Small note for around 8:34: The pieces of data striped across the array in a RAID0 configuration are typically a power of 2 bytes, not individual bits. A common strip size (the piece of data put onto each drive in a stripe) is 16KB, but you can usually adjust it to be bigger or smaller depending on your use case.
I've never heard the "independent" variant until now. I'm quite sure this is another case of enough people saying it wrong that they began accepting both. This was definitely a question on my A+ exam
The term "RAID" was invented by David Patterson, Garth A. Gibson, and Randy Katz at the University of California, Berkeley in 1987. In their June 1988 paper "A Case for Redundant Arrays of Inexpensive Disks (RAID)"...Industry manufacturers later redefined the RAID acronym to stand for "redundant array of independent disks" Wiki
Why you're confused with RAID5 results? Here's some points to consider. 1. To calculate parity it's required to send data to and from GPU, or calculate with CPU. It's a possible bottleneck. 2. Was our array in optimal state before tests? It should calculate parity data(process called build/rebuild/resync) before going to optimal state. That can limit performance too, and it explains 100% GPU usage without host I/O.
7:07 12v 7a, probably sounds like a fighter jet with the afterburners on, of COURSE it's made by Delta. Their little 60mm monsters and the sound they made is part of why we started getting 80mm and then 120mm CPU coolers.
@@jrcowboy1099 It's not a piece of silicon which almost entirely dissipates its power input as waste heat. It's an electric motor showing an efficiency above 95%. Of those 84W, maybe 5 are converted to heat, the rest moves air. A LOT of air.
GPU is going to start standing for "general processing unit" before long. They do physics, they do math and hashing, they do rendering and texture mapping, and now they are freaking hardware controllers too!
Yeah, like how GPUs being incredibly fast and having stupidly fast memory that the processor can directly access from less than an inch away happened to be perfect for running ML tasks at fast speeds.
it makes me really happy that these are exactly the same as GPUs, because that means that A: The software could run on ANY gpu, and B: you can use these cards AS FUCKING GPUS if needed, meaning they won't go to waste if you no longer need them for running a RAID with. No doubt you'll need some kind of hack because of the license thing, but yea that's wild lol
It would work the other way around. These are older GPUs. It's more likely that industry would repurpose existing but outdated GPUs to use for this application, instead of what you described.
@@ViciousTuna2012 god I hope so lol, then I might actually be able to access this technology within my lifetime. With how the setup is described, it makes me wonder just how much processing and bus bandwidth you really need for this. Presumably memory is a factor, but for simple user-oriented tasks, you legit do not need 100gb a second for *anything*. I wonder how this would interact with different solid state tech like Optane and optane hybrid. You might be able to blow the price, capacity, features, and performance of the top end raw pcie drives on the market using an old junker display processor, and 2-5 consumer optane hybrid m.2s. Of course adapters to fill that many slots may ramp up costs prohibitively. Pcie bifurcation and slot adaptation is way more expensive than it should be.
@@curvingfyre6810 I wouldn't think of this in terms of user-oriented speed requirements. This tech is going to be most useful for server applications, which will be able to utilize that level of bandwidth (100GB/s).
except he has some personality traits that are.. fairly annoying imo. hard to put words to it, but I'd have to disagree on the charisma part. he comes across as a smug know-it-all who likes to tell people how much he knows because he thinks it's impressive. unlike Linus who does it because he genuinely wants other people to learn. not always, just most of the time.
Lately Jake has been very good on camera, he got a lot better since 1-2 years ago when he was no good at all, to the point when I remember Linus telling the viewers and i am paraphrasing "I promise Jake is very good at what he does". Now we see he is knowledgeable and he manages to be entertaining too.
Its really neat when they do tests on stuff for their studio, that they ACTUALLY USE. Like the investment in the stuff isn't wasted, and the reaction is genuine to the performance to the point they they would definitely use/not use a thing is nice.
I'm pretty sure they often have to return some of the more sophisticated nicknacks they show off. They directly talked about it when they played around with that petabyte of nvmes.
I don't know any of the computing stuff that he was talking about in this episode, but I just enjoy watching linus get excited about stuff. It brightens my day.
Funny comment, but yeah enterprises pay way more for software licenses than for hardware, oracle, esxi, etc can easily become 100s of thousands or millions a year and are priced per socket or per core
pair of Epyc 75F3 cost over $12k, pair of Xeon 8380HL will drain you of 28k Euro. And you still can't buy 5950X for less than a kidney - if you find one, that is. Edit: It may depend on use case, but usually RAM becomes the biggest money sink, unless you are fine with dropping 64 GB in a dual socket server.
GPU accelerated raid, NVME drives attached with a GPU, PCIe 5 coming. I still say there could be some really interesting devices coming that could handle several different things at once, without performance issues. e.g. A single GPU, with 4 drives, in Raid, with the raid being taking care of by the GPU, and the GPU sending video output (perhaps some other computation happening as well). All without a performance being impacted with any of it.
You are underestimating the software complexity of these solutions. Anyone can duct tape together a bunch of components, but making them actually work together seamlessly is not trivial, and when something goes wrong, they are just too difficult to debug. If you look at the history of computing, simpler solutions tend to win. Only the most performance critical applications tolerate some complexity.
For a few minutes of this video, the server fan tone fit well with the key of the background music like a droning common tone. Enough that it distracted me while they were talking
LTT: We just had a massive storage failure where data checksums were critical to recovery. Also LTT: Let's switch to this new RAID system that has no data checksums!
Well the server that had the issues would have a completely different use case than this server. The failure was on a long term archival storage server using HDDs. This type of server would be used for short term storage of ongoing work on SSDs. So, what they are discussing here uses more reliable equipment in a situation where any error would almost certainly be detected extremely quickly and relies on high transfer speeds and much more expensive storage.
@@konzo5942 The nice thing about zfs (and btrfs) is that they can detect corruption within raid pools and AUTOMATICALLY CORRECT it. That probably won't work with this setup as everything is exposed as one single drive. You will know something failed, but not which copy is correct and you won't be able to restore it.
@@kommentator1157 a purely filesystem software that does checksumming is going to work to correct bit rot irrespective of the underlying hardware. Also it's less of an issue with SSDs which this is only going to be used for.
@@konzo5942 not really. This solution (like most raid only systems) presents the pool as a single logical disk. The filesystem could do checksumming and tell you something is wrong, but I doesn't have the necessary information to correct it. This solution only uses the parity data if there is a read error, regardless of whether the data from a particular drive is correct, because the raid abstraction does not utilize checksumming. Wendel from L1 techs was able to corrupt the data easily during his tests (by inserting a discrepancy on one drive to simulate the random changes to the saved values due to bitrot) and because the drive didn't report an error it was not detected. ZFS on the other hand caught and corrected the same scenario. I agree this is not a huge issue for short term live data, but if there are no protections, such discrepancies could find their way into the backups and archives because they were not detected/corrected
With the kind of airflow going through that chassis, would the RAID GPU card stay cooler if you took the shroud and fan off? Assuming the heatsink fins are oriented horizontally.
From experience, there is no difference. A lot of bus cards, FPGAs, expanders and other PCIe appliances come in with a fan, so they can be put in a tower desktop, mostly for development purposes, but in general when you put it inside a server (especially in 1U or half-width 1U on dual node systems) it just doesn't matter.
gotta remember that it's a positive pressure system in that thing so it's forcing air though those fins even with the integrated fan turned off. when it's on it's more boosting the speed over the server's insane baseline by helping air maintain speed. it's only ever going to be a single digit percentage diffrence
The part of this whole concept that I find most interesting is the fact that someone basically found a useful way to repurpose old GPUs that would otherwise be eWaste. Someone that’s into farming Chia might find something like this useful as well, so there’s still a way to earn some crypto with it.
Surprised to hear them talking so positively about RAID 5 and dismissing RAID 10 as "expensive". Yes, you do lose half your capacity with RAID 10, but you get more performance, more resiliency, and importantly rebuilds are faster / less resource intensive. Rebuilding your array from a failed disk doesn't matter so much when all the drives are new, but when they're all end of life, the extra strain from a RAID 5 rebuild can cause a chain of failures, potentially causing a totally array loss. Coming from industry, I'd never deploy a RAID 5 array again, RAID 10 all the way. RAID 1 also has some good use cases.
the extra stress for raid5-6 etc while rebuilding, is more of a problem for the mechanical hard drives. ssd's, especially of that class, don't suffer the same drawbacks, only write wear matters, reads don't affect them. also due to much higher transfer speed, the rebuild will complete way faster. still, with a bit of patience and thinking, you can use a combination of different raid levels, that will fill the needs better...
Seems like an exciting step in making alternative hardware accelerators common place in systems. It's wild to think that despite how far innovation has pushed the capabilities of modern gpus in terms of video encoding and rasterization, they are but one manifestation of hardware accelerators in a wide field with new opportunities coming from things like fpgas or asics.
It's interesting because this is much like an echo of the past. In the 90s we had co-processors and such. Sound cards were also hip until everyone had their baked-in Realtek chip. We kinda went backwards and are now seeing the start of a boom of new accelerators, possibly due to the progress of performance improvement slowing down.
@@Naetrox Good point, it doesn’t seem as viable to just throw more general compute at a problem. Instead better specified architecture provides much more real performance. I wonder if someday the almighty cpu might become a compute analog to a clock generator.
@@uncle-ff7jq I'm holding out for quantum computers, I think we'll see adoption of them in a few decades as we figure out how to operate them at room temperature.
You should run some tests with degraded RAID5, see what really happens with CPU utilization at that point. Expect it'll still offload to the GPU. Or a fully degraded RAID6 (two disks failed)
Gotta love the knowledge and expertise of Jake and the videos he does where Linus is a spectator. Jake’s skills are impressive, he seems to know everything!
I bought relatively cheap Kioxia SSD this month - Toshiba NAND and 4GB DRAM controller from Samsung. it was even cheaper than a lot of low end drives from Chinese manufacturers without DRAM. Its performance is pretty good, but I'm not much of storage purist.
This is like the opposite of ZFS in every way. ZFS is like "ecc for resiliency on top of hashing everything"! Graid be like "speed!!!!". It'd be interesting to see how much data corruption would happen on one of those in a real world heavy write workload.
While I agree, you response may lead a novice reader to mistakenly believe ZFS protects against corruption, when infact it an be the opposite. - The hash in ZFS detects corruption, can't actually be used to recover data - ZFS has next to zero protection against bitrot - ZFS can't detect errors introduced by the host, so ECC is critical at all layers, RAM. ie: Don't run ECC on cheap hardware, overclock ect.. as errors introduced on the host will get written to the ZFS and NOT detected. ZFS is perfect to host VMs but the ZFS server itself should never be virtual due to higher risk of corruption. So while ZFS is about as good as you can get in the enterprise, due to the extreme robustness, ease of recovery and tons of features. Running it on consumer hardware introduces a lot more risk over other filesystems and will generally perform poorly without a crap ton of memory.
@@timsmith3743 ZFSs protection against bitrot is literally a focus in it's initial design. If your pool is using raidz, ZFS not only protects from bit rot, it eliminates it in real time. The hash that ZFS stores with every write, and the parities of your pool config allows ZFS to determine which version of data is correct and only return that data. Not only does ZFS protect against bit rot, it also protects against file corruption. ECC is recommended for ZFS systems because it's the only part of the system that is not protected against corruption. ECC protects the data while it is processed by ZFS.
@@gamebrigada2 , ZFS detects bit rot, but can only perform the bitflip if a replica exists. To understand your risk, you need to review the results of the scheduled scrubs and often need to replace the lost data where it can not be repaired automatically.
My prediction for the future- this type of card will quickly be obsolete and become integrated into CPUs or motherboards. This is a PhyX card for storage. They might continue to exist for servers though.
Perhaps, but it's using GPU Compute at 100% and I don't think it'll be in CPUs anytime soon since it seems to be very much an enterprise solution. It might be compared to like a PhysX card but this is more akin to using a dedicated GPU for tensor applications.
"This type of card" is just a GPU. There doesn't seem to be any special fixed-function hardware or special firmware as explained, It's just offloading what the CPU would do onto what's effectively a GTX 1650 with the TDP pared back to fit under a smaller cooler. Integrating it into CPUs would just be doing GPU compute on the iGPU. For home users, the processor overhead isn't that high anyway. For servers there isn't an iGPU to offload on, and even if they did add one, it would be fighting the main CPU for die space and power budget where an add-in card would not.
@@jameslake7775 chiplets change this equation slightly. AMD could put a dedicated RAID accelerator on the I/O die of Epyc without affecting the rest of the products.
@@Prophes0r I don't think that hardware getting more capable and more general purpose is ever a bad thing. If a TON of people start using features like this in the short run it could affect the supply chain. But honestly I don't think there'll be enough people using this that it'll affect anyone's ability to get a GPU. And honestly it just means your card you have can do more things. Why would that be a bad thing?
@@nicekeyboardalan6972 Aight rich kid, just know if war comes, you'll be first to get eaten. That said, expensive GPUs make for half decent impromptu weapons. See you on the battlefield, come sit down for some soup and bread with us.
The main interest of RAID for hard drives is the ability to hot swap any faulty drive on a running server. Could you test hot swap in this beefy configuration ?
Amazes me nobody knows about the T400, low profile, full Ampere NVenc/dec in effectively the 1630 that never happened for $150 US, in this market its a great card for HTPC/Encode servers/plex/streaming etc etc
This was such a cool concept and video!I would love to see the open source community come up with their own solution of this! Please see what else you can do with this, and I await the videos that will relate to it!
IBM and Nvidia are working on a much better version of this concept, enables a GPU to connect directly to a computer's SSD to bypass the CPU completely.
This system is completely a different use case to what Nvidia recently announced with IBM. What Nvidia-IBM are making are something akin to DirectStorage where pulling dataset blocks from storage for number crunching/training are quite slow going via the CPU so like MSDS is cutting out the middle man. What this GRAID system is doing is using a relatively inexpensive workstation card (stability etc) to be the write target for the disk are so it can use the massive amount of mathematical capability of the GPU cores to calculate parity data which is pure math of block A + block B = block C and if you were to lose either A or B you could get the missing data back because you have 2 parts of the 3 sided equation. In short, it's just using the Nvidia card as a write accelerator which is pretty impressive and am also a little surprised this hasn't been posited before now.
Love the video. While I've never ran anything greater than Raid 1, between your channel and others I have been running Unraid for almost 2 years. Yes I know there are pro's and con's of both, but at this time Unraid has been great for me needs.
Wow, on a RAID 5 getting 30+GB/s with barely any CPU usage is insane! This card is definitely going to help us all in the DC! Edit: 6.5M IOPS?! Whooaaa
But hardware raid has one huge, show-stopping (at least for me) drawback: If the RAID *controller* goes out, your need the EXACT same chipset and version, for the most part, in order to access your drives again. Any deviation and it will not see the data, unless things have changed significantly in recent times. Software raid wouldnt suffer from this, since if the hardware (or windows/linux) host dies, you can just stick the drives in another computer and as long as you install the same software, it should work.
This seems very solvable. This solution seems to just be a cuda-capable software raid, only bundled with a gpu it's been validated with. Barring the license, you should be able to swap the hardware for anything you want if it's capable in regard to cuda versions
The disk load should ideally be created on another system, as it's likely a decent % is used by the benchmark took, to produce the output. As performed, you can't estimate the CPU impact.
That's absolutely insane, and the fact it really is just a GPU is even more insane. Only a matter of time before folks are working around the license & just running this sorta setup with whatever their last GPU was. I'd love to finally build a a new system & then give my now old 1080TI a new lease on life as a raid controller for a home server to hold all of my own footage & assets.
Not going to lie, I found Jake incredibly annoying when he first started. But now the videos with Jake and Linus are my favorite videos. Those two have found a really nice chemistry.
Agreed. I went back to a few older videos just to see what changed and it seems that he's definitely mellowed out a lot. They've had this "close friends" type relationship for awhile but I think Jake might've been overdoing it for the camera back then. Seems like he's a lot more comfortable now.
Meh, I'll take ZFS with NVMe drives instead. Plenty of performance, low CPU overhead and excellent data integrity. TrueNAS would work just fine on that server.. Pull the power cord out while you're running that benchmark, let's see how it handles recovery.
@@chedatomasz I see. I'd assume the (probably) massive differences in whatever the hardware equivalent of an API is might make it a bit more difficult than a simple lock removal, though.
A modern filesystem like ZFS has its own RAID-like function and it's advised that you don't use any RAID card and expose the drives directly to the filesystem. And from the video, it seems it has LVM-like function.
It's crazy how even a little bit of GPU acceleration can impact a "high throughput" application like this. Frankly, you'll eventually be bottlenecked by DDR4 memory bandwidth. HBM CPUs anyone?
Octa-channel ddr5 will be interesting for bandwidth. I would also like to see some HBM3 CPUs. It wouldn't be out of the question for somebody to cram 16 or 32gb on a package and still have lot of space for a CPU left in there.
"redundant array of independent disks" has a k in disks, not a c, as in "Hard Disks". Disc with a "c" is used to refer to optical media like Compact Discs or DVDs.
Small note for around 8:34: The pieces of data striped across the array in a RAID0 configuration are typically a power of 2 bytes, not individual bits. A common strip size (the piece of data put onto each drive in a stripe) is 16KB, but you can usually adjust it to be bigger or smaller depending on your use case.
First, love you LTT
Ok
Now how am I supposed to “Well Achtually” correct people in the comments??
I have to know. Does he actually accidentally drop stuff or does he do it for show these days
Were is scrapyard wars
I want to believe Linus dropping stuff is a bit now, but deep down, I know it’s real.
seriously lol.
Jesus Christ please stop the cringe
Honestly it's both
Linus is just like me -> clumsy 😆
linus is a bit energetic
Linus almost dropping it in the first two minutes, just classic.
more like first 30 seconds. It's Linus we're talking about.
lol linus is what we're talking about now?
Just business as usual
Linus has worse hand/eye coordination than my grandma who has Parkinsons.
Linus dropping things and his lame segues into into sponsor and LTT Store shilling. Name a more iconic duo.
Fun facts: RAID originally meant "Redundant Array of Inexpensive Disks", but "inexpensive" didn't sell as well as "independant"
I've never heard the "independent" variant until now. I'm quite sure this is another case of enough people saying it wrong that they began accepting both. This was definitely a question on my A+ exam
I've never heard of inexpensive tbh
The term "RAID" was invented by David Patterson, Garth A. Gibson, and Randy Katz at the University of California, Berkeley in 1987. In their June 1988 paper "A Case for Redundant Arrays of Inexpensive Disks (RAID)"...Industry manufacturers later redefined the RAID acronym to stand for "redundant array of independent disks"
Wiki
And LTT messed it up by writing "Discs"
Thankyou! I had a moment where I doubted myself as I thought have I been saying it wrong all this time? Turns out I'm just old instead.
Why you're confused with RAID5 results? Here's some points to consider.
1. To calculate parity it's required to send data to and from GPU, or calculate with CPU. It's a possible bottleneck.
2. Was our array in optimal state before tests? It should calculate parity data(process called build/rebuild/resync) before going to optimal state. That can limit performance too, and it explains 100% GPU usage without host I/O.
7:07 12v 7a, probably sounds like a fighter jet with the afterburners on, of COURSE it's made by Delta. Their little 60mm monsters and the sound they made is part of why we started getting 80mm and then 120mm CPU coolers.
HOLY... 84W?!?! WOW! It need a lot of that power to cool itself!
@@jrcowboy1099 It's not a piece of silicon which almost entirely dissipates its power input as waste heat. It's an electric motor showing an efficiency above 95%. Of those 84W, maybe 5 are converted to heat, the rest moves air. A LOT of air.
@@ujiltromm7358 Lol, yeah you're right, I guess I was apparently brain dead when I wrote that.
GPU is going to start standing for "general processing unit" before long. They do physics, they do math and hashing, they do rendering and texture mapping, and now they are freaking hardware controllers too!
You know what? Thats pretty cool
Yeah, like how GPUs being incredibly fast and having stupidly fast memory that the processor can directly access from less than an inch away happened to be perfect for running ML tasks at fast speeds.
Applications beyond graphics....now the language is "accelerator" as GPU seems outdated/limited usage.
That's kind of what Xeon Phi was for. Parallel processing taken to the max. I cast my vote for PPU
for some reason i read harrasing and hashing...
As someone who works in enterprise tech, I enjoy these videos significantly more than 70+ slide power points from our engineers.
Oh you poor soul. If there’s nothing corps love more than profit it’s PowerPoints.
The vibe you guys were in during the filming of this episode made it really enjoying to watch. more of this!
it makes me really happy that these are exactly the same as GPUs, because that means that A: The software could run on ANY gpu, and B: you can use these cards AS FUCKING GPUS if needed, meaning they won't go to waste if you no longer need them for running a RAID with. No doubt you'll need some kind of hack because of the license thing, but yea that's wild lol
I wonder if you could use it as a basic display output at the same time.
It would work the other way around. These are older GPUs. It's more likely that industry would repurpose existing but outdated GPUs to use for this application, instead of what you described.
Mining only GPUs perhaps. Nvidia's thing is CUDA, I expect if you can run CUDA code then you can run this RAID thing.
@@ViciousTuna2012 god I hope so lol, then I might actually be able to access this technology within my lifetime. With how the setup is described, it makes me wonder just how much processing and bus bandwidth you really need for this. Presumably memory is a factor, but for simple user-oriented tasks, you legit do not need 100gb a second for *anything*. I wonder how this would interact with different solid state tech like Optane and optane hybrid. You might be able to blow the price, capacity, features, and performance of the top end raw pcie drives on the market using an old junker display processor, and 2-5 consumer optane hybrid m.2s. Of course adapters to fill that many slots may ramp up costs prohibitively. Pcie bifurcation and slot adaptation is way more expensive than it should be.
@@curvingfyre6810 I wouldn't think of this in terms of user-oriented speed requirements. This tech is going to be most useful for server applications, which will be able to utilize that level of bandwidth (100GB/s).
I remember when a "T1000" was something else completely different. And deadly. But not good with extreme temperatures...
That's a nice Terminator reference you've got there👌
Wonder if sky net used these
@@nathanlowery1141 Skynet is a LTE internet service provided by Skytel telecommunications company in Mongolia.
You just made me laugh harder than I should, office mates looking at me having a geek attack
He'll be back
Killing the videos lately, been great content, love Alex and Jake's hosting/co-hosting.
Did they recently stop using the intros though? I could’ve sworn I hadn’t seen it on another recent upload.
Right? These two together are electric
I like the videos where Jake does cool stuff and Linus just watches and reacts.
They have great on camera chemistry.
linus: with these enterprise drives, they're not likely to fail
jake: they usually all fail at the same time
:D
Jake really has a similar combination of knowledge and charisma that made Linus so popular and enjoyable to watch.
except he has some personality traits that are.. fairly annoying imo. hard to put words to it, but I'd have to disagree on the charisma part. he comes across as a smug know-it-all who likes to tell people how much he knows because he thinks it's impressive. unlike Linus who does it because he genuinely wants other people to learn. not always, just most of the time.
Lately Jake has been very good on camera, he got a lot better since 1-2 years ago when he was no good at all, to the point when I remember Linus telling the viewers and i am paraphrasing "I promise Jake is very good at what he does". Now we see he is knowledgeable and he manages to be entertaining too.
he would be even better on camera if he eliminated the pedostache
@@Shadowcast140 your momma loves it though
He's finally filling the role that Luke had on camera.
I remember the "Jake needs a sack" phase but he's massively stepped up his game, it's amazing to see!
I'd love to see him healthy!
Its really neat when they do tests on stuff for their studio, that they ACTUALLY USE. Like the investment in the stuff isn't wasted, and the reaction is genuine to the performance to the point they they would definitely use/not use a thing is nice.
I'm pretty sure they often have to return some of the more sophisticated nicknacks they show off. They directly talked about it when they played around with that petabyte of nvmes.
Linus said on the WAN show they would have to return most of the stuff for this project
Please read my name💕
Yeah, i know right? like when Linus was so happy with the Framework laptop that he fraking bought it on cammera lol
As a tech geek it always makes me smile how giddy they are when talking about extraordinary hardware / software solutions. 😄
I don't know any of the computing stuff that he was talking about in this episode, but I just enjoy watching linus get excited about stuff. It brightens my day.
That wardrobe change though. Super high production quality!
Linus says "CPUs are fricking expensive" in 2022 while literally waving a GPU in his hand. The audacity
Haha
we are all clear that he speaks about CPUs that are like 2 3090ti with enough money left to buy a descend car second hand.
@@satakrionkryptomortis yup. And licenses are significant in VM infrastructure cost measured per CPU core.
Funny comment, but yeah enterprises pay way more for software licenses than for hardware, oracle, esxi, etc can easily become 100s of thousands or millions a year and are priced per socket or per core
pair of Epyc 75F3 cost over $12k, pair of Xeon 8380HL will drain you of 28k Euro. And you still can't buy 5950X for less than a kidney - if you find one, that is.
Edit: It may depend on use case, but usually RAM becomes the biggest money sink, unless you are fine with dropping 64 GB in a dual socket server.
30 secs in and Linus's already dropping the thing. Legend.
What?
@@bustex1 what do you mean "what?" just read it again
@@BillyBonesYT read it and watched the first 30 seconds five times. Happened at like 34 second mark not first 30 seconds. What.
He's just too good. Somebody stop this man
Clumsy=Linus
GPU accelerated raid, NVME drives attached with a GPU, PCIe 5 coming.
I still say there could be some really interesting devices coming that could handle several different things at once, without performance issues.
e.g. A single GPU, with 4 drives, in Raid, with the raid being taking care of by the GPU, and the GPU sending video output (perhaps some other computation happening as well). All without a performance being impacted with any of it.
It's like a return to the past when some Enterprise GPUs had inbuilt SSDs.
Certainly not common, but not a new idea.
Back to the days of a math coprocessor
Please read my name❣️
Are we going to see less of a push for improvements to the CPU in the future then? maybe that's why Intel are moving into the GPU space.
You are underestimating the software complexity of these solutions. Anyone can duct tape together a bunch of components, but making them actually work together seamlessly is not trivial, and when something goes wrong, they are just too difficult to debug. If you look at the history of computing, simpler solutions tend to win. Only the most performance critical applications tolerate some complexity.
You just gotta love Jake's laugh - even if he remembers me to the grumpy Jake of Two and a half men (8. Seasons). 🤭
For a few minutes of this video, the server fan tone fit well with the key of the background music like a droning common tone. Enough that it distracted me while they were talking
I kinda had a feeling this man was able to look into the future.
Yep
Why does he keep dropping stuff then
@@WayStedYou Dropping stuff is required to look into the future
That's the downside of being able to look into the future
I wonder if this would be possible to do with any GPU with custom firmware. Would be a cool purpose for spare GPUs...
In b4 nvidia makes this a software feature and obsoletes this thing. Probably doesn't even need to buy them to kill them.
Spare gpu's.
HOW DARE YOUUUUUUU
@@firefly2472 I'm sure they're talking about 10 year old cards
@@doggo_woo I know :)
Have a few myself.
And no, not gonna sell them :p
@@firefly2472 Gotta collect em all lol
LTT: We just had a massive storage failure where data checksums were critical to recovery.
Also LTT: Let's switch to this new RAID system that has no data checksums!
Well the server that had the issues would have a completely different use case than this server.
The failure was on a long term archival storage server using HDDs. This type of server would be used for short term storage of ongoing work on SSDs.
So, what they are discussing here uses more reliable equipment in a situation where any error would almost certainly be detected extremely quickly and relies on high transfer speeds and much more expensive storage.
this is block level only, so use a checksumming filesystem? a bit unfair to compare a filesystem to a raid only software.
@@konzo5942 The nice thing about zfs (and btrfs) is that they can detect corruption within raid pools and AUTOMATICALLY CORRECT it. That probably won't work with this setup as everything is exposed as one single drive. You will know something failed, but not which copy is correct and you won't be able to restore it.
@@kommentator1157 a purely filesystem software that does checksumming is going to work to correct bit rot irrespective of the underlying hardware. Also it's less of an issue with SSDs which this is only going to be used for.
@@konzo5942 not really. This solution (like most raid only systems) presents the pool as a single logical disk. The filesystem could do checksumming and tell you something is wrong, but I doesn't have the necessary information to correct it. This solution only uses the parity data if there is a read error, regardless of whether the data from a particular drive is correct, because the raid abstraction does not utilize checksumming. Wendel from L1 techs was able to corrupt the data easily during his tests (by inserting a discrepancy on one drive to simulate the random changes to the saved values due to bitrot) and because the drive didn't report an error it was not detected. ZFS on the other hand caught and corrected the same scenario. I agree this is not a huge issue for short term live data, but if there are no protections, such discrepancies could find their way into the backups and archives because they were not detected/corrected
Good job you guys getting a telus sponsor thats pretty big
The Server Rig in the background
Ready for Take of
😂
That's actually really cool. I'd love to see a 3090 or that super top of the line A1000 run the software raid stuff.
Yeahhhh gonna need you guys to either spoof or secure a license for that big boy card to see what would happen. That would be insane lol
With the kind of airflow going through that chassis, would the RAID GPU card stay cooler if you took the shroud and fan off? Assuming the heatsink fins are oriented horizontally.
From experience, there is no difference. A lot of bus cards, FPGAs, expanders and other PCIe appliances come in with a fan, so they can be put in a tower desktop, mostly for development purposes, but in general when you put it inside a server (especially in 1U or half-width 1U on dual node systems) it just doesn't matter.
gotta remember that it's a positive pressure system in that thing so it's forcing air though those fins even with the integrated fan turned off. when it's on it's more boosting the speed over the server's insane baseline by helping air maintain speed. it's only ever going to be a single digit percentage diffrence
Informative answers from both of you, thanks. I simply don't have rackmount server experience, and am happy to learn from those who do.
I love how that G710+ has hung around for so long. I bought mine like a week before your video covering it came out.
Linus, you are an insider and the number one Tech channel, of course you knew...
The part of this whole concept that I find most interesting is the fact that someone basically found a useful way to repurpose old GPUs that would otherwise be eWaste. Someone that’s into farming Chia might find something like this useful as well, so there’s still a way to earn some crypto with it.
Linus now: On these enterprise level devices it should be fine.
5 years later: Yeah guys it failed and we didn’t see it through
5 years of INCREDIBLY heavy, daily use.
And no way to detect bit rot, unlike ZFS. Having backup of bit rot is kinda pointless. :P
Surprised to hear them talking so positively about RAID 5 and dismissing RAID 10 as "expensive". Yes, you do lose half your capacity with RAID 10, but you get more performance, more resiliency, and importantly rebuilds are faster / less resource intensive. Rebuilding your array from a failed disk doesn't matter so much when all the drives are new, but when they're all end of life, the extra strain from a RAID 5 rebuild can cause a chain of failures, potentially causing a totally array loss. Coming from industry, I'd never deploy a RAID 5 array again, RAID 10 all the way. RAID 1 also has some good use cases.
Rebuilding an array is the worst part of a process when you do get a failure.
How do you feel about raid 6?
He did specify that there are use cases for RAID 10, but it was severe overkill for the hardware/use case they were considering.
the extra stress for raid5-6 etc while rebuilding, is more of a problem for the mechanical hard drives. ssd's, especially of that class, don't suffer the same drawbacks, only write wear matters, reads don't affect them. also due to much higher transfer speed, the rebuild will complete way faster. still, with a bit of patience and thinking, you can use a combination of different raid levels, that will fill the needs better...
I prefer RAID 50 if you wanna go with the mirroring route.
Holy shit is exactly what I said right before he said I was going to say that. So Linus, you predicted the future twice in this video 😂
Man I was not expecting telus as the sponsor! That's a really big company
RAID originally stood for "Redundant Arrays of Inexpensive Disks"
My experience has been "Risking All Important Data"...
@@majorbogart3476 thats the lack of raid shadow legend..
And it still means that if you compare any kind of RAID to a dedicated, proprietary storage machine in a datacenter.
@@satakrionkryptomortis you could have not said that but you did it anyway
@@majorbogart3476 yea, if RAID 0 and not any other form
I love that Linus is still mega hyped about tech even after all these years 😀
Seems like an exciting step in making alternative hardware accelerators common place in systems. It's wild to think that despite how far innovation has pushed the capabilities of modern gpus in terms of video encoding and rasterization, they are but one manifestation of hardware accelerators in a wide field with new opportunities coming from things like fpgas or asics.
It's interesting because this is much like an echo of the past. In the 90s we had co-processors and such. Sound cards were also hip until everyone had their baked-in Realtek chip. We kinda went backwards and are now seeing the start of a boom of new accelerators, possibly due to the progress of performance improvement slowing down.
@@Naetrox Good point, it doesn’t seem as viable to just throw more general compute at a problem. Instead better specified architecture provides much more real performance. I wonder if someday the almighty cpu might become a compute analog to a clock generator.
@@uncle-ff7jq I'm holding out for quantum computers, I think we'll see adoption of them in a few decades as we figure out how to operate them at room temperature.
It's nice to see the Level 1 Techs merch on display.
You should run some tests with degraded RAID5, see what really happens with CPU utilization at that point. Expect it'll still offload to the GPU. Or a fully degraded RAID6 (two disks failed)
Based on the diagram, in that read does not go through the GPU. I would expect it to hit the CPU. That's a great question.
Two minute's in and linus is already dropping thing's. Absolute legend.
Even less than 2 minutes; more like 40 seconds
Gotta love the knowledge and expertise of Jake and the videos he does where Linus is a spectator. Jake’s skills are impressive, he seems to know everything!
I bought relatively cheap Kioxia SSD this month - Toshiba NAND and 4GB DRAM controller from Samsung. it was even cheaper than a lot of low end drives from Chinese manufacturers without DRAM. Its performance is pretty good, but I'm not much of storage purist.
how much did you pay?
what drive?
These are my favorite types of videos on this channel. Crazy stuff, thanks
For the Kioxia BG5, I'm surprised you didn't mention its application in the steam deck, as it uses M.2-2230 SSDs as well.
I wish I could find one too. Seems like everywhere I look they’re still unavailable
This is nuts. I know I'll never get to work on these kinds of projects, but man I'd love to. I have no need for it, I just want it lol Amazing stuff
This is like the opposite of ZFS in every way. ZFS is like "ecc for resiliency on top of hashing everything"! Graid be like "speed!!!!". It'd be interesting to see how much data corruption would happen on one of those in a real world heavy write workload.
While I agree, you response may lead a novice reader to mistakenly believe ZFS protects against corruption, when infact it an be the opposite.
- The hash in ZFS detects corruption, can't actually be used to recover data
- ZFS has next to zero protection against bitrot
- ZFS can't detect errors introduced by the host, so ECC is critical at all layers, RAM. ie: Don't run ECC on cheap hardware, overclock ect.. as errors introduced on the host will get written to the ZFS and NOT detected.
ZFS is perfect to host VMs but the ZFS server itself should never be virtual due to higher risk of corruption.
So while ZFS is about as good as you can get in the enterprise, due to the extreme robustness, ease of recovery and tons of features. Running it on consumer hardware introduces a lot more risk over other filesystems and will generally perform poorly without a crap ton of memory.
@@timsmith3743 ZFSs protection against bitrot is literally a focus in it's initial design. If your pool is using raidz, ZFS not only protects from bit rot, it eliminates it in real time. The hash that ZFS stores with every write, and the parities of your pool config allows ZFS to determine which version of data is correct and only return that data. Not only does ZFS protect against bit rot, it also protects against file corruption.
ECC is recommended for ZFS systems because it's the only part of the system that is not protected against corruption. ECC protects the data while it is processed by ZFS.
@@gamebrigada2 , ZFS detects bit rot, but can only perform the bitflip if a replica exists. To understand your risk, you need to review the results of the scheduled scrubs and often need to replace the lost data where it can not be repaired automatically.
My prediction for the future- this type of card will quickly be obsolete and become integrated into CPUs or motherboards. This is a PhyX card for storage. They might continue to exist for servers though.
Perhaps, but it's using GPU Compute at 100% and I don't think it'll be in CPUs anytime soon since it seems to be very much an enterprise solution. It might be compared to like a PhysX card but this is more akin to using a dedicated GPU for tensor applications.
If there is a market for this, AMD and Intel will integrate it into their CPUs in a more refined way.
i can see RISC type chips made specifically for this task becoming popular.
"This type of card" is just a GPU. There doesn't seem to be any special fixed-function hardware or special firmware as explained, It's just offloading what the CPU would do onto what's effectively a GTX 1650 with the TDP pared back to fit under a smaller cooler. Integrating it into CPUs would just be doing GPU compute on the iGPU.
For home users, the processor overhead isn't that high anyway. For servers there isn't an iGPU to offload on, and even if they did add one, it would be fighting the main CPU for die space and power budget where an add-in card would not.
@@jameslake7775 chiplets change this equation slightly. AMD could put a dedicated RAID accelerator on the I/O die of Epyc without affecting the rest of the products.
I don't know crap about what they are talking about but I really enjoyed seeing how pumped they were about what was going on.
I just bumped into one today and I can't believe that I am learning more about it today(all in the same day)
I am looking forward to the day someone is releasing an open source driver accelerating storage with any GPU.
This would be awesome
@@Prophes0r I don't think that hardware getting more capable and more general purpose is ever a bad thing. If a TON of people start using features like this in the short run it could affect the supply chain. But honestly I don't think there'll be enough people using this that it'll affect anyone's ability to get a GPU. And honestly it just means your card you have can do more things. Why would that be a bad thing?
@@nicekeyboardalan6972 Aight rich kid, just know if war comes, you'll be first to get eaten. That said, expensive GPUs make for half decent impromptu weapons. See you on the battlefield, come sit down for some soup and bread with us.
This thing is sick. Could be a small revolution in scientific computing.
The main interest of RAID for hard drives is the ability to hot swap any faulty drive on a running server.
Could you test hot swap in this beefy configuration ?
i love whenever you guys work on server stuff. just absolutely top-tier stuff
that hair in your eye at 4:15 man that would have drove me kitchen roll bonkers
The T400/600/1000 line of GPUs are underrated gems. Tiny little powerhouses in an otherwise completely screwed GPU market!
Amazes me nobody knows about the T400, low profile, full Ampere NVenc/dec in effectively the 1630 that never happened for $150 US, in this market its a great card for HTPC/Encode servers/plex/streaming etc etc
This was such a cool concept and video!I would love to see the open source community come up with their own solution of this! Please see what else you can do with this, and I await the videos that will relate to it!
Hi @Hobbes, there is the mergers + snapraid combo of you want open source projects. That's what I use in my OpenMediaVault server.
This!
"Compared to the documentation of anything Microsoft" I felt that Jake, I felt that
its not gonna be a LTT video without linus actually or almost dropping stuff... its kind of a tradition now
nice F1 simulation at 18:10
0:34 Linus drop this strikes again
IBM and Nvidia are working on a much better version of this concept, enables a GPU to connect directly to a computer's SSD to bypass the CPU completely.
@@marcogenovesi8570 Will this be open sourced like IBM/Nvidia plan to do with what they are doing?
@@karehaqt does Nvidia have any open source products?
This system is completely a different use case to what Nvidia recently announced with IBM. What Nvidia-IBM are making are something akin to DirectStorage where pulling dataset blocks from storage for number crunching/training are quite slow going via the CPU so like MSDS is cutting out the middle man.
What this GRAID system is doing is using a relatively inexpensive workstation card (stability etc) to be the write target for the disk are so it can use the massive amount of mathematical capability of the GPU cores to calculate parity data which is pure math of block A + block B = block C and if you were to lose either A or B you could get the missing data back because you have 2 parts of the 3 sided equation.
In short, it's just using the Nvidia card as a write accelerator which is pretty impressive and am also a little surprised this hasn't been posited before now.
@@marcogenovesi8570 The software is what I was implying, not the hardware.
We need a continuation making it work with other GPUs
I think that telus ad was the first canadian targetted ad I've seen in a long time on this channel
This is super exciting. Its awesome where storage is right now!
We didn't even have to wait a minute before Linus almost dropped it.
"This server from Gigabyte handles all that through a backplane here in the front of the chassis."
Wouldn't that be a "frontplane" then? Har Har.
It's in the back of the front.
You see, David, this is why your mom left
Hello and welcome to another fine installment of Linus loves ridiculously awesome storage solutions.
I'm not surprised, by how catching things throughout years, Sebastian became sort of master of catching things. That 🐈 agility xD
Love the video. While I've never ran anything greater than Raid 1, between your channel and others I have been running Unraid for almost 2 years. Yes I know there are pro's and con's of both, but at this time Unraid has been great for me needs.
Wow, on a RAID 5 getting 30+GB/s with barely any CPU usage is insane! This card is definitely going to help us all in the DC! Edit: 6.5M IOPS?! Whooaaa
But hardware raid has one huge, show-stopping (at least for me) drawback: If the RAID *controller* goes out, your need the EXACT same chipset and version, for the most part, in order to access your drives again. Any deviation and it will not see the data, unless things have changed significantly in recent times. Software raid wouldnt suffer from this, since if the hardware (or windows/linux) host dies, you can just stick the drives in another computer and as long as you install the same software, it should work.
This seems very solvable. This solution seems to just be a cuda-capable software raid, only bundled with a gpu it's been validated with. Barring the license, you should be able to swap the hardware for anything you want if it's capable in regard to cuda versions
The disk load should ideally be created on another system, as it's likely a decent % is used by the benchmark took, to produce the output. As performed, you can't estimate the CPU impact.
That's absolutely insane, and the fact it really is just a GPU is even more insane. Only a matter of time before folks are working around the license & just running this sorta setup with whatever their last GPU was. I'd love to finally build a a new system & then give my now old 1080TI a new lease on life as a raid controller for a home server to hold all of my own footage & assets.
I love these server tech videos! Keep them coming guys!!!
Not going to lie, I found Jake incredibly annoying when he first started. But now the videos with Jake and Linus are my favorite videos. Those two have found a really nice chemistry.
yeah they kinda cute ngl 🥴
Agreed. I went back to a few older videos just to see what changed and it seems that he's definitely mellowed out a lot. They've had this "close friends" type relationship for awhile but I think Jake might've been overdoing it for the camera back then. Seems like he's a lot more comfortable now.
@@sune9578 gay erasure
@@chocolate_maned_wolf Excuse me. I think it is you that is doing the erasing of the gay.
@@sune9578 ok boomer 🤓
Meh, I'll take ZFS with NVMe drives instead. Plenty of performance, low CPU overhead and excellent data integrity. TrueNAS would work just fine on that server.. Pull the power cord out while you're running that benchmark, let's see how it handles recovery.
Yes Linus, you indeed called it...
Loved the more technical aspects of this video :D
Man, the noise at 18:03! I don't miss the server room noise which this reminded me of. Geez, its really working!
I kind of really want to see this on something like an RTX 3090. Probably wouldn't work, but imagine if it did.
3090 is cuda capable and afaik has a compatible cuda version, so barring the device version lock it should work.
@@chedatomasz I see. I'd assume the (probably) massive differences in whatever the hardware equivalent of an API is might make it a bit more difficult than a simple lock removal, though.
A modern filesystem like ZFS has its own RAID-like function and it's advised that you don't use any RAID card and expose the drives directly to the filesystem.
And from the video, it seems it has LVM-like function.
It's crazy how even a little bit of GPU acceleration can impact a "high throughput" application like this.
Frankly, you'll eventually be bottlenecked by DDR4 memory bandwidth.
HBM CPUs anyone?
We already are bottlenecked by it tbh.
Octa-channel ddr5 will be interesting for bandwidth. I would also like to see some HBM3 CPUs. It wouldn't be out of the question for somebody to cram 16 or 32gb on a package and still have lot of space for a CPU left in there.
Wow I haven't heard of such a thing in years
Did he just wink at the camera after that awesome recovery catch?! Haha Linus you're freaking awesome!
"This only runs on Linux server operating systems"
Me: Yes!
"So we are running Ubuntu Server 20.04 LTS"
Me: Ubuntu? Noooooo
I've heard tape drives are still around as a backup because while slow, they are super cheap. Would you guys ever do a video on those?
With more and more people running network storage I'm surprised tape drive backups are not more common.
They made a video on that years ago.
"redundant array of independent disks" has a k in disks, not a c, as in "Hard Disks". Disc with a "c" is used to refer to optical media like Compact Discs or DVDs.
Thank you. I was hoping someone else noticed that.
Love your videos I'm addicted to them love the t-shirt swap at 15:30 on timeline 🤣
Jake's shirt! Sysadmin haiku! Brilliant!
Welcome to another episode of things I won't buy as I can't afford thanks to Nvidia.
I just got a glitch in the matrix. All the comments disappeared.
Anyone taking bets on how long it'll take for an open source version of this software to show up?
Im exited to see the next video on this.
Very exciting results
I like that this thing sounds like it’s taking off.