A dual GPU RX580 AMD never released? Radeon Pro 7350x2 Review

cagoblex

n00b
Joined
Aug 28, 2020
Messages
63
What is the last time you see a high-end dual GPU consumer video card? For team green it was Titan Z released back in 2014. For team red it was the Fiji based Radeon Pro Duo, not to mention that it’s technically not a consumer grade card.

Today let’s take a look at something interesting. The Radeon Pro V7350x2.
v7350.jpg.b608f9801682cd0f94e445dda7699a2c.jpg
It is a dual slot, passively cooled card, with only one mini Displayport output. It is designed to be used in GPU servers with high airflow. So what is this thing? Techpowerup has this card in their database. It is a Polaris 10 based card paired with 16GB of GDDR5 memory for each GPU. That makes a total of 32GB GDDR5 memories on board.


powerup.thumb.PNG.58aa2fc5c63dbe3dc1b563b6fda14006.png

echspec.thumb.PNG.9751ac9c827fd9c180a9fd0c0fddde9e.png
It is basically two RX580 on a single GPU. It has exactly double the specs except for the TDP. It has a 6pin and a 8pin power connector. And it’s rated at a 200W TDP. This is very impressive considering even a single Polaris 10 GPU would consume 185W of power.



peccomparison.png.32c012c1d77cb9b68f299c0a7a81f503.png
Let’s start by taking the card apart. Here is the front of the PCB. It has a very similar layout as the Polaris based Radeon Pro Duo but it has a shorter PCB. Unlike early over engineered ES cards, this cards actually looks pretty close to retail version. It has two Polaris GPU with no markings on it, at 45 degree angle.



MG_4532.thumb.jpg.021486d540e79bdb533615f8ec061814.jpg


Here is the back of the PCB, the only indication of it being an Engineering Sample card is the MicroSEMI FPGA card on the back for debugging purposes.


MG_5727.thumb.jpg.3a45daab8a675bb6cb6db612f50f334b.jpg



debug.jpg.267b552b9afcde355cd8704d090b86c3.jpg
Let’s plug it in and see if it works. To my surprise, the driver installation is pretty painless. Unless the unreleased Nvidia cards, this card works with default AMD driver without having to ticker with anything. The driver picks up the video card right away and shows the correct model in the driver panel.
device.PNG.a13c9b694ef58304f6657bd4556e6e60.png



control.PNG.e5695dfe6f5a4c7d8375cf81be77601b.png


Now it’s time for benchmark. In comparison, we test it against the Titan Z, since that’s the last dual GPU consumer card Nvidia has to offer.

_titanz.thumb.PNG.aad0a80aa5bc8eb1295bbbb0df82555a.png

Here you can see the benchmark results. Please note, I’m not really cooling it with the appropriate way, so the result is just FYI. It is thermal throttling quite a bit.

In 3Dmark Timespy, it is about 20% faster than a Titan Z, and remember that’s with thermal throttling, according to the seller I got this card from, it is able to achieve about 1000 points higher if you cool it properly.


3dmark.thumb.png.12c54a8b4fd71663b9033bb9d08e1d7b.png

In actual game testings, it has a significant lead over Titan Z.

hitman.thumb.png.01ff026ea529b91a82b1b8735cb7965f.png


metro.thumb.png.dddb180c8266a2f0290d32f201f3dab9.png

In Geekbench 5 compute testings, it almost doubles the score of Titan Z in OpenCL, but it’s still considerably slower than the GTX1080.

ekbench.thumb.png.964fbb9606bd885b38cf708410408fbe.png

This concludes part one of the review, in the next part I will work on a better cooling solution, try to overclock it, and compare it against two RX480 cards in more professional workloads.

You can check out more about this card in this Youtube video:




Thanks for reading.
 
This is probably a variant on the Radeon Pro Duo Polaris card, IIRC with 32GB ram...
https://www.techpowerup.com/gpu-specs/radeon-pro-duo-polaris.c2953
Yes they have the same specs and similar PCB layouts. However the 7350x2 has slightly shorter PCB and only one display output. Using Nvidia for comparison, Radeon Pro is more like a Titan level card, but the V7350X2 is more like a Tesla card. They have the same underlying GPU architecture but they are aimed at different market segments.
 
Or the S7150x2 - which was out for a bit for the same reason (MxGPU).
Yes it is indeed a replacement for 7150X2. It has GCN4.0 architecture instead of GCN3.0 on 7150x2. However it was cancelled at the very last minute for some reason that we will never know.
 
Yes it is indeed a replacement for 7150X2. It has GCN4.0 architecture instead of GCN3.0 on 7150x2. However it was cancelled at the very last minute for some reason that we will never know.
Not a useful card. MxGPU didn’t include real time H.264 encoding; the cards were not useful for their intended purpose. We sold and then returned a lot of them- didn’t have the necessary compatibility for almost any real use case.
 
Not a useful card. MxGPU didn’t include real time H.264 encoding; the cards were not useful for their intended purpose. We sold and then returned a lot of them- didn’t have the necessary compatibility for almost any real use case.
Exactly. Even Intel did a better job in terms of real time H.264 encoding. I'm working on a review for an Intel VCA 2 card which does a pretty good job at this.

And they are very hot at 200W TDP. I use the same thermal solution for Nvidia K80 which has 300W TDP, and they run at about the same temperature.
 
Exactly. Even Intel did a better job in terms of real time H.264 encoding. I'm working on a review for an Intel VCA 2 card which does a pretty good job at this.

And they are very hot at 200W TDP. I use the same thermal solution for Nvidia K80 which has 300W TDP, and they run at about the same temperature.
Well, it was totally missing - Nvidia had vGPU with NVENC for sending a desktop stream as far back as Kepler (K1/K2 cards). they made it licensed software in Maxwell (M10/M60), which drove AMD to try to enter the market - but without the hardware encoder, you did the blast encoding in CPU - which cut your guest ratio by 33%, and you had to have higher clock speeds to do that at 30FPS (2.6ghz minimum to be safe), so that eliminated the higher core count CPUs... and so on. Their most recent ones were supposed to have it, but I haven't ever SEEN one, and MxGPU using SR-IOV was also a limitation, vs a software division.

I architect this stuff for a living - AMD cards were something we really hoped would rock, but were really hard to utilize.
 
Well, it was totally missing - Nvidia had vGPU with NVENC for sending a desktop stream as far back as Kepler (K1/K2 cards). they made it licensed software in Maxwell (M10/M60), which drove AMD to try to enter the market - but without the hardware encoder, you did the blast encoding in CPU - which cut your guest ratio by 33%, and you had to have higher clock speeds to do that at 30FPS (2.6ghz minimum to be safe), so that eliminated the higher core count CPUs... and so on. Their most recent ones were supposed to have it, but I haven't ever SEEN one, and MxGPU using SR-IOV was also a limitation, vs a software division.

I architect this stuff for a living - AMD cards were something we really hoped would rock, but were really hard to utilize.
I see your point. I haven't had any experience with AMD in terms of virtualization but I agree Nvidia did a great job on this, both in high density environments and high performance environments. The only bad thing however, is that you have to pay the software licensing fee for Nvidia vGPU per client. I am working on a review of Tesla P10, which is also never released. I assume it would perform much better than the v7350x2.
 
Oh yes. VGPU licensing sucks... but at the same time, they wrote it- and it rocks. Hell, you can even live migrate then now.
 
Would the AMD Radeon Pro Vega II Duo in the Mac Pro count as dual GPU release, albeit one for a specific client? Intel and AMD have released other defacto-Apple-only SKUs over the years (midrange Navi with HBM, Intel CPU with sideport memory/cache).
 
Would the AMD Radeon Pro Vega II Duo in the Mac Pro count as dual GPU release, albeit one for a specific client? Intel and AMD have released other defacto-Apple-only SKUs over the years (midrange Navi with HBM, Intel CPU with sideport memory/cache).
Well technically yes. But since they do not have the start PCI-E connector so I wouldn't consider as a "public released product". For AMD and Intel I guess yes, actually I'm using a Xeon W-2150B to type this reply right now, which is a customized Apple SKU from Intel. Although they are kinda hard to get, but they do in fact work in other consumer motherboards.
 
Oh yes. VGPU licensing sucks... but at the same time, they wrote it- and it rocks. Hell, you can even live migrate then now.
I don't even know that...Maybe I would need your help on the Intel VCA2 if possible. I'm having a really hard time making it to work through virtual desktop.
 
Back
Top