm, obscenely rich at the expense of everyone else.
@@SlawcioD Tell how is mining a good thing?
Currently in a mechanical engineering major at a top university in canada, I think the fins linus was confused about was intended to oriente the air into a lamar flow which would increase air flow and air speed through the actual heatsink mounted to the chip, with all the air moving in the same direction (as the other air and the fins) less energy is lost between particle interactions compared to if the air had to oriente in the heatsink. This assumes that the fans output is turbulent which I dont know for sure.
I have actually heard that closely packed GPUs can cause mining to slow due to electromagnetic interference. I wonder if that was one of the reasons to encase the whole card in metal?
That'd be fun to test! :D Regardless of whether it is true or not, the great questions lead to great discoveries!
@@tradingnichols2255 it would be fairly expensive if you don’t already have the cards but it should be just a matter of getting two of those rigid connecting bus card things filling them with mining cards and putting one upside down over the others or just getting the flexible PCIe cables and sticking the cards as close to tether as you can. And then fiddling with the shielding on the cards and the card positions.
@@glenecollins I would assume so he always encases the shell
holy shit the blender part drove me insane, "LOADING RENDER KERNEL, THIS CAN TAKE A FEW MINUTES"
You just had to wait two minutes and maybe READ the text on the screen
You are talking to guy that bricked GUI from installing Steam by not reading things.
@@thischannelisforcommenting5680 lmao true
@@thischannelisforcommenting5680 and then blamed linux
What you guys should do is a watt for watt comparison of 30 series vs 40 series cards and include some odd ones like workstation cards. Cards like the rtx a5000 and that a100, if you still have it. Target certain break points, like no aux power, 1 6-pin, 1 8-pin, 2 8-pin, etc.
Alex the actual engineer. "Just rip it"
Linus looked a bit upset just at that moment. The pressure of breaking a $5k card might have blocked his understanding of humor right then.😄
_Sigh_
I just want a 700 dollar GPU.
Same
me too
Yep
Same. Even $700 is stretching it by quite a bit.
I still remember when we can get rx580s and gtx 1060s for a reasonable price
Linus, please stop making videos about anything else! This guy in particular doesn't care, his opinion is the most relevant out of all
Not sure where to ask this? Would LOVE LOVE to see a complete how to vid 😉 for properly setting gaming settings? Especially for mid to low range computer set ups??? This would be soooo helpful as issues frequently comes up.
As an aeronautical engineer i kinda cringed when i saw that 3d printed part. It will perform wat better if you decrease the angle of the slope and give it more space after passing it through the fan. If your any good at cad you could make it elliptical but at these airspeeds thats just for minor gains
I am thinking of making one of these ducts for a similar project and was wondering how long it should be?
Imagine Someone buying this for their son not knowing anything about it.
Son: “I guess my time as a gamer is up. Now I must become a minor”
I remember back in the days when my dad bought me a Quadro GPU as a present 😂
Then their son has a chance of becoming a data scientist.
I can't cuz nobody can afford 5k gpu for their sons lol
There is a reason Nvidia doesn't sell it alongside there normal cards, they don't want that to happen
Coming back to this man's channel and the intro music still the same. Epic love.
Another reason for the efficiency bump is HBM also uses much less power than typical graphics memory.
NVidia: We need to change something.
Designer: Minimalist logo?
NVidia: No, minimalist GPU.
Designer: [Exploding Head]
Designer: [dies]
@@lioscar80 yes, he would die after his head being blown off well done
Why the hell would you buy this? You can get more hashrate spending your 5k on two 3090's.
Availability. Not to mention you're paying more than 5k for 2 3090s.
@@MrHakubi but you can always sell them anytime you want.
@@benthelion7105 you don't need to worry about re-sale value if your mining in a farm
I don't know if linus mentioned it, but I'd say it safe to assume they have lower electricity cost than two 3090's
You're also not being a douche and depriving people that actually are using them for a setup.
I am waiting for the result of that meeting linus got set up with UA-cam management. I believe it is going to be a 15 minute "$hit management says" episode XD
My money is on UA-cam not doing shit, even with Linus on their case. They don't give a fuck. Only about corporations and advertisers.
I imagine Nvidia got a REALLY good deal on those copper heats, then realized it didn't match length of the enclosure of the card which were already in production. So got those "fake" fins to fill out the card more and give it a professional look
Another comment I read said they were probably flow straighteners which help even out the distribution of air across the internal channels for better cooling efficiency.
6:17
What are you laughing at?
This flippin' circuit board Jan.
I understood that reference
The aluminum grates might induce Laminar flow. Does that make air convection better or worse?
Generally worse
Miners stealing GPUs from gamers to mine, Linus stealing mining GPUs from miners to game. Linus has played the Uno Reverse card.
its not stealing lol.
@@finkyfamboni4333 You are correct. Mining is not stealing, it is a war crime
@@1000_Gibibit innit I legit was going to build my pc last year for around 1.8k AUD with a 2070 super OC, now I have to build it for 2.6k this year with a 2060 OC, I hate miners
@@tobias8351 fam you didn’t pay 929 AUD for a 3 year old gpu
You're correct that is not gaming silicon, it would be compute anyway, but that is only so because nvidia choose for it to work that way! It's one thing to save on the output (they could have put only one like some old firepros but whatever) but to completely lock it down when people have been asking for years for server features like sriov just shows once again that they simply do not care
But does the output matter when the way it’s configured already makes it a horrible gpu for gaming. (I’m rather new to the hobby so correct me if I’m wrong)
11:50 why not just use a massive thermal pad on the back for slightly better cooling because the backplate not serves a function
I'd say those black aluminium fins are to clear out air vortexes before the actual copper cooler to get better air flow and there for efficiency thru the chassis. They look good, but could also have a function too.
now people have pallets loaded with GPUs they cant sell. "trust me bro this is the future."
"How can this be 40-50% faster than the top tier Geforce GPU and double the efficiency?" You forgot that Nvidia restricted the vbios/drivers.
Well said. It's their drivers restrictions
only on LHR cards all the rest are full hash rate the reason they are 40 - 50% faster is because of memory bandwidth like said in the video.
The problem with this efficiency is it's not extreme enough. Two 3080's will mine $13-15/day and use $1.5 in energy. This card will use $0.6 and mine $11-12.5. Thus the two 3080's will net hundreds more profit but cost hundreds less.
In less words, efficiency only matters when it limits hash due to heat, you've maxed out your electrical service, or the extra energy costs exceed hash returns. Even $1 more in energy for $1.01 extra mined is profit, and across 2000 cards in a month it's not a trivial amount.
@@hydrocarbon82 my 3080s are happy to run 102mh at 226w. I don't see how this card makes sense for anything.
12:30 next up: we turned this LTT Waterbottle into a Gaming PC!
That is actually a good idea it would be more possible if it was huge
A spanish youtuber by the name of SFDX show gave new life to one of this cmp cards its insane.
Watching the guys geek out over this thing was a joy to watch. It seemed like they were pumped! More of these please!
Yes, exactly!
This was better than the $10,000 TV monitor. At least it didn't have a dead pixel in a noticeable location, just hand solder job.
Well thank you, I hope the miners are going to buy more of these cards and give us back the Gaming GPU’s
Noone is going to buy these cards at 5000 usd they are useless because you cant sell them
Those miners gotta be some giant meatheads to even consider this card. I mean, get 2 RTX 3090s and you get much more than one of these, for a considerable less chunk of cash.
@@tridiots3681 ssssht don’t spoil the beans. We need them GPU’s back. 😉
they will when Nvidia wakes up and reduces the prices by 5x on those cmp cards. until then, gaming cards are going to be the cards that miners buy.
@@tridiots3681 this just shows how little you understand mining. Double the upfront cost for double the efficiency is far better financially before you even take into account that it only takes up one slot, one power cable and you can stack them close and not worry about cooling.
For the serious miner these cards a god send.
These youtube ads are getting ridiculous - 25 seconds of unskipable monstrosity
GMiner needs to update their readme to say "used by LTT."
"I dont want to twist on an component on this card"
"Dont worry, its a big component" :D
Linus: "$5000 Graphics Card"
Me: "Oh, you got a 3060? Nice"
In this video i learned about HBM. That its mounted closer to GPU, gives way higher bandwidth and its expensive. I even learned a little about crypto mining.
I watch You Linus for years, for humor and education - keep it Up.
If its possible, put even more knowledge in the videos. I think Your viewers brains can take it ;)
HBM unfortunately breaks very quickly. It's also why Vega 56/64 and Radeon VII are insane for mining. Could technically mine on an R9 Fury as well, but they only have 4GB of HBM and consume tons of power.
$5000 graphics card?
You mean the 3060 this guy in my town was trying to sell for a profit lmao
Love the black aluminium fins covering a copper heatsink xD
Next time: "today we are going to try water-cooling this bad boy"
i'd love to get some cheap cast off CMP cards for off-coin stuff like doge or monero. not gonna pay thousands though
i'd love to get some cheap cast off RTX card for gaming stuff like for wukong: black myth or stalker 2. not gonna pay thousands though
@@Akkbar21 uses less power than gaming lmao. Making a living is more important than mindless recreation anyway. I have an effectively 'free' 3080Ti + 10850k gaming PC now because of mining, even with the crazy market prices today. Only an idiot with no financial sense wouldn't mine on any GPUs they have when they're not being used for gaming. You really thing an astute businessman like Linus doesn't have a mining rig set up with all the spare GPUs they have lying around?
Also I don't see the point of CMP when four 2060 supers or 5600XTs have the same hashrate of this card for half the price. Sure it's more efficient but the difference in power costs will take years to make up for the difference in GPU costs.
@@KingOfForest22 You, people like you specifically, are why a graphics card can cost more than an entire pre-built.
@@The_Bird_Bird_Harder actually, the cards are basically free after 6-8 months of mining, even in this inflated market. Do you not understand how the financial system works kid? It's called return on investment and it's why I haven't paid a cent of my own money for my gaming PC. Don't be stupid.
Linus: I respect your effort to keep this thing intact and fingerprint clean.
I would love to see more videos on weird crypto cards. This kind of video fascinates me because it shows how often preexisting technology can be used in a different way, usually for applications different to those of a consumer. I really enjoy seeing technology that we don't usually hear about because it is not designed for us to buy.
These aren't weird crypto cards (by the way crypto stands for cryptography not cryptocurrency), they are cards for general purpose computing useful for accelerating many tasks of which mining cryptocurrencies is the least important.
@@tokiomutex4148 Thank you for your comment. I assumed their main use was for cryptocurrency mining, based on the content of the video. Now I know better. Now I would enjoy a video from LTT outlining the other uses of this card and the general technology behind it.
@@christopherparsons7125 They could show some video transcoding, 3D rendering and physics simulations maybe accelerated compression, I can't imagine Linus training AI in an LTT video
@@tokiomutex4148 No, I can't see Linus training AI. I would be interested if they did a video on the technology that makes this card so special in some workloads (I know this is touched on in this video).
When you launch Blender in Optix with a new graphics card, it will spend 5+ minutes sorting itself out the first time. Not sure how long they waited, but yeah, it just stares at you for a while the first time out.
The front black fins are there to line up the airflow to go through the copper heatsink more efficiently. The back looks like it ties into the housing to help dissipate some of the heat it will collect out the exhaust instead of allowing it all to radiate into the case.
I love this kind of video, looking at weird tech is always awesome. Would love to see more stuff like this.
YOU GET SIGNAL!
Thumbs up for the AYBABTU reference.
That milled alu chassis and hand soldering are consistent with low volume production methods. Perhaps you were given one of the pre-production test mules or and early production model that's been around the block once or twice, and they've improved the manufacturing efficiency since then?
“Don’t touch the copper!” he says as he’s mushing his ungrounded fingers all over the PCB while wearing an ESD generating hoodie and no ESD strap. 😂
Tbh i think that efficiency is what is so cool about these cards. Imagine a mining farm running of renewable energy using efficient cards like this. Now imagine a hack or new driver for these cards in 2 years time when they are otherwise e waiste allowing them to be used for compute in other applications at similar efficiency.
alternatively, ditch crypto bc the computations themselves are pointless, make at least the firmware and drivers open (if not the hardware itself) and use them to run useful things off renewable energy, like protein folding and deep learning applications
the efficiency would be impressive if applied elsewhere, that's the big thing
@@ashlyy1341 Crypto is actually really useful. It's used in everything security related. So I imagine these cards could get reused for security purposes.
@@teddysmith8725 true, for vpns or stuff like this they might be useful but I'm entirely not sure if it would work well for real time because of the pcie 1.1 x4
@@leeroyjenkins0 Yes, he was referring to crypto mining, which is of practically zero real world use. But these cards are not ASICs and their compute cores can be used for a lot more than mining.
The lack of memory bandwidth will be a problem for a lot of applications, but it will be useful for applications that do a lot of computing and return relatively small tensors/matrix arrays/values. ML could be a decent application for this, depending on the model as well as VPNs/server related setups.
(Edit: Also, ASICs can be used for cryptography research)
@Linus Tech Tips Those cards are 30 million USD minimum order at 3000 USD a piece. So a minimum order of 10k pieces. I was looking for those mining cards earlier this year as we had a client that possibly wanted 600 gpu's. When we contacted an Nvidia rep (this was months ago) they basically laughed at us when we said we needed 600 and would not put us in contact with a reseller.
The metallurgy on this card costs around 3000$ 😂
"All its base are belong to us"
Linus, I want you to know you are heard and appreciated :)
boston food sucks
Linus set up us the meme.
We get signal _!!_
What you say!
The black fins are for laminar flow, it makes the air flow less turbulent thus more efficient
I'd love to see more information on the 3D load (supposedly) caused by passing graphics though the card running the display. Seeing a similar issue on an XPS 15 9510 with both integrated and dedicated graphics, and it leaves the dedicated card almost useless - as the passthrough causes 100% load on the integrated card, bottlenecking the whole process.
,.