AMD Radeon Pro W7900 Workstation Graphics Card Spotted

[GFreeman]

According to the latest leak, it appears that AMD will soon launch its Navi 31-based Pro W7900 workstation graphics card. Currently, AMD's workstation graphics card lineup is limited to RDNA 2-based Radeon Pro W6000 series, with Radeon Pro W6800 as the flagship. AMD also has the Radeon Pro W6900X, but this was limited to Apple's systems. Now, it appears that AMD is preparing to launch RDNA 3-based Radeon Pro W7000 series, as the Radeon Pro W7900 was spotted in PugetBench database.

The Radeon Pro W7900 managed to score 135.3 in GPU score, putting it in line with the Radeon Pro W6900X in the same benchmark, which scored around 138. Of course, bear in mind that one was running on Mac Pro, while other was running on Windows 11 OS, so we'll wait to see the official specifications and reviews to get some better idea regarding the actual performance.



View at TechPowerUp Main Site | Source

 

[Punkenjoy]

Strange, the 32 GB of VRAM suggest a 256 bit memory bus, a chip yet to be released by AMD or just a rebranding of NAVI 21?

 

[HBSound]

I am incredibly excited about this announcement. I wanted to try the last-generation pro card, but no one offered a water block.
Hopefully, this time around, someone does, and I will professionally skip the PNY A card to the side.

 

[Vya Domus]

Strange, the 32 GB of VRAM suggest a 256 bit memory bus, a chip yet to be released by AMD or just a rebranding of NAVI 21?

You don't have to use all of the memory interfaces on a chip, even though Navi 31 has a 384 bit bus they can totally use just a portion of it. So this could be using 8 double density chips (32*8 = 256 bits), I don't know if this really is the case, I am just saying it's plausible.

 

[Haile Selassie]

You don't have to use all of the memory interfaces on a chip, even though Navi 31 has a 384 bit bus they can totally use just a portion of it. So this could be using 8 double density chips (32*8 = 256 bits), I don't know if this really is the case, I am just saying it's plausible.

Ring bus consists of 6x 64bit channels. Half of the IC is double density without any problem to the memory controller.

 

[Punkenjoy]

Indeed, they could just use 4 of the 6 MCD, but that still feel strange to me. They either decided that more bandwidth was useless for the workload they aim for. Or maybe it's Navi 32, or even maybe it's still Navi 21. Who know.

 

[HBSound]

Is there a reason any one intentionally uses the AMD Pro over the PNY A series of GPUs?

 

[Vya Domus]

Ring bus consists of 6x 64bit channels. Half of the IC is double density without any problem to the memory controller.

I don't how RDNA3 is configured but the 7900XTX has 12 memory chips, so every memory chip has to be connected by a 32 bit interface, either way the point is they can use the same chip. If you want 32 GB you can get that with 8 x 32 gigabit modules, I don't think they'd bother with an asymmetric configuration.

 

[Wirko]

I don't how RDNA3 is configured but the 7900XTX has 12 memory chips, so every memory chip has to be connected by a 32 bit interface, either way the point is they can use the same chip. If you want 32 GB you can get that with 8 x 32 gigabit modules, I don't think they'd bother with an asymmetric configuration.

By asymmetric I understand that 4 channels have 2 RAM dies each, and the other 8 have one die each, all dies are of the same capacity but maybe some in 2-layer packages, is that what you mean?
So why not do that? By using all channels the card should achieve greater bandwidth on average, even if inconsistent, than with 8 symmetrical channels. I don't see big technical obstacles here.

 

[laptop-hpc]

Memory
32GB (2x16GB) 4800MHz

That is the only mention of memory anywhere in the screenshot or in the sources, which is just from PugetBench listing the system RAM. So far no confirmation of VRAM amounts.

 

[Vya Domus]

By asymmetric I understand that 4 channels have 2 RAM dies each, and the other 8 have one die each, all dies are of the same capacity but maybe some in 2-layer packages, is that what you mean?
So why not do that? By using all channels the card should achieve greater bandwidth on average, even if inconsistent, than with 8 symmetrical channels. I don't see big technical obstacles here.

It's just not something that's usually done, I can't think of any example of a card mixing memory modules of different densities, I don't know if it's even possible.

That is the only mention of memory anywhere in the screenshot or in the sources, which is just from PugetBench listing the system RAM. So far no confirmation of VRAM amounts.

Yeah and the amount of VRAM that would make most sense is actually 48GB.

 

[Haile Selassie]

It's just not something that's usually done, I can't think of any example of a card mixing memory modules of different densities, I don't know if it's even possible.

Very popular GTX 970 was one of them.

 

[Chrispy_]

I am incredibly excited about this announcement. I wanted to try the last-generation pro card, but no one offered a water block.
Hopefully, this time around, someone does, and I will professionally skip the PNY A card to the side.

Is there a reason any one intentionally uses the AMD Pro over the PNY A series of GPUs?

You seem to be confusing graphics card manufacturers with GPU manufacturers. PNY don't make GPUs at all, and they use exclusively Nvidia GPUs in their cards. Since Nvidia and AMD pro/workstation GPUs do not have the same API support or features, you'd be unlikely to use one in place of the other; You wouldn't try to cross an ocean in a car, the same was as you wouldn't try to get down the highway in a boat.

• If your applications require CUDA, RTX, or other Nvidia proprietary APIs, then you can't use a Radeon Pro.
• If your applications do not require CUDA, RTX, or other Nvidia proprietary APIs, Radeon Pros offer around 2.5x more performance/$ and far more VRAM for equivalent price points.

Not having the API support you need is a deal-breaker.
Not having enough VRAM to complete your task is a deal-breaker.

You'd pick the card you need for your workflow and there's not a huge amount of crossover between them unless you happen to have unlimited funding, in which case just get the biggest and most expensive RTX A6000 (or whatever will replace it) and ignore the fact that it's obscenely expensive for what it is.

 

[Vya Domus]

Very popular GTX 970 was one of them.

All memory chips were the same density as far as I know, it was a different issue with that.

 

[Chrispy_]

All memory chips were the same density as far as I know, it was a different issue with that.

Yeah, last asymmetrical RAM density cards from Nvidia were Kepler in 2012:
The 660 and 660Ti had four 256MB chips and two 512MB chips each, for 2GB total, but only the first 1.5GB was 192-bit, the remaining 512MB was 64-bit.

The GTX970 issue wasn't RAM density, but that the 970 was trying to squeeze 8 VRAM chips into 7 memory channels - should really have just been officially a 3.5GB card.

 

[HBSound]

You seem to be confusing graphics card manufacturers with GPU manufacturers. PNY don't make GPUs at all, and they use exclusively Nvidia GPUs in their cards. Since Nvidia and AMD pro/workstation GPUs do not have the same API support or features, you'd be unlikely to use one in place of the other; You wouldn't try to cross an ocean in a car, the same was as you wouldn't try to get down the highway in a boat.

• If your applications require CUDA, RTX, or other Nvidia proprietary APIs, then you can't use a Radeon Pro.
• If your applications do not require CUDA, RTX, or other Nvidia proprietary APIs, Radeon Pros offer around 2.5x more performance/$ and far more VRAM for equivalent price points.

Not having the API support you need is a deal-breaker.
Not having enough VRAM to complete your task is a deal-breaker.

You'd pick the card you need for your workflow and there's not a huge amount of crossover between them unless you happen to have unlimited funding, in which case just get the biggest and most expensive RTX A6000 (or whatever will replace it) and ignore the fact that it's obscenely expensive for what it is.

Thank you so much for the reply. I currently use the A6K and get great performance - no issues at all. I wanted to try the W6800 Pro, but EKWB or not one else provided a water block for the W6800. My primary use - CAD/Civil 3D/Revit/BIM/Rendering

 

[Chrispy_]

CAD/Civil 3D/Revit/BIM

Will run happily on a potato for most software, since they're widely optimised for thin-clients and Remote Desktop Connection. It's been a while since the "pro" tax has provided any meaningful gains outside of a few wallet-critting CAM suites. Autodesk stuff actually works close to optimally on integrated graphics these days (now that IGPs aren't complete junk any more)

Rendering

This is where an A6000 might be necessary. Not for the drivers, or the features, or any of the other nonsense - but because it's the only way to get >24GB VRAM and if your scene requires more than 24GB you have to use an A6000 or fallback to software rendering which isn't usually an option as it'll take hundreds of times longer to render.

 

[HBSound]

Will run happily on a potato for most software, since they're widely optimised for thin-clients and Remote Desktop Connection. It's been a while since the "pro" tax has provided any meaningful gains outside of a few wallet-critting CAM suites. Autodesk stuff actually works close to optimally on integrated graphics these days (now that IGPs aren't complete junk any more)

This is where an A6000 might be necessary. Not for the drivers, or the features, or any of the other nonsense - but because it's the only way to get >24GB VRAM and if your scene requires more than 24GB you have to use an A6000 or fallback to software rendering which isn't usually an option as it'll take hundreds of times longer to render.

Thank you so much. I will hold on to what I have.
Thank you

 

[RadeonProVega]

If the W6800 is still selling for over $1000, I'm sure the w7900 will sell around $2000 lol.
Next GPU I'm buying is a AMD Radeon Pro VII 16GB card.

 

[HBSound]

I can never find any professional water blocks for any of those AMD cards. The gamers' cards have water blocks, but they are 2 and 3 PCI slots wide. The EKWB pro water blocks for the Nivida AX series of cards are all single-lane.

 

[Haile Selassie]

Yeah, last asymmetrical RAM density cards from Nvidia were Kepler in 2012:
The 660 and 660Ti had four 256MB chips and two 512MB chips each, for 2GB total, but only the first 1.5GB was 192-bit, the remaining 512MB was 64-bit.

True, also one of such designs is XboX Series X:
10 memory ICs for a total of 16GB RAM - 10GB on a 320bit wide bus and 6GB on a 192bit bus.

 

[Wirko]

Yeah, last asymmetrical RAM density cards from Nvidia were Kepler in 2012:
The 660 and 660Ti had four 256MB chips and two 512MB chips each, for 2GB total, but only the first 1.5GB was 192-bit, the remaining 512MB was 64-bit.

The GTX970 issue wasn't RAM density, but that the 970 was trying to squeeze 8 VRAM chips into 7 memory channels - should really have just been officially a 3.5GB card.

It's still better to use all the available channels from the GPU chip than to disable some of them just to make the VRAM arrangement symmetrical.
I wasn't aware that asymmetrical configs were so rare. Did the 660 actually have two different RAM dies or were those just different packages (stacked and non-stacked)?

Yeah and the amount of VRAM that would make most sense is actually 48GB.

It's totally possible that the PCB is designed for 24 chips but the W7900 has 8 spaces left empty.

 

[Chrispy_]

It's totally possible that the PCB is designed for 24 chips but the W7900 has 8 spaces left empty.

I'm expecting the W7900 to be a double-density variant of the 7900XT or XTX, so either 40 or 48GB VRAM.

This whole thread is a trainwreck about VRAM size because the very first comment was someone misinterpreting 32GB of system RAM in the test PC as VRAM, and people don't read the thread.

TL;DR we don't know how much VRAM it'll have. Probably 48GB, but also possibly 40GB, 24GB, 20GB too!

I wasn't aware that asymmetrical configs were so rare. Did the 660 actually have two different RAM dies or were those just different packages (stacked and non-stacked)?

They're rare now because they're a driver nightmare, but it was a thing for the 550Ti as well. Nvidia tried it and it gave their driver team additional workload, stuttering performance if the VRAM usage spilled into the lower-bandwidth region, and didn't really save any money, and so they stopped doing it. I believe it was just a marketing tactic as the competition from AMD was all 256-bit 2GB cards and 1.5GB looked weaker for the average joe comparing GPUs on a store shelf.

The 550Ti, 660, and 660Ti all share the asymmetrical layout but it was handled differently between generations. The 550Ti had six GDDR5 packages of mixed density, the 660 and 660Ti had 8 identical packages where four of them competed for bandwidth on two of the controllers:

(credit, Anandtech)



In theory, there is no real performance difference between these two approaches since data can only be striped/strided across three-quarters of the VRAM before having to access the remaining VRAM over just 2 channels instead of 6. Whichever way you look at it, the "extra" VRAM beyond what you'd expect was hindered by only having a third of the expected bandwidth and therefore used as a last resort by the driver. It may as well have not been there because by the time games needed all the VRAM, they needed all the bandwidth too.

 

[Punkenjoy]

Yep my bad!

But still a nice conversation about bus size, memory package etc.

 

[Vya Domus]

https://www.amd.com/en/graphics/workstations?fbclid=IwAR3Gi9rUT-cDftmGwvqS4Jo0cCVQeOV3hZ5aIYrH40wGGKBia2NtTA7hrTc

They've officially announced these cards and as expected the W7900 is fully enabled and has 48GB, the W7800 has 32GB and only a 256 bit interface, so no asymmetric strangeness just double density modules.