AMD WX 8200

LaurV · 2019-01-06, 16:45

Anybody had the chance to benchmark a WX-8200? I have the opportunity to get one below $1k and I would like to know if it worth the try. There are no numbers on James' site for it, and neither close to it, with ~700 GFLOPS DP (double compared with 1080Ti/2080Ti), and 10 TFLOPS SP (about the same as 1080Ti/2080Ti), and 2048 bits memory (this seems to be like high Tesla/Volta cards, 6-7 times wider than a "normal" GTX/Titan card) this should score something like over 100GHzD/D of LL and over 1THzD/D of TF. Of course, I know I am comparing pears and oranges, but if I am close with the assumption, that it would worth the money, compared with a Titan RTX or Titan V100 which costs triple.

nomead · 2019-01-06, 17:06

Just looking at the listed specs suggests it is somewhere between the consumer Vega 56 and Vega 64. Same core config as on Vega 56 but clocked a bit faster, same amount of HBM2 memory but clocked higher, and a higher TDP.

Also the Titan RTX and Titan V (Tesla V100) are vastly different things.

So just a guess, by looking at how the Vega cards have performed in benchmarks. LL performance should be slightly less than a Titan RTX, but vastly slower than a Titan V / Tesla V100 because those Volta chips have lots of DP performance. For some reason the higher DP GFLOPS rating doesn't seem to translate directly into LL performance.

And in factoring the Titan RTX and Titan V are pretty much at the same level, and that is about 3.5x faster than an ordinary Vega 64.

From purely a performance perspective, I wouldn't think it's worth any more than a Vega 64. But there are other reasons why these workstation cards exist.

LaurV · 2019-01-06, 17:31

Thanks. I will have to dig more into this.

The "v100" was a typo. I was not thinking to Tesla V100 (which costs in the $9-10k ballpark) but to Titan V which costs <$3k (and yes, I know, no ECC RAM, etc).

preda · 2019-01-07, 07:39

I'd look here https://en.wikipedia.org/wiki/List_o...ocessing_units for a comparison to Vega64. At first look, seems pretty similar to consumer Vega64 in power and memory bandwidth. (1/16 DP). May have ECC RAM (I don't know), but that's not such an important factor anymore. Probably a better deal would be a Vega64 at half price.

nomead · 2019-01-07, 08:13

Yes it has ECC support. That's usually the main distinguishing characteristic between consumer and workstation products. Also different display connectors, but that's again irrelevant when using it for computation only.

https://www.amd.com/en/products/prof...on-pro-wx-8200

LaurV · 2019-01-09, 03:14

Quote:

Originally Posted by preda

Probably a better deal would be a Vega64 at half price.

Thanks. That is what I "needed" to hear. (Not as in "wanting" to hear, but better safe than sorry - also not as in "expected" to hear, because actually I had no expectations in either direction, I don't know much about AMD cards, I only "hoped" (or is it "hopped"?

) that the WX Pro cards are better than "normal" cards. But it seems they are not different from Nvidia cards is this respect, "Pro" generally means safer, but not higher performance.

M344587487 · 2019-01-09, 08:25

Quote:

Originally Posted by LaurV

...it seems they are not different from Nvidia cards is this respect, "Pro" generally means safer, but not higher performance.

Some generations ago the DP ratio was different but not a lot else, you pay double largely for the fabled validation. The next generation looks set to diverge again, rumour has it that there'll be a consumer 7nm Vega stop gap that's not a cut-down version of the workstation MI50/MI60.

LaurV · 2019-01-09, 09:05

Yeah, right.

What took my eyes and inflated my imagination was the "2048 bit memory interface" (compared with 352 known from cuda cards), I went like "Waaaaa !" (see my avatar on the left). But it seems this is common too, for AMD cards, they use a different approach.

M344587487 · 2019-01-09, 13:35

HBM is the difference and it's the direction all compute cards will likely go eventually as it can offer more bandwidth, some nvidia cards like the V100 use HBM2. AFAIK Vega was the first to use HBM2 in any meaningful quantity on consumer cards but it's still relatively expensive, we're unlikely to see it in cheaper consumer cards anytime soon. I'm hoping a consumer Vega refresh is coming as it might be good for us, HBM2 bins have improved a lot since Vega was released and a 7nm Vega should be a lot more energy efficient. An MI50/MI60 would be nice but I don't have enough kidneys to sell.

firejuggler · 2019-01-09, 15:23

I might be late to the party but

https://hothardware.com/reviews/amd-...pro-gpu-review

LaurV · 2019-01-10, 13:28

Thanks. Nice.

Articles similar to the one you posted is what sparked this thread in the first place, stimulated by the fact that we bought some other hardware and had the opportunity to get this for a cheaper price. But after more web studies and replies in this thread, we also have seen that this card is "usually" under $1k, so our trade would not be as discounted as we assumed - as said, our reference point for comparison (expensive cuda professional cards like Tasla or Quadro) was wrong from the start. We decided not to take it.

2019-01-06, 16:45	#1
LaurV Romulan Interpreter "name field" Jun 2011 Thailand 41·251 Posts	AMD WX 8200 Anybody had the chance to benchmark a WX-8200? I have the opportunity to get one below $1k and I would like to know if it worth the try. There are no numbers on James' site for it, and neither close to it, with ~700 GFLOPS DP (double compared with 1080Ti/2080Ti), and 10 TFLOPS SP (about the same as 1080Ti/2080Ti), and 2048 bits memory (this seems to be like high Tesla/Volta cards, 6-7 times wider than a "normal" GTX/Titan card) this should score something like over 100GHzD/D of LL and over 1THzD/D of TF. Of course, I know I am comparing pears and oranges, but if I am close with the assumption, that it would worth the money, compared with a Titan RTX or Titan V100 which costs triple. Last fiddled with by LaurV on 2019-01-06 at 16:46

2019-01-06, 17:31	#3
LaurV Romulan Interpreter "name field" Jun 2011 Thailand 41·251 Posts	Thanks. I will have to dig more into this. The "v100" was a typo. I was not thinking to Tesla V100 (which costs in the $9-10k ballpark) but to Titan V which costs <$3k (and yes, I know, no ECC RAM, etc). Last fiddled with by LaurV on 2019-01-06 at 17:31

2019-01-06, 17:06	#2
nomead "Sam Laur" Dec 2018 Turku, Finland 317 Posts	Just looking at the listed specs suggests it is somewhere between the consumer Vega 56 and Vega 64. Same core config as on Vega 56 but clocked a bit faster, same amount of HBM2 memory but clocked higher, and a higher TDP. Also the Titan RTX and Titan V (Tesla V100) are vastly different things. So just a guess, by looking at how the Vega cards have performed in benchmarks. LL performance should be slightly less than a Titan RTX, but vastly slower than a Titan V / Tesla V100 because those Volta chips have lots of DP performance. For some reason the higher DP GFLOPS rating doesn't seem to translate directly into LL performance. And in factoring the Titan RTX and Titan V are pretty much at the same level, and that is about 3.5x faster than an ordinary Vega 64. From purely a performance perspective, I wouldn't think it's worth any more than a Vega 64. But there are other reasons why these workstation cards exist.

2019-01-07, 07:39	#4
preda "Mihai Preda" Apr 2015 2²·3·11² Posts	I'd look here https://en.wikipedia.org/wiki/List_o...ocessing_units for a comparison to Vega64. At first look, seems pretty similar to consumer Vega64 in power and memory bandwidth. (1/16 DP). May have ECC RAM (I don't know), but that's not such an important factor anymore. Probably a better deal would be a Vega64 at half price.

2019-01-07, 08:13	#5
nomead "Sam Laur" Dec 2018 Turku, Finland 13D₁₆ Posts	Yes it has ECC support. That's usually the main distinguishing characteristic between consumer and workstation products. Also different display connectors, but that's again irrelevant when using it for computation only. https://www.amd.com/en/products/prof...on-pro-wx-8200

2019-01-09, 09:05	#8
LaurV Romulan Interpreter "name field" Jun 2011 Thailand 41×251 Posts	Yeah, right. What took my eyes and inflated my imagination was the "2048 bit memory interface" (compared with 352 known from cuda cards), I went like "Waaaaa !" (see my avatar on the left). But it seems this is common too, for AMD cards, they use a different approach.

2019-01-09, 13:35	#9
M344587487 "Composite as Heck" Oct 2017 2·5²·19 Posts	HBM is the difference and it's the direction all compute cards will likely go eventually as it can offer more bandwidth, some nvidia cards like the V100 use HBM2. AFAIK Vega was the first to use HBM2 in any meaningful quantity on consumer cards but it's still relatively expensive, we're unlikely to see it in cheaper consumer cards anytime soon. I'm hoping a consumer Vega refresh is coming as it might be good for us, HBM2 bins have improved a lot since Vega was released and a 7nm Vega should be a lot more energy efficient. An MI50/MI60 would be nice but I don't have enough kidneys to sell.

2019-01-09, 15:23	#10
firejuggler "Vincent" Apr 2010 Over the rainbow 2920₁₀ Posts	I might be late to the party but https://hothardware.com/reviews/amd-...pro-gpu-review

2019-01-10, 13:28	#11
LaurV Romulan Interpreter "name field" Jun 2011 Thailand 41·251 Posts	Thanks. Nice. Articles similar to the one you posted is what sparked this thread in the first place, stimulated by the fact that we bought some other hardware and had the opportunity to get this for a cheaper price. But after more web studies and replies in this thread, we also have seen that this card is "usually" under $1k, so our trade would not be as discounted as we assumed - as said, our reference point for comparison (expensive cuda professional cards like Tasla or Quadro) was wrong from the start. We decided not to take it.