AMD Radeon Pro W6800 appears on Geekbench with 32GB of VRAM

[Random_Person1234]

Summary

The AMD Radeon Pro W6800, which features the same amounts of CU and Stream Processors as the RX 6800, but with 32GB of VRAM, has appeared on Geekbench. The graphics card's listed maximum frequency of 2554 MHz is relatively high, however it is unknown how accurate this is and how long this frequency can be sustained. This card scored 137,230 points on the OpenCL benchmark, which is higher than the normal RX 6800. This GPU is expected to be announced at a special event on June 8th.

 

Quotes

Quote

The graphics card has just appeared on Geekbench, where it was tested in the OpenCL benchmark. According to the benchmark, this SKU would feature 30 Compute Units (for AMD RDNA2 GFX1030 architecture this means 60 Compute Units), which means it will have 3840 Stream Processors, just as many as Radeon RX 6800 for gaming. A major change comes to the memory configuration though. AMD has added 32GB of memory to the graphics card, which is twice as much as Radeon RX 6800/6900 series.

Quote

The graphics card relatively high maximum core clock of 2554 MHz. However, we do not know how accurate this reading is and for how long can such a frequency can be sustained.

Quote

The graphics card scored 137230 points in the OpenCL test, which is actually higher than Radeon RX 6800 according to official charts.

 

My thoughts

Nice to see more competition in the workstation GPU segment of the market. I wonder if they will release a W6900 with even more VRAM, maybe even with HBM?

 

Sources

https://browser.geekbench.com/v5/compute/2934489

https://videocardz.com/newz/amd-radeon-pro-w6800-to-be-the-first-navi-21-graphics-card-with-32gb-memory

[emosun]

Reminds me of those 2gb HD3450s they sold back in the day with way too much vram

[AbydosOne]

6 minutes ago, Random_Person1234 said:

maybe even with HBM?

I would be surprised to see HBM in anything that already supports GDDRn. It's such a different interface that you need to design the core for it, not just stick it on as an option later.

[jaslion]

I mean it is "competition" but really its not. Nvidia currently has such a massive monopoly over almost everything workstation related as they really invested there with cuda, their optx renderer that is actually good and whatnot. Just having good hardware isn't enough here when you also need good software.

[BuckGup]

9 minutes ago, jaslion said:

I mean it is "competition" but really its not. Nvidia currently has such a massive monopoly over almost everything workstation related as they really invested there with cuda, their optx renderer that is actually good and whatnot. Just having good hardware isn't enough here when you also need good software.

I agree but AMD has been making great strides lately. Frontier the 1.5+ exaflop cluster uses all AMD MI200 cards and they are very impressive. I'm hoping that trickles down to the consumer side

[da na]

If it's remotely affordable, or if there are lower end models in the series, this card is definitely something I'll be looking out for. I've been in the market for a higher end workstation card for a while now and this piques my interest.

 

Edit: Just saw it's like what they did last year with the W5500/W5700. It was a 5500xt/5700(non xt) but with a blue shroud, better drivers, tweaked clocks, and around 10-20% higher price from most retailers than the gaming variant. I think the pricepoint for this card is promising for smaller creators like me who need a good, reliable workstation card but can't afford the obscene prices of Quadros. (Don't believe me? The Quadro 7000 was $15,000 at launch)

[igormp]

21 minutes ago, BuckGup said:

I agree but AMD has been making great strides lately. Frontier the 1.5+ exaflop cluster uses all AMD MI200 cards and they are very impressive. I'm hoping that trickles down to the consumer side

It is impressive, but still lacks any support software-wise.

My bet is that people who will work on it are already used to OpenCL, in case they're doing anything GPU-related.

[jasonvp]

As a Mac Pro 7,1 owner, I'd love to see Apple and AMD spin up an MPX version of this.  Wonder if it'll happen?

[Levent]

41 minutes ago, emosun said:

Reminds me of those 2gb HD3450s they sold back in the day with way too much vram

Pretty sure HD3xxx series cards didnt have 2G per GPU. First consumer single GPU card ever to come with 2G vram was HD4850/HD4870 2G (2008). First ever 2G Pro/consumer card was V8650 2G (2007).

[DankDoodles]

Maybe this video from this post be correlated with this new "Radeon Pro" GPU?

 

[emosun]

36 minutes ago, Levent said:

Pretty sure HD3xxx series cards didnt have 2G per GPU. First consumer single GPU card ever to come with 2G vram was HD4850/HD4870 2G (2008). First ever 2G Pro/consumer card was V8650 2G (2007).

I got the number wrong it was the 5450 that had 2gb of vram lol

[Dabombinable]

4 hours ago, emosun said:

Reminds me of those 2gb HD3450s they sold back in the day with way too much vram

It's not too much for a web browsing machine these days...even with the 5450 (mobile or desktop - 512MB sucked)

[Doobeedoo]

They need make another card that can have a NVMe SSD for it's VRAM though.

[thechinchinsong]

9 hours ago, igormp said:

It is impressive, but still lacks any support software-wise.

My bet is that people who will work on it are already used to OpenCL, in case they're doing anything GPU-related.

That's the point though then right? Even if there's not many people who are working with mostly just OpenCL, but if they are, AMD is a good option for them. Yeah, a ton of applications and people use CUDA, but its not like AMD is gonna come out with a card that does CUDA soon. Might as well still do professional cards that people are still gonna buy since non-CUDA applications still exist.

[igormp]

1 hour ago, thechinchinsong said:

That's the point though then right? Even if there's not many people who are working with mostly just OpenCL, but if they are, AMD is a good option for them. Yeah, a ton of applications and people use CUDA, but its not like AMD is gonna come out with a card that does CUDA soon. Might as well still do professional cards that people are still gonna buy since non-CUDA applications still exist.

AMD does have a half-assed try at getting CUDA applications to work on their GPUs with HIP, but it's still not good enough (nor anything on the entire ROCm stack other than OpenCL itself).

[thechinchinsong]

21 minutes ago, igormp said:

AMD does have a half-assed try at getting CUDA applications to work on their GPUs with HIP, but it's still not good enough (nor anything on the entire ROCm stack other than OpenCL itself).

Yeah, I thought so. AMD is never gonna be able to compete using CUDA or CUDA emulation, but it might as well still tap into the market of non-CUDA professionals which even if small, is probably still a decent size for AMD to target, seeing as it is a smaller volume company compared to Nvidia or Intel anyways.

[leadeater]

52 minutes ago, thechinchinsong said:

AMD is never gonna be able to compete using CUDA or CUDA emulation

AMD hasn't and won't do CUDA emulation, they have done CUDA translation to HIP and/or OpenCL code as mentioned but that's EXTREMELY hard to do while also outputting optimized HIP/OpenCL code for the original purpose of the code. There aren't like for like optimizations between CUDA and OpenCL as they are very different technologies and optimizations are done for hardware specific reasons so when you're translating for a different target hardware actually doing it with the best possible performance is really really hard.

 

AMD simply has to improve the software ecosystem for OpenCL to compete, like Nvidia did with CUDA.

[Jet_ski]

1 hour ago, leadeater said:

AMD simply has to improve the software ecosystem for OpenCL to compete, like Nvidia did with CUDA.

Totally. OpenCL is great and all but nobody is going to do AMD’s work for them. They just have to invest in the software side of their business. Hopefully now that they are making tons of money, they’ll allocate the budget for it.

[jaslion]

22 hours ago, igormp said:

It is impressive, but still lacks any support software-wise.

My bet is that people who will work on it are already used to OpenCL, in case they're doing anything GPU-related.

Pretty much this amd cards have always been pretty darn good in opencl cases. Also a big cluster cannot be compared to normal desltop/workstations or render farms as those are made to run common software whilst a cluster like this has the software made for it usually so then drivers matter a lot less. I do hope they get better soon with it to just have another usefull option.

[jaslion]

15 hours ago, Jet_ski said:

Totally. OpenCL is great and all but nobody is going to do AMD’s work for them. They just have to invest in the software side of their business. Hopefully now that they are making tons of money, they’ll allocate the budget for it.

The bonus about openCL is that well any card can do it quite well amd is just focused on it and does it better than others for now so if they do invest into it they can make it at least a valid competitor for all tasks where Cuda is the main option now. They've taken over the console space. Are taking over the server cpu space and are steadily rising in rank on the desktop and laptop cpu space. So next it is time to truely make their purchase of ATI worth it and become an actual big player instead of the number 2 by a very large gap. The first GCN launch was a great step where they started going wrong later on, vega was well vega and sort of worked out but now with their rx lineup they are becoming an actual first pick for gamers again instead of a you can always get an amd card vibe.

[igormp]

5 hours ago, jaslion said:

has the software made for it usually so then drivers matter a lot less. I do hope they get better soon with it to just have another usefull option.

Indeed. Drivers are okay for that case (AMDGPU is pretty much rock solid nowadays), but framework support and whatnot is what kills them.

Unless you're building your own stuff from scratch (which happens a lot in HPC scenarios, tbh), you're out of luck when it comes to the big frameworks (be it for data processing or ML).

[cj09beira]

On 6/4/2021 at 5:43 PM, AbydosOne said:

I would be surprised to see HBM in anything that already supports GDDRn. It's such a different interface that you need to design the core for it, not just stick it on as an option later.

navi 21 seems to support hbm, we figured this out by looking at linux kernel code, so it is possible, not sude they would launch it though