HBM memory. The ability to extract more performance out of a Vega 56 by flashing the Vega 64 VBIOS on it. The very existence of Radeon VII. The RX Vega cards were some of the quirkiest mainstream GPUs ever made.
And it would draw 800W and it would also set your house on fire as a bonus, but before you die to the burns, it'll deafen you with the most jarring coil whine you'll ever hear.
the connector on the top edge is for infinity fabric which is similar to Nvidia's NVlink so each core can access the memory of the other core and depending on how the driver is setup it could be doing that transparently making the dual gpu act like a single gpu to programs being run on it. As for the MPX connector the pinout while not publicly fully known it carries video down to the motherboard to be routed to thunderbolt connections on the system and brings thunderbolt pcie lanes into the card to be exposed on the rear.
@@hombrepepega3472 Built into what? It's literally routing it to the mobo IO which I'm sure will depend on how the specific machine is configured. It also makes it proprietary which is kind of Apple's thing.
I believe it's possible to run two of these modules in the mac so I'm pretty sure that's what those IF connectors are for... now I'm really curious if a quad GPU setup could get hacked together in Win.
What I used to do, is modify the driver ini file, with the PCI ID that the card is showing. That usually lets it skip by the check and install the driver, to try it. I've done that lots of times to get drivers working for cards using the same chipset, but different PCI ID. SO if you dig into the driver files, you may be able to do this. The thing with the packaged drivers now, is you usually have to run the installer, then copy the unpacked driver from the temp folder, and then modify the copy, and manually install the driver.
I did that with a 1660 Ti when I was having problems in AC Odyssey. There was only 2 drivers available at the time so I took the driver from a month before the 1660 Ti came out and "Whitelisted" the 1660 Ti so the drivers would install and used them for 6 weeks until Ubisoft released the next content update which fixed the problem so I could run the most current drivers again. That's a trick from about 15 years ago we used to use before Nvidia unified the drivers. A lot of times you could take the drivers for a GPU that was the next one up the stack and whitelist your lower tier card and gain a little extra performance as it would change the clock speed and power profile but that trick no longer worked once they unified the drivers. With the 1660 Ti (Which is basically a 2060 without RT and Tensor cores) I suspected the code for it would already be in the drivers a month or two before it released, they just didn't whitelist the card and apparently I was correct.
This Is How you install the newest Mesa good graphics driver on GNU/Linux on PS4 to work perfectly with PS4's AMD GPU with the same driver used for AMD GPU in GNU/Linux on PCs
The extra MPX pins are for thunderbolt - you could plug your monitor into one of the thunderbolt ports on the Mac Pro itself and it would still work, because those extra MPX pins can be used to carry the DisplayPort signals back into the motherboard.
Yeah it does both. It brings 8 PCIe lanes to the Thunderbolt controllers embedded in the MPX module to allow using them for other peripherals than displays and brings back DisplayPort signals to the motherboard to allow connecting displays into any Thunderbolt connectors.
This is why we are fans. Nobody wants to see the 10,000th PC build with the same 10 parts. The unique and off the wall tinkering is what makes it interesting. PS. I was going to suggest that there was some Bootcamp driver for the card since bootcamp lets you run windows on Mac hardware.
@@ryanjofre Yeah, i remember those times when Nvidia and ATI bashed eachother if a Dual-PCB card could be called one card. Fun times, when performance actually made a difference in gameplay settings, and you ACTUALLY could see the differences.
@@Gunni1972 I kinda remember Nvidia wanting to sue Amd/Ati and Vice versa. Those dual cards weren’t perfect but they where bad ass. People greatly exaggerated the micro stuttering. I was gonna buy one but then the 4890 eventually came out and it was so bad ass & affordable you only needed one. I use to play Crysis on ultra settings on that card with just a dual core Phenom II.
Hey Roman, You should check out the XFX BC-160 Accelerator. It uses the obscure AMD Navi12 GPU with 8Gb of HBM2. Its similar to the 5700XT but uses HBM2 instead of GDDR6. They were built for Blockchain usage in mining farms. A bit of background: The Navi12 GPU was only supposed to be used in Apple 2020 Macbook Pros. Apparently there was a massive oversupply of the Navi12 GPU and they started using them in Mining accelerators such as the BC-160. The Navi12 features basically ideantical specs to the 5700XT but with HBM2 and a 4096Bit memory bus. I'd be very interested so see a video on the card. I've never seen the Navi12 GPU in the flesh before.
The drivers for this card are not from AMD but rather Apple in their Boot Camp collection. This is the software that Apple bundles to boot and run Windows on their Mac Pro. You got the GPU drivers from it installed but you will also need to the Thunderbolt drivers installed. The HDMI port on the card is fed from one of those Thunderbolt controllers. The MPX controller should also have a few pins for handling interrupts for Thunderbolt hot plug. Not sure if the HDMI functionality requires those pins to be connected. Depending on the task, video memory is indeed additive instead of mirrored across both GPUs. This is due to the on board Infinity Fabric link between the GPUs. Having two of these cards with the Infinity Fabric bridge would enable quad GPU functionality and the ability to pool all the VRAM together depending on the application. The bridge should be attached so that it can properly terminate the Infinity Fabric link. There are four PCIe devices on that card: two GPUs and two Thunderbolt controllers. Conceptually there would be 16 lanes per GPU, 4 lanes to each Thunderbolt controller, 16 lanes to the PCIe connector and presumably 16 lanes to the MPX connector. That'd be 72 lanes in total. What is worth noting is the PCIe bus on the Mac Pro is only PCIe 3.0, though the GPUs support PCIe 4.0. The bridge chip could be limiting things to PCIe 3.0 speeds.
Honestly Id love to see something like this on pc where the add in cards are these modules which has shrouds fully covering the length of the case as air ducts. While the fans upfront are cooling everything. It just looks so neat and clean.
That would be hell to standardize though, the number of cases with different lengths would be impossible to support. Although perhaps for multi-GPU systems it could provide a benefit, but considering shrouded GPUs would be using more ideal blower fans and coolers anyways, it probably wouldn't be that different in terms of cooling, at least compared to all of the expansion cards being cooled as their own collection.
@@jamesbuckwas6575 There is no need to standardize, some case manufacturer could theoretically sell this thing as a 3rd party cooler for some specific gpu+case line. Realistically not viable to sell as a product, but could be an interesting marketing stunt to maybe manufacture limited pieces at ridiculously high prices just for the sake of it
neat and clean is basically what the Mac Pro is. Clearly, it's a Mac and tech guys are supposed to hate it under penalty of manhood shrink. But if the Mac Pro was a PC we will have to accept it's an engineering wonder. The target is easy to use, so it's easier to install an MPX module than a normal card (no need to move around cables), but it's not designed for hacking.
Good luck trying to get enough airflow trough without bajilion fans on case. Btw apple stuff is already castrated to not melt the entire thing, there is just not enough cooling for full power card, much less overclocked one, which is compensated with apple ""magic"" (having personal OS perfectly tailored to work with your machine and software).
This was super exciting!!! I love these super exotic PC parts, you're on fire with these new videos, and I can certainly hear your excitement, I wonder what else is next, Micron 9400 Workstation U.2 SSD maybe? I'm ready for anything!
The card is correctly stated as 64GB because both GPUs can access all 64GB (although if the memory is in the other GPU the latency is higher, in which _usually_ the CL try using only local memory). What the Infinity Fabric, SLI, Crossfire, NVlink does is to memory map each-other GPU memory space in a high bandwidth dedicated link. It's somewhat limited in consumer grade SLI/Crossfire, but for these Pro series cards, basically any compute language that runs on the card can access full memory, regardless of using a single gpu or all of them.
Love the different mechanical design of this to what we're used to in the PC space. It's got more elegance to it than the stuff that happens in servers imo (eg the release/removal lever), and reminds me a it of how the gear @CuriousMarc takes apart fits together, but with fewer wire looms!
Nice ingenuity getting this to work. I only recently got into doing custom PC mods and I enjoyed seeing someone who knows what they're doing attempting to do something unique. Subscribed!
I think the MPX also has a link to the SMC in the MacPro. Apple stopped implementing GPU metrics in the driver as they communicate directly with the SMC on all other Macs to control power and cooling.
It’s nice to see Techtubers helping each other and not hiding their secrets for themselves 🥰😇💪👍. I enjoyed seeing this gpus pcb ….. what a thing of beauty 🤩
@@Karti200 I wonder the same, because card was meant for Unix based system, but also not sure if Apple's BootCamp mode more reliably supports Linux based OS instead of Windows OS
@@tadasnanartonis9092 There is no really an issue running Linux on "pre-M series" Apple devices. Yes, there is still BootCamp, but honestly there is nothing really making it harder for you to install Linux on those. Just download iso, make bootable iso and launch from it - poof and install :) There are even videos with people running Ubuntu off Apple products from 2019/2020
This was awesome.! Always wanted to see inside this card and see how it done under load! Thanks! Glad you got it going. Would love to see a waterblock for it!
yeah It's kinda annoying that this is still an obscure thing for many people. It's so convenient but unfortunately there's a lack of benchmark support on linux. There are some but it's not comparable to the usual ones. I guess Wine could run those without a problem never checked though.
There's a difference between support and performance - generally speaking you'll get higher framerates with the windows drivers - considerably more time and effort has been spent there.
@@JohnnyWednesday My 6950xt is faster in linux than windows for hogwarts legacy, it's gotten to the point that many games are actually faster in linux than windows now, it's largely down to when they're slower it's because proton isn't playing well with the game itself.
@@JohnnyWednesday This is true for games but for everything else Linux runs it better. Especially Blender. Edit: specifically in regards to running something on an AMD gpu
Reminds me of my old gtx 690 dual GPU card ...that was a production pc card but still an absolute pain to get working in 99% of applications and games ... I sold it after just a few weeks it wasn't worth the headaches ...went back to the gtx680 I had as it had more consistent performance in most games.
That is a neat video, I am not surprised there is not much information about the MPX module, getting any sort of electric diagram out of Apple has been the bane of every repair shop in existance. but its cool to see it working. I have also seen videos where the 6800x duo is claimed to be thermal trottling once installed on a Mac Pro and running a productivity suit, looks like the passive cooling and the few fans the mac pro has is not enough.
Of cource it is not enough. Same card still runs hot with 3 fans at full power, now imagine having 2 for both cards, with very few fins, and their RPM is gipmed af because apple doesnt want their customers to be annoyed at the fact that their computer is struggles to cool itself.
Would love to see more of this card. I think that dual gpu power is underrated by the market, it would be cool to see it coming back to personal builds some day with the new technology that we have in the present.
Is not about being underrated. Is about not being needed. There is just no reason for it in consumer space other than a few enthusiasts trying to pull 120 fps in 8k with ray tracing. If you go to lower chips then why bother by putting two together instead of just buying a higher spec one.
@@SIPEROTH Adding to that, especially games have a lot of issues with dual-GPU setups, it's the entire reason why SLI/NVLink failed there too. You need to put a lot of effort into synchronizing rendering between the two GPUs, and that kind of task management is hard.
OH WOW!!!! THAT WAS AWESOME!!!!! II can't believe you got it to work!!!! YES!!! MORE TESTING!!!! I just LOVE stuff like this!!! Your videos are always SO MUCH FUN!!!! THANK YOU!!! :D
Hi Roman your videos and enthusiasm is just insane and keeps me watchinng and sometimes even rewatching for zen-purposes during my lunch breaks etc. P.s. whenever someone says the word insane many of us immediately think about your channel haha!
Awesome video! Would be really fun if you'd manage to put this on a waterblock and attempt some overclocking with it. Also to put it against the 4090 for gaming and productivity work, try raytracing and other features!
@@FixedFunction Thank you :) Apple is not so closed as most people think. macOS X is OpenSource for most of its code. If you look for the white paper for the 2019 MacPro, Apple does provide schematics (high level) for the cards and explains why they do MPX. I had posted the link but it got removed by UA-cam. Just Google it.
Back when I used a Mac Pro 2006 and a 5000 series AMD GPU (official from Apple), I couldn't install the official AMD Radeon drivers because of the vendor ID on the card was from Apple. My options were to change the vendor ID via a BIOS mod or to install the drivers that are included in the Apple Bootcamp driver. EDIT: I'd love to see a follow-up video of this, especially if you end up making a water block. Would be really cool if you were able to design a custom PCB that the card just slots into and adapts it for use with a normal PC, without needing to solder wires.
Wow it's awesome AMD are still going with dual GPU in 2021! I can't see this as stable for gaming but for professional workloads perhaps this could be a thing. I wonder once PCIE 5.0 is more widely adopted if AMD would go with dual GPU options for the high-end as a single slot card.
Not PCIe gen 5 but MCM / die-to-die interconnect could definitely make it happen again. That's what Apple is doing with the M-Ultra series SoCs. Connecting two 420mm2 (not including interconnect) or 460mm2 interconnect included) dies together two form a 920mm2 die that would otherwise not be possible via traditional reticles.
This is fascinating. The Mac Pro allows two (2) Duo cards with a fabric bridge installed. The older generation had HBM memory too (instead of GDDR6). This is awesome experimentation nonetheless. Bravo!
I would be interested to see more compute oriented workloads. Blender rendering or folding@home for example. This might be a worthwhile deal for a diy setup. Could also be interesting to see if the card can be modified (MPX connector cut off) to fit in a normal motherboard. This thing with a custom watercooling solution would probably be a beast.
Blender would be expected since it can utilize multi-GPU with pretty much plug-and-play (Non-Eevee at least) and also doesn't chugged PCIE bandwidth. I would be more interested in AI accelerator.
There are some "PCIe extenders" which are basically a PCIe x16 card edge connector with a PCIe x16 slot soldered to them. I would think that one of them would raise the card up from the motherboard enough that the MPX connector would clear it. I would do that rather than try to cut off the connector. Then you could use a stand off of the right height in the screw hole on the cards backplate.
I did similar tests with an MPX 580 Pro a few years ago. Sadly the display outputs will never work in a PC, they're all routed through the MPX connector to a thunderbolt mux on the MacPro 7 logic-board (even the HDMI port). You can use the EDID emulation feature in the Radeon Pro driver to create a dummy display, needed to make the card boost properly in some 3D apps, but you still loose some performance using the iGPU for display.
So epic; the knowledge and skills to make this happen is just awesome. Makes me chuckle, thinking about another vid regarding his 4090 Strix not running at full PCIE x16, people in the comments were like "it could be your motherboard, you should try another motherboard to be sure." COME ON, REALLY? This is der8auer folks.
Der8auer you just showed me one of my wishes is actually somewhat achievable! Mad respect to you for spending so much money on something that is pretty much destined to be a showpiece! If I may ask, do you perhaps plan on doing something similar with a Vega II Duo? I believe that is the closest you can get to a full Radeon VII without going for an Instinct MI60. Awesome video!
Thank you for your content DeBauer, for the thermal grizzly and you're snazzy cases plus all the crazy content. You've been on my radar since that weird ad you did for Asus on their benchmarking program. Do you still dub?
What if the MPX connector is for quad crossifire? If I remember correctly you could put 2 of such cards in the MAC pro. So if you want a quad card solution you will need ugly bridges. But for apple this is a no go. Do we have Crossfire pinout? The Crossfire jumper I have has 10 pairs, this matches with the 10 pairs with decoupling caps I can see in the video. If the other pairs (the one without caps are also 10 it means that is CrossfireX bridge betweent the 4 GPUs. I have an really old HD ATI card and I've seen that the corossfire is indeed 10 pairs but no caps in the pairs. Someone should check how/if the lanes are split in 5tx+5rx on newer cards and if they have have the standard 22nF caps in TX lanes of hi speed busses. About the pcie gen3 it might be the cable the PLX (*pfx) chip is marked G4 soo... riser cable issue? EDIT: I think that the cardss will be connected as such: CARD_A: GPU_1 GPU_2 ▲MPX▼ ▲MPX▼ CARD_B: GPU_3 GPU_4 If some hi res photos would be available I could dig a little bit more... in the "reverse engineering".
Its for the thunderbolt ports, but maybe could be used for other things. All thunderbolt ports on the mac pro can be used for display out, including the front panel
On cards from older Mac Pros, you would need to flash the vBios with a modded bios. Usually people do it the other way round, flashing Windows cards with modified vBios ie the GTX680 runs perfect on Mac Pros including display during boot. Maybe there's a modding scene for these GPUs.
Bro, make a project around this card a full PC build. 1. Water block, maybe dual sided since it has ram on the back. 2. instead of soldering wired created a plug-in 3D printed adapter with prover wires going into adapter and all connected
Versuchs mal mit MS Hybrid. Damit sollte man alle Anwendungen über iGPU ausgeben können aber über die "richtige" GPU, also die DualGPU, rendern. Hat jedenfalls bei den Tesla Karten funktioniert.
The MPX data slot is used for 2 things 1) to rout video out back to the motherboard so you can attach your screen to any TB port on the macPro you do not need to attach to the one on the back of your GPU 2) to rout TB connections to the TB ports, those are not just display they are full TB including PCIe tunnelling. You're not going to be able to make use of it even with a pinout since there are no intel xeon platforms out there with the needed firmware that would enabled this type of TB tunnelling. These gpus do not use crossfire. The multi gpu support they have is differnt and is forced on compute. Apple expose a set of multi gpu apis in metal that let one gpu read and write memroy from the other (without going over the system) including scheduling tasks on each other without involving the cpu.
I really want to see this GPU in other workloads, especially in games, to see if it is running crossfire, and only a few games are supported, or maybe it is something else. I know it is not likely, but if this multiGPU is visible to programs as just one, this could mean some future for multiGPU, this time with infinity fabric.
I wonder if a custom edge connector PCB for the power contacts with VERY thick traces that break out to a 2x8 or 3x8 PEG connector might let this thing run at full tilt?
This is really cool. A powerful dual gpu for sure. It’s a shame that multi gpu is dead. The writing was on the wall when it required game developers to implement it in their games for dx12 and vulkan.
Pretty sure with DX12 and Vulcan, it's that developers DON'T have to implement it themselves as with previous APIs they couldn't be bothered to do the work.
@@MaddJakd I think what he's referring to was that Dx12 implemented somekind of multiple discrete GPU method. That would definitely require developers to implement it. Basically it would mean that GPUs are just separate devices and it would be entirely up to the software developer to utilize them all or not. Anyway, any time microsoft takes control of a feature away from the hardware manufacturers it's usually the death nail for that technology or feature. They do it their way, which typically sucks and nobody wants to support it, then it dies. That's par for the course for microsoft. They want your PC to be dull and boring and fully under their control!
It's technically possible if it's done like what Apple did with the M1 Ultra. By connecting both chips together with such a fast connection, the system treats them as a single chip.
I believe you must check that MPX connector to confirm EVERY power connection, I believe this is shutting down due to lack of power so yes lack of wires but also connected elsewhere. Another important detail would be the missing SMBUS drivers etc, power wise budget your cabling for at least 500watts over 12v, and check please check for additional power connections and their voltages. 5v should no be going over 75w and 3.3v max at 45/55w
Search for TAGG XOC he have some eye watering jaw dropping videos about modding motherboards and video cards, lots of soldering, cutting, scrapping and dremeling involved!
Fascinating video. It's very interesting to see the technology these companies hide away (through exclusive prices and availability) as the scaling against the 6800xt's performance looked, alright, weirdly.
I'd love to see a follow up to this. Its super interesting that this Apple AMD Dual gpu actually functions at all on a windows system but also completely stomps the 7900 XTX. Maybe dual gpus could make a comeback in the future using whatever technology AMD used in these modules? Clearly its working much more seamlessly than typical crossfire solutions. Though, that's just with Timespy. I wonder if the performance gain over the 7900 XT will hold true in real production applications or games.
its huge difference in setup...its base on CDNA and consumers are using different branch RDNA architecture ...CDNA is base on older tech from GCN3.0 that been renamed to CDNA...compute processing vs RDNA is old raster processing that base way older tech from ati since hd6900...hence this how far they can get versus raja koduri GCN tech can go further is unexpected for amd...
yes please do a follow up video on this. I would love to see waterblock added to this or some sort of cooler to lock down its max performance. Very impressive
I'm all for dual GPUs! 1kHz OLED monitors are absolutely possible and absolutely nothing we have today from either Nvidia or AMD can touch 4K @ 1000Hz (kHz) on a single, much less, even a double GPU. If Nvidia brought back 4-way SLI I guarantee you people running it regardless of the expense.
I really wish multi gpu setups aren't a thing of the past. when I first started building I always dreamed of a crossfire setup then years down the road when I'm able to afford a dual GPU setup its pointless unless I want to show off 3dmark
The strange bridge on top of the gpu is the infinity fabric link. That bridge is allowing one gpu to have access to the other gpu memory
I think you can link more of those cards .
This way .
Its not make for a pc .
It make for professional systeem
u guessing
@@dyslectische i think you can do exactly that with Radeon Pro and Instincts
@@dyslectische "pRoFesSioNaL SyStEm", FTFY.
On Mac that bridge allows both cards to access ALL resources between them since they route all I/O through a set of MUXs.
I think I have found a worthy successor to my RX Vega obsession. Long live weirdly performant AMD GPUs!
thing is if i really wont get anything, for whatever, justifying, reason, then i can just stop trying and try other work
HBM memory. The ability to extract more performance out of a Vega 56 by flashing the Vega 64 VBIOS on it. The very existence of Radeon VII.
The RX Vega cards were some of the quirkiest mainstream GPUs ever made.
@@d4nyllradeon VII server farm for FP64🤤
A dual GPU 7900XT that actually worked for gaming would be a monster
Everything in multi hardware setup would be monster if it works
If
nvidia 5090 heres it comes
But the price tho
@@cap_napu7757 probably still cheaper than a 4090 😆
And it would draw 800W and it would also set your house on fire as a bonus, but before you die to the burns, it'll deafen you with the most jarring coil whine you'll ever hear.
the connector on the top edge is for infinity fabric which is similar to Nvidia's NVlink so each core can access the memory of the other core and depending on how the driver is setup it could be doing that transparently making the dual gpu act like a single gpu to programs being run on it. As for the MPX connector the pinout while not publicly fully known it carries video down to the motherboard to be routed to thunderbolt connections on the system and brings thunderbolt pcie lanes into the card to be exposed on the rear.
That's very cool. Thanks for that info
Why not just build it in?
that's quite amazing what they managed to build with this card and the mac. If only we could see these kinds of things on more vendors.
@@hombrepepega3472 Built into what? It's literally routing it to the mobo IO which I'm sure will depend on how the specific machine is configured. It also makes it proprietary which is kind of Apple's thing.
I believe it's possible to run two of these modules in the mac so I'm pretty sure that's what those IF connectors are for... now I'm really curious if a quad GPU setup could get hacked together in Win.
You're an absolute legend. I had thought of this back when Apple released their dual GPU cards lol and you actually did it! Madlad!
YES!!!
So, naturally, we would also like a follow-up with more 12V wires and appropriate cooling (industrial turbine fan level at least).
thank you!
Nothing new 295 Nvidia Dual old project or 550 Ti Dual... . Mac shit and price for Bankster childs.... .
@@der8auer-enI always thought this card was interesting, dual LN2 cooled would be insane!
Haha reminds me of blackwell but more accessible.
What I used to do, is modify the driver ini file, with the PCI ID that the card is showing. That usually lets it skip by the check and install the driver, to try it. I've done that lots of times to get drivers working for cards using the same chipset, but different PCI ID. SO if you dig into the driver files, you may be able to do this. The thing with the packaged drivers now, is you usually have to run the installer, then copy the unpacked driver from the temp folder, and then modify the copy, and manually install the driver.
Hm... pretty much the same way i did it with NVIDIA. Where i can read about Radeons? Just interesting.
👍🏼 nice
I did that with a 1660 Ti when I was having problems in AC Odyssey. There was only 2 drivers available at the time so I took the driver from a month before the 1660 Ti came out and "Whitelisted" the 1660 Ti so the drivers would install and used them for 6 weeks until Ubisoft released the next content update which fixed the problem so I could run the most current drivers again. That's a trick from about 15 years ago we used to use before Nvidia unified the drivers. A lot of times you could take the drivers for a GPU that was the next one up the stack and whitelist your lower tier card and gain a little extra performance as it would change the clock speed and power profile but that trick no longer worked once they unified the drivers. With the 1660 Ti (Which is basically a 2060 without RT and Tensor cores) I suspected the code for it would already be in the drivers a month or two before it released, they just didn't whitelist the card and apparently I was correct.
This Is How you install the newest Mesa good graphics driver on GNU/Linux on PS4 to work perfectly with PS4's AMD GPU with the same driver used for AMD GPU in GNU/Linux on PCs
but its a little more complicated on GNU/Linux
The extra MPX pins are for thunderbolt - you could plug your monitor into one of the thunderbolt ports on the Mac Pro itself and it would still work, because those extra MPX pins can be used to carry the DisplayPort signals back into the motherboard.
or bring some PCIe over the existing TB3.0 ports, depending where it is going.
Yeah it does both. It brings 8 PCIe lanes to the Thunderbolt controllers embedded in the MPX module to allow using them for other peripherals than displays and brings back DisplayPort signals to the motherboard to allow connecting displays into any Thunderbolt connectors.
I think I just purchased the last last last AMD dual GPU card the never released but I somehow got an engineering sample of Radeon pro v540
@@beatyoubeachyt8303
Whut? How?
Can you put two pcie risers and plug it into two pcie slots?
I definitely love me some Dual-GPU cards! I've been wanting to test one of these cards for a while. Glad you were able to get your hands on one!
This bad boy can fit so much cloud gaming server in it
Cheers mate!
This is why we are fans. Nobody wants to see the 10,000th PC build with the same 10 parts. The unique and off the wall tinkering is what makes it interesting.
PS. I was going to suggest that there was some Bootcamp driver for the card since bootcamp lets you run windows on Mac hardware.
Cool my system if you are fan god damnit! Get back in my case and stop writing UA-cam comments!
I'd be interested to see a custom block and power metering. I miss dual GPU setups, this was a fun episode!
Back when you could get more frames in exchange for horrible stuttering.
Yeah, allot of gamers are to young to remember when Ati/Amd sold the 4850x2 & 4870 x 2👌
@@ryanjofre Yeah, i remember those times when Nvidia and ATI bashed eachother if a Dual-PCB card could be called one card. Fun times, when performance actually made a difference in gameplay settings, and you ACTUALLY could see the differences.
@@Gunni1972 I kinda remember Nvidia wanting to sue Amd/Ati and Vice versa.
Those dual cards weren’t perfect but they where bad ass. People greatly exaggerated the micro stuttering.
I was gonna buy one but then the 4890 eventually came out and it was so bad ass & affordable you only needed one. I use to play Crysis on ultra settings on that card with just a dual core Phenom II.
Hey Roman, You should check out the XFX BC-160 Accelerator. It uses the obscure AMD Navi12 GPU with 8Gb of HBM2.
Its similar to the 5700XT but uses HBM2 instead of GDDR6. They were built for Blockchain usage in mining farms.
A bit of background:
The Navi12 GPU was only supposed to be used in Apple 2020 Macbook Pros. Apparently there was a massive oversupply of the Navi12 GPU and they started using them in Mining accelerators such as the BC-160. The Navi12 features basically ideantical specs to the 5700XT but with HBM2 and a 4096Bit memory bus.
I'd be very interested so see a video on the card. I've never seen the Navi12 GPU in the flesh before.
The drivers for this card are not from AMD but rather Apple in their Boot Camp collection. This is the software that Apple bundles to boot and run Windows on their Mac Pro. You got the GPU drivers from it installed but you will also need to the Thunderbolt drivers installed. The HDMI port on the card is fed from one of those Thunderbolt controllers. The MPX controller should also have a few pins for handling interrupts for Thunderbolt hot plug. Not sure if the HDMI functionality requires those pins to be connected.
Depending on the task, video memory is indeed additive instead of mirrored across both GPUs. This is due to the on board Infinity Fabric link between the GPUs. Having two of these cards with the Infinity Fabric bridge would enable quad GPU functionality and the ability to pool all the VRAM together depending on the application. The bridge should be attached so that it can properly terminate the Infinity Fabric link.
There are four PCIe devices on that card: two GPUs and two Thunderbolt controllers. Conceptually there would be 16 lanes per GPU, 4 lanes to each Thunderbolt controller, 16 lanes to the PCIe connector and presumably 16 lanes to the MPX connector. That'd be 72 lanes in total. What is worth noting is the PCIe bus on the Mac Pro is only PCIe 3.0, though the GPUs support PCIe 4.0. The bridge chip could be limiting things to PCIe 3.0 speeds.
That was very interesting. Please do make followup with proper power delivery and more tests!
Honestly Id love to see something like this on pc where the add in cards are these modules which has shrouds fully covering the length of the case as air ducts. While the fans upfront are cooling everything.
It just looks so neat and clean.
Also be a ton easier to stack the cards with out suffocation some of them.
That would be hell to standardize though, the number of cases with different lengths would be impossible to support. Although perhaps for multi-GPU systems it could provide a benefit, but considering shrouded GPUs would be using more ideal blower fans and coolers anyways, it probably wouldn't be that different in terms of cooling, at least compared to all of the expansion cards being cooled as their own collection.
@@jamesbuckwas6575 There is no need to standardize, some case manufacturer could theoretically sell this thing as a 3rd party cooler for some specific gpu+case line. Realistically not viable to sell as a product, but could be an interesting marketing stunt to maybe manufacture limited pieces at ridiculously high prices just for the sake of it
neat and clean is basically what the Mac Pro is. Clearly, it's a Mac and tech guys are supposed to hate it under penalty of manhood shrink. But if the Mac Pro was a PC we will have to accept it's an engineering wonder. The target is easy to use, so it's easier to install an MPX module than a normal card (no need to move around cables), but it's not designed for hacking.
Good luck trying to get enough airflow trough without bajilion fans on case. Btw apple stuff is already castrated to not melt the entire thing, there is just not enough cooling for full power card, much less overclocked one, which is compensated with apple ""magic"" (having personal OS perfectly tailored to work with your machine and software).
This was super exciting!!! I love these super exotic PC parts, you're on fire with these new videos, and I can certainly hear your excitement, I wonder what else is next, Micron 9400 Workstation U.2 SSD maybe?
I'm ready for anything!
The card is correctly stated as 64GB because both GPUs can access all 64GB (although if the memory is in the other GPU the latency is higher, in which _usually_ the CL try using only local memory). What the Infinity Fabric, SLI, Crossfire, NVlink does is to memory map each-other GPU memory space in a high bandwidth dedicated link.
It's somewhat limited in consumer grade SLI/Crossfire, but for these Pro series cards, basically any compute language that runs on the card can access full memory, regardless of using a single gpu or all of them.
Love the different mechanical design of this to what we're used to in the PC space.
It's got more elegance to it than the stuff that happens in servers imo (eg the release/removal lever), and reminds me a it of how the gear @CuriousMarc takes apart fits together, but with fewer wire looms!
Nice ingenuity getting this to work. I only recently got into doing custom PC mods and I enjoyed seeing someone who knows what they're doing attempting to do something unique. Subscribed!
You just subscribed to one of the two top modders on UA-cam and probably in existence. Smart move.
@@asiatravel2010 who is the other person? Thanks!
@@toddsherman812 in my opinion that would be Steve at Gamers Nexus.
@@toddsherman812 Buildzoid? (Definitely for his GPU necromancy series.)
This is seriously cool! Would love to see a custom water block!
I think the MPX also has a link to the SMC in the MacPro. Apple stopped implementing GPU metrics in the driver as they communicate directly with the SMC on all other Macs to control power and cooling.
It’s nice to see Techtubers helping each other and not hiding their secrets for themselves 🥰😇💪👍. I enjoyed seeing this gpus pcb ….. what a thing of beauty 🤩
Maybe running that gpu off linux could work better?
@@Karti200 I wonder the same, because card was meant for Unix based system, but also not sure if Apple's BootCamp mode more reliably supports Linux based OS instead of Windows OS
@@tadasnanartonis9092 There is no really an issue running Linux on "pre-M series" Apple devices. Yes, there is still BootCamp, but honestly there is nothing really making it harder for you to install Linux on those. Just download iso, make bootable iso and launch from it - poof and install :) There are even videos with people running Ubuntu off Apple products from 2019/2020
Its always a pleasure to enjoy interesting contents like this. Awesome video as always 🙏🏻🧡
Aha that's dope! Would love to see a part 2!
This was awesome.! Always wanted to see inside this card and see how it done under load! Thanks! Glad you got it going. Would love to see a waterblock for it!
Linux has native support for amd gpus, I wonder if it would’ve worked there out of the box
yeah It's kinda annoying that this is still an obscure thing for many people. It's so convenient but unfortunately there's a lack of benchmark support on linux. There are some but it's not comparable to the usual ones. I guess Wine could run those without a problem never checked though.
@@akosv96 by benchmark you most likely mean gaming benchmark, as linux has a ton of production benchmarks, including ones that use the gpu
There's a difference between support and performance - generally speaking you'll get higher framerates with the windows drivers - considerably more time and effort has been spent there.
@@JohnnyWednesday My 6950xt is faster in linux than windows for hogwarts legacy, it's gotten to the point that many games are actually faster in linux than windows now, it's largely down to when they're slower it's because proton isn't playing well with the game itself.
@@JohnnyWednesday This is true for games but for everything else Linux runs it better. Especially Blender. Edit: specifically in regards to running something on an AMD gpu
I've been waiting for years for a PCB breakdown of this card!
Reminds me of my old gtx 690 dual GPU card ...that was a production pc card but still an absolute pain to get working in 99% of applications and games ... I sold it after just a few weeks it wasn't worth the headaches ...went back to the gtx680 I had as it had more consistent performance in most games.
This is a nice look into what graphics cards could be. I don't know why there aren't more like this.
That is a neat video, I am not surprised there is not much information about the MPX module, getting any sort of electric diagram out of Apple has been the bane of every repair shop in existance. but its cool to see it working. I have also seen videos where the 6800x duo is claimed to be thermal trottling once installed on a Mac Pro and running a productivity suit, looks like the passive cooling and the few fans the mac pro has is not enough.
Of cource it is not enough. Same card still runs hot with 3 fans at full power, now imagine having 2 for both cards, with very few fins, and their RPM is gipmed af because apple doesnt want their customers to be annoyed at the fact that their computer is struggles to cool itself.
@@alexturnbackthearmy1907what a waste of card :S
7:50 Cat inspection, the cat approve 🐱
Would love to see more of this card. I think that dual gpu power is underrated by the market, it would be cool to see it coming back to personal builds some day with the new technology that we have in the present.
Is not about being underrated. Is about not being needed. There is just no reason for it in consumer space other than a few enthusiasts trying to pull 120 fps in 8k with ray tracing.
If you go to lower chips then why bother by putting two together instead of just buying a higher spec one.
@@SIPEROTH Adding to that, especially games have a lot of issues with dual-GPU setups, it's the entire reason why SLI/NVLink failed there too. You need to put a lot of effort into synchronizing rendering between the two GPUs, and that kind of task management is hard.
Insanely interesting thanks ! I hope to see a followup video
What a beautiful looking GPU
OH WOW!!!! THAT WAS AWESOME!!!!!
II can't believe you got it to work!!!!
YES!!! MORE TESTING!!!! I just LOVE stuff like this!!! Your videos are always SO MUCH FUN!!!!
THANK YOU!!! :D
It’s always interesting to see dual GPU designs. Being Mac just adds extra mystery.
Hi Roman your videos and enthusiasm is just insane and keeps me watchinng and sometimes even rewatching for zen-purposes during my lunch breaks etc. P.s. whenever someone says the word insane many of us immediately think about your channel haha!
Awesome video! Would be really fun if you'd manage to put this on a waterblock and attempt some overclocking with it. Also to put it against the 4090 for gaming and productivity work, try raytracing and other features!
AMD cancelling Crossfire was a disservice to gamers everywhere.
A crossfire 7900xtx would be monstrous and would put serious fire onto nvidia.
The MPX connector is used to reroute PCIe to Thunderbolt ports throughout the machine or the Afterburner card.
Finally, yes. Somebody who understands it's not just, "Apple proprietary bullshit," it actually serves a purpose.
@@FixedFunction Thank you :) Apple is not so closed as most people think. macOS X is OpenSource for most of its code. If you look for the white paper for the 2019 MacPro, Apple does provide schematics (high level) for the cards and explains why they do MPX. I had posted the link but it got removed by UA-cam. Just Google it.
Would love to see a follow up video, really interesting to see very unique hardware that rarely gets on the spotlight
Back when I used a Mac Pro 2006 and a 5000 series AMD GPU (official from Apple), I couldn't install the official AMD Radeon drivers because of the vendor ID on the card was from Apple. My options were to change the vendor ID via a BIOS mod or to install the drivers that are included in the Apple Bootcamp driver. EDIT: I'd love to see a follow-up video of this, especially if you end up making a water block. Would be really cool if you were able to design a custom PCB that the card just slots into and adapts it for use with a normal PC, without needing to solder wires.
Great to see someone test one of these out, I remember being very surprised to see it being released at all during the height of the mining craze.
Yes , please make more video's about this card
This is the most exciting video I have ever seen in this channel so far.
thanks :)
Wow it's awesome AMD are still going with dual GPU in 2021! I can't see this as stable for gaming but for professional workloads perhaps this could be a thing. I wonder once PCIE 5.0 is more widely adopted if AMD would go with dual GPU options for the high-end as a single slot card.
Not PCIe gen 5 but MCM / die-to-die interconnect could definitely make it happen again. That's what Apple is doing with the M-Ultra series SoCs. Connecting two 420mm2 (not including interconnect) or 460mm2 interconnect included) dies together two form a 920mm2 die that would otherwise not be possible via traditional reticles.
8:15 - “Already the backside is such a beauty”. I wondered whether he was talking about the card, the cat, or someone standing off screen?!?!
12:57 did you measure voltage drop on those two wires? seems too little for running dual GPU card even accounting for PCI-E power
Super interesting video! Thank you for sharing and doing something special with this beautiful dual GPU
This is fascinating. The Mac Pro allows two (2) Duo cards with a fabric bridge installed. The older generation had HBM memory too (instead of GDDR6).
This is awesome experimentation nonetheless. Bravo!
Looking forward to your watercooling solution and more tests or experiments on this amazing card)
I would be interested to see more compute oriented workloads. Blender rendering or folding@home for example. This might be a worthwhile deal for a diy setup.
Could also be interesting to see if the card can be modified (MPX connector cut off) to fit in a normal motherboard. This thing with a custom watercooling solution would probably be a beast.
Blender would be expected since it can utilize multi-GPU with pretty much plug-and-play (Non-Eevee at least) and also doesn't chugged PCIE bandwidth.
I would be more interested in AI accelerator.
There are some "PCIe extenders" which are basically a PCIe x16 card edge connector with a PCIe x16 slot soldered to them. I would think that one of them would raise the card up from the motherboard enough that the MPX connector would clear it. I would do that rather than try to cut off the connector. Then you could use a stand off of the right height in the screw hole on the cards backplate.
Thats amazing, if only AMD or Nvidia could make a dual GPU function as one GPU.
Amazing piece of tech, I`d love to see more testing with it. Maybe even a custom waterblock?
It’s so nice to see you all grown up, Jimmy! Where’s Goddard??
He's built numerous goddards since his youth
@@BBWahoo
Yea I guess he’s too grown up for something so simple now 😂😢
I did similar tests with an MPX 580 Pro a few years ago. Sadly the display outputs will never work in a PC, they're all routed through the MPX connector to a thunderbolt mux on the MacPro 7 logic-board (even the HDMI port). You can use the EDID emulation feature in the Radeon Pro driver to create a dummy display, needed to make the card boost properly in some 3D apps, but you still loose some performance using the iGPU for display.
So epic; the knowledge and skills to make this happen is just awesome. Makes me chuckle, thinking about another vid regarding his 4090 Strix not running at full PCIE x16, people in the comments were like "it could be your motherboard, you should try another motherboard to be sure." COME ON, REALLY? This is der8auer folks.
Der8auer you just showed me one of my wishes is actually somewhat achievable!
Mad respect to you for spending so much money on something that is pretty much destined to be a showpiece!
If I may ask, do you perhaps plan on doing something similar with a Vega II Duo? I believe that is the closest you can get to a full Radeon VII without going for an Instinct MI60.
Awesome video!
Thank you for your content DeBauer, for the thermal grizzly and you're snazzy cases plus all the crazy content. You've been on my radar since that weird ad you did for Asus on their benchmarking program. Do you still dub?
I believe the mpx slot is used to get display signals to all the thunderbolt ports on the MAC chassis, along with other functionality I’m sure.
yep it is also used so that the TB ports on the back of the GPU are full TB including data etc not just for display!
Amazing video, love when you have unique content like this. Thank you Sir.
17:15 the MPX connector is there to make the card proprietary. It's even in the name MPX: Mac Pro Extension.
It provides a direct and dedicated I/O path for the MUX so that every Type-C connector on the machine can be a display output from the card(s).
Looking forward to a follow-up!
What if the MPX connector is for quad crossifire?
If I remember correctly you could put 2 of such cards in the MAC pro. So if you want a quad card solution you will need ugly bridges. But for apple this is a no go.
Do we have Crossfire pinout? The Crossfire jumper I have has 10 pairs, this matches with the 10 pairs with decoupling caps I can see in the video. If the other pairs (the one without caps are also 10 it means that is CrossfireX bridge betweent the 4 GPUs.
I have an really old HD ATI card and I've seen that the corossfire is indeed 10 pairs but no caps in the pairs. Someone should check how/if the lanes are split in 5tx+5rx on newer cards and if they have have the standard 22nF caps in TX lanes of hi speed busses.
About the pcie gen3 it might be the cable the PLX (*pfx) chip is marked G4 soo... riser cable issue?
EDIT: I think that the cardss will be connected as such:
CARD_A: GPU_1 GPU_2
▲MPX▼ ▲MPX▼
CARD_B: GPU_3 GPU_4
If some hi res photos would be available I could dig a little bit more... in the "reverse engineering".
Its for the thunderbolt ports, but maybe could be used for other things. All thunderbolt ports on the mac pro can be used for display out, including the front panel
there is a quad fabric connector for dual duos, search for apple a2667
Yes definitely do a follow up video. Would love to see what this card can really do
Glad the Bootcamp driver made this GPU work on Windows. While ARM has its advantages, the new Macs are far more locked down, so it has its down sides.
the madlad casually running alien hardware
we need more THIS WAS INTRESTING
a mad scientist indeed
never regretti getting his o11-dxl pc case
On cards from older Mac Pros, you would need to flash the vBios with a modded bios. Usually people do it the other way round, flashing Windows cards with modified vBios ie the GTX680 runs perfect on Mac Pros including display during boot. Maybe there's a modding scene for these GPUs.
What a super interesting video! Well done, Sir.
Definitely would love to see a follow up on this GPU. Wonder if waterblock + OC could beat the 4090 😂
Its still a 6800xt level card. It could never beat a 4090. Even in render workloads
Not even a chance...
@@KG_BM it is ideally twice the performance than a 6800 XT
Bro, make a project around this card a full PC build.
1. Water block, maybe dual sided since it has ram on the back.
2. instead of soldering wired created a plug-in 3D printed adapter with prover wires going into adapter and all connected
I'd love to see a custom waterblock for this. I've owned several dual GPU cards. I've had a 5870x2, 7990, 2 295x2s.
still have a 295 gtx and a 3870x2........cool novelty,good to run crysis back in the day
I miss dual cards and crossfire/sli. Really wish that they were still supported
Versuchs mal mit MS Hybrid. Damit sollte man alle Anwendungen über iGPU ausgeben können aber über die "richtige" GPU, also die DualGPU, rendern. Hat jedenfalls bei den Tesla Karten funktioniert.
The MPX data slot is used for 2 things
1) to rout video out back to the motherboard so you can attach your screen to any TB port on the macPro you do not need to attach to the one on the back of your GPU
2) to rout TB connections to the TB ports, those are not just display they are full TB including PCIe tunnelling.
You're not going to be able to make use of it even with a pinout since there are no intel xeon platforms out there with the needed firmware that would enabled this type of TB tunnelling.
These gpus do not use crossfire. The multi gpu support they have is differnt and is forced on compute. Apple expose a set of multi gpu apis in metal that let one gpu read and write memroy from the other (without going over the system) including scheduling tasks on each other without involving the cpu.
This.
I really want to see this GPU in other workloads, especially in games, to see if it is running crossfire, and only a few games are supported, or maybe it is something else.
I know it is not likely, but if this multiGPU is visible to programs as just one, this could mean some future for multiGPU, this time with infinity fabric.
Zoltar would flip watching your solider skills.
I wonder if a custom edge connector PCB for the power contacts with VERY thick traces that break out to a 2x8 or 3x8 PEG connector might let this thing run at full tilt?
This is really cool. A powerful dual gpu for sure. It’s a shame that multi gpu is dead. The writing was on the wall when it required game developers to implement it in their games for dx12 and vulkan.
Pretty sure with DX12 and Vulcan, it's that developers DON'T have to implement it themselves as with previous APIs they couldn't be bothered to do the work.
@@MaddJakd I think what he's referring to was that Dx12 implemented somekind of multiple discrete GPU method. That would definitely require developers to implement it. Basically it would mean that GPUs are just separate devices and it would be entirely up to the software developer to utilize them all or not. Anyway, any time microsoft takes control of a feature away from the hardware manufacturers it's usually the death nail for that technology or feature. They do it their way, which typically sucks and nobody wants to support it, then it dies. That's par for the course for microsoft. They want your PC to be dull and boring and fully under their control!
Astonished that the driver website doesn't have a banner "Hi, der8auer. Good to see you"
dont we all wish single board dual gpu card still a thing. just for fun
Dual GPU card, very cool, and then of course you can have two of them.
It's technically possible if it's done like what Apple did with the M1 Ultra. By connecting both chips together with such a fast connection, the system treats them as a single chip.
I know is not a consumer-grade GPU, but it isn't what the CDNA2 cards are doing?
what a majestic creature.
the cat is also nice.
Damn, I did not even know this gpu was a thing. What a beast.
I believe you must check that MPX connector to confirm EVERY power connection, I believe this is shutting down due to lack of power so yes lack of wires but also connected elsewhere. Another important detail would be the missing SMBUS drivers etc, power wise budget your cabling for at least 500watts over 12v, and check please check for additional power connections and their voltages. 5v should no be going over 75w and 3.3v max at 45/55w
I have seen some really scuffed setups on youtube to go fast, but soldering power to the card is just wild.
Ever heard of volt-modded GPUs?
You disable a card's voltage regulator, and solder a whol different VRM on like EVGA's E-POWER series.
Search for TAGG XOC he have some eye watering jaw dropping videos about modding motherboards and video cards, lots of soldering, cutting, scrapping and dremeling involved!
Fascinating video. It's very interesting to see the technology these companies hide away (through exclusive prices and availability) as the scaling against the 6800xt's performance looked, alright, weirdly.
I'd love to see a follow up to this. Its super interesting that this Apple AMD Dual gpu actually functions at all on a windows system but also completely stomps the 7900 XTX. Maybe dual gpus could make a comeback in the future using whatever technology AMD used in these modules? Clearly its working much more seamlessly than typical crossfire solutions. Though, that's just with Timespy. I wonder if the performance gain over the 7900 XT will hold true in real production applications or games.
its huge difference in setup...its base on CDNA and consumers are using different branch RDNA architecture ...CDNA is base on older tech from GCN3.0 that been renamed to CDNA...compute processing vs RDNA is old raster processing that base way older tech from ati since hd6900...hence this how far they can get versus raja koduri GCN tech can go further is unexpected for amd...
@@user78405 That's simply not true. The W6800X Pro Duo uses the same Navi 21 cores as the RX 6800/6900 family. CDNA is used in AMD's instinct cards.
Wow this was incredible amounts of effort (and patience)
Try using a current clamp to see the current
Almost impossible to remove cover, that may very well damage the product just by being removed for maintenance: literally Apple in a nutshel.
Maybe running that GPU off linux could work better?
yes please do a follow up video on this. I would love to see waterblock added to this or some sort of cooler to lock down its max performance. Very impressive
I wonder if the open source AMDGPU graphics driver for Linux would work...
I'm all for dual GPUs! 1kHz OLED monitors are absolutely possible and absolutely nothing we have today from either Nvidia or AMD can touch 4K @ 1000Hz (kHz) on a single, much less, even a double GPU.
If Nvidia brought back 4-way SLI I guarantee you people running it regardless of the expense.
Could it potentially run on Linux then?
I really wish multi gpu setups aren't a thing of the past. when I first started building I always dreamed of a crossfire setup then years down the road when I'm able to afford a dual GPU setup its pointless unless I want to show off 3dmark
This was very interesting. AMD should have brought out this card for PC use to break the nVidia dominance.
AMS has a way of shooting themselves in the foot like that.
do a follow up video pls! this is crazy good content
I think the mpx connector is for pcie and usb directly connected to the cpu for thunderbolt.
would love to see more on this!
Oh man, you gota get this fully working. Love these types of vids.