AMD RX Vega

So RX Vega is around the corner. At least professional one. To be discussed in more detail at the upcoming Computex at the end of this month.

I wanted to start a discussion what can we expect.

AMD official announced :

VX Vega FE (Frontier Edition)
-64 NCU (next-gen compute units)
-4096 shader processors
-16GB HBM2
-12.5 TFLOPS FP32 @ 275w? I hope so
-25 TFLOPS FP16

For direct comparison with other AMD GPUs

[TABLE=“class: grid, width: 500, align: center”]

GPU
TFLOPS FP32
Power
Blender
Blender
LuxMark
LuxMark

Blender BMW (1 car)
Perf Per Watt (*100)
LuxMark 3.1 (hotel)
Perf Per Watt
Price (usd) @ newegg
Perf per dollar
Source

Radeon VEGA Frontier Edition
12.5
300W
53s
17.6
4690
15.63
1000
4.69
https://www.pcper.com/image/view/833…n=node%2F68037

RX VEGA 64 (water)
13.6
345W

https://www.pcper.com/reviews/Graphics-Cards/AMD-Radeon-RX-Vega-64-and-Vega-56-Specs-Prices-Power-Detailed

RX VEGA 64
12.7
295W

https://www.pcper.com/reviews/Graphics-Cards/AMD-Radeon-RX-Vega-64-and-Vega-56-Specs-Prices-Power-Detailed

RX VEGA 56
10.5
210W

https://www.pcper.com/reviews/Graphics-Cards/AMD-Radeon-RX-Vega-64-and-Vega-56-Specs-Prices-Power-Detailed

RX VEGA Nano
???
150W

??? just screenshots across the web and in the official rx vega capacian event, one was handed out to Tim Swinny? or whats his name… either way… no direct info.

Radeon Pro Duo (single gpu = R9 Nano)
8.1
175W

3331
19.03
450 (for nano)
7.40
[TABLE=“class: cms_table_grid, width: 500, align: center”]
[TR]
[TD]https://www.pcper.com/image/view/833…n=node%2F68037

[/TABLE]
[/TD]
[/TR]

Radeon Pro Duo (Fiji - dual R9 nano)
16.38
350W

7260
20.74
1175 (ebay)
6.17

Radeon Pro Duo (Polaris - dual RX 480)
11.5
250W

RX 480
5.8
150W
65s(?)
43.3
2758
18.38
180 (25)
15.32 (11.03)
https://hothardware.com/reviews/amd-…et-spot?page=4

R9 Fury X
8.6
275W

3529
12.83
500?
7.05
http://www.anandtech.com/show/11180/…0-ti-review/14

GTX 1080TI

250W
54s
19.28
5288
21.15
800
6.61
http://www.anandtech.com/show/11180/…0-ti-review/14

[/TABLE]

So seeing how it fits against current performance of RX 480 (with latest builds) that will be one powerful singe GPU card.

Nvidia TFLOPS ratings are not equivalent for direct comparison but also OpenCL is not mature on NVidia cards as compared to AMD’s.

Thoughts?[TABLE=“class: cms_table_grid, width: 500, align: center”]

https://www.pcper.com/image/view/833…n=node%2F68037

[/TABLE]

Going to wait for reviews and benchmarks before making any decisions. It’s interesting to see if and how well the card is going to perform on deep learning. Nvidia has a huge head start on that front.

If you have the money, it’s better to go for a pro duo indeed, but the Vega should have about the same performances (less Tflops in FP32, but it will have faster FP16 and it will have a faster memory) for half the price (500€ instead of 1000€). Also the consumption may be better as it’s a single GPU card = 2x less physical memory for the same available memory and also HBM is much more energy efficient than GDDR5.
But better decide after benchmarks have been done, hopefully on buildbots and not outdated 2.78c, but few benchmark sites know about Blender development sadly… So I don’t know how representative of real performance it will be. Best will be results from forum users.

If you are deciding to get the Pro Duo or the Vega, and willing to buy the Duo at 2x the cost, just buy 2 Vegas.

Not really interested in cost or deciding to get Pro Duo (only 4gb per core is a bit limiting on larger scenes). There is also Pro Duo Polaris cards with 16GB of memory. a bit better but performance is sub par over its predecessor with Fiji GPUs.

Just more trying to figure out Vega, and where it stands when it comes to performance in Blender.

Price wise, the Pro Duo in certain shops is nearly 1/2 price. So it is more equivalent to the cost that Vega might be.

@bliblubli - I really like what you mentioned about benchmarking. Sites that use Blender will use the official blender builds, showing a significant speed deficiency of the AMD cards. I really hope 2.79 comes out before then though that is a non chance…

Awesome guys :slight_smile:

Big number are good for Blender. I think most of the reviews are way to concentrated on gaming and that has nothing to do with cycles.

Two rx 480 is 11.6 Tflops and the same is for one 1080 TI. The cycles results are practically the same.

1200 and 1700 USD for AMD VEGA FE accordingly for air and liquid cooling in preorder

Best Regards!
Albert

bigbad… do you mean that Rx 480 performance is on par with 1080ti? Not sure I understood your post…

I would like to see driver support first. For some reason I can’t use Cycles with my AMD gpu.

2xRX480=1080Ti

Ahh, that makes more sense to me.

Now the bigger problem is AMD cards for crypto mining get these RX cards cost nearly as much as one 1080 (non-ti)… crazy times we live in.

Hopefully AMD can ensure good supplies when it comes out…

I find it funny that AMD will maybe surpass nvidia in a half year because of mining. There is nowhere to buy amd stuff in sweden. That’s great because now more open standards will compete with nvidia.

What is this crypto mining you’re talking about ?

It’s not gonna change much really. This is actually the third time that this has happened, the last time AMD overproduced chips that ended up not being in demand by the time they shipped.

Here’s the deal: Everyone is buying up GPUs to mine for a fixed amount of digital currency. Within a short amount of time, it will be unprofitable again and then those GPUs may end up on ebay, eating into the profits for new GPUs. That’s why you see both AMD and NVIDIA producing crypto-cards without display outputs.

If the crypto-market crashes (as it always does when there’s a craze like now), the whole thing might be over very quickly. If you have one of these AMD GPUs right now, the smartest thing might be to sell it.

It’s when you pretend you’re rugged and butch and work in a coal mine when really you’re quite effeminate and work as a receptionist.

This is the last gen pro duo with 4GB of HBM per GPU

Would be useful if someone posts performance of AMD’s Vega Frontier Edition on Cycles rendering. Also, multi gpu performance will be interesting. I expect the FE cards to be significantly faster than the 1080 ti, based on their FP16 performance. Let’s see…

Performance is supposedly going to be bad (just like with most other benchmarks), because AMD reportedly hasn’t finished writing their drivers for Vega yet.

I expect the FE cards to be significantly faster than the 1080 ti, based on their FP16 performance. Let’s see…

Cycles doesn’t use FP16 and it would require changing lots and lots of datatypes in the code. For geometric operations, FP16 is not going to be enough. It’s enough for some shader operations, but the results are going to be slightly different. Does the user get a choice? Compared to the overhead of interpreting shaders, the speedup from FP16 is likely modest anyway.

I googled and blender mostly uses FP32. FP32 is quite high on the vega card. It should perform really good on blender. Also the opengl is fast on vega also. Great news for eevee rendering.

I’m no expert on compiling but couldn’t the compiler make the convertion FP32 to FP16 while compiling. I’d assume it will break