AMD Vega (FE and RX) Benchmarks [Updated Aug 10 - RX Vega 64 Unboxing]

[leoneazzurro]

But that game is one where Fiji shows one of the best performances, generally beating the 1070

http://www.guru3d.com/articles_pages/asus_geforce_gtx_1080_strix_oc_11_gbps_review,15.html

Considering the increases shown in the slide, Vega in that game should stay in the middle between 1080 and 1080Ti.
But it is very likely that in other cases we can see larger increases in Vega, in some rare cases even less. Depends on where the bottleneck is, and sometimes the bottleneck is not even inside the GPU.

 

[antihelten]

AMD's own numbers have it @20~25% at low resolutions, and 34% at 4K.

I missed that one. It is of course only one game, but I would expect AMD to cherry pick their games to show Vega in as good a light as possible, so I wouldn't be surprised if this is one of the better cases for Vega.

So 25-35% is probably realistic.

But that game is one where Fiji shows one of the best performances, generally beating the 1070

http://www.guru3d.com/articles_pages/asus_geforce_gtx_1080_strix_oc_11_gbps_review,15.html

Considering the increases shown in the slide, Vega in that game should stay in the middle between 1080 and 1080Ti.
But it is very likely that in other cases we can see larger increases in Vega, in some rare cases even less. Depends on where the bottleneck is, and sometimes the bottleneck is not even inside the GPU.

So what if this is a game where Fiji does well compared to the 1070. Raghu's claims were with regards to Fiji vs Vega.

Also I don't know if you can really say that Fury X generally beats the 1070 in BF1, at least not if you go by TPUs numbers, where it basically trades blows at 1080P and 1440P (0-5% difference) and loses at 4K (13% slower). Also the improvement shown by AMD in their slide, would put Vega neck and neck with the 1080, not between the 1080 and 1080 Ti.

Also why would you expect to see larger increases for Vega in other cases. I would expect the marketing department of AMD to show off the best case scenario wherever possible, unless they are engaged in an epic case of sandbagging. I mean when have you ever seen a product turn out better in real life than in the marketing material?

 

[tviceman]

Also why would you expect to see larger increases for Vega in other cases. I would expect the marketing department of AMD to show of the best case scenario wherever possible, unless they are engaged in an epic case of sandbagging. I mean when have you ever seen a product turn out better than in real life than in the marketing material?

Never before have truer words been spoken.

 

[Muhammed]

Vega 64 liquid time spy score : 7580
a GTX 1080 non oc'ed is about: 7450
http://www.3dmark.com/compare/spy/2192756

 

[leoneazzurro]

I missed that one. It is of course only one game, but I would expect AMD to cherry pick their games to show Vega in as good a light as possible, so I wouldn't be surprised if this is one of the better cases for Vega.

So 25-35% is probably realistic.

So what if this is a game where Fiji does well compared to the 1070. Raghu's claims were with regards to Fiji vs Vega.

Also why would you expect to see larger increases for Vega in other cases.

Because it is ONE benchmark where Fiji already does quite well (so its bottlenecks are not affecting it so much, where other architectures's bottlenecks can affect them, i.e. if there is a memory bandwidth bottleneck in an application you will have lower gains going from Fiji to Vega, if at all, where if the shader or polygon throughput is the bottleneck you'll see probably larger gains, due mainly to clock speed increse I'm afraid) , and there are other benchmarks if you look in the endnotes of the presentation deck where you have gains (on minimum framerates) where gains are already larger than that (and other show also less gain, but as told before it must be seen, where the bottleneck is), and because in other non-gaming tasks there are increases (i.e. Polygon, compute) much larger than that. Not saying that it will become a wonderful products, but at least one must not see only the black side of things.

 

[Qwertilot]

To be fair they've even an alibi for their marketing department here - even if the increase from Fiji was larger in other games they might well not have wanted to show any kind of benchmark for those games if the overall performance was still rubbish.

 

[Despoiler]

No, I'm not. One product can win 3 benchmarks and lose 7. That is still trading blows, even if it's 5-10% slower overall.

Ahh yes you are. You literally just did it again. Anytime you post a speculative it's AMD is losing more than it wins. In this last chuckle post your scenario is a clear loss scenario that you are trying to present as trading blows. That's what we call bias.

 

[Rasterizer]

Vega 64 liquid time spy score : 7580
a GTX 1080 non oc'ed is about: 7450
http://www.3dmark.com/compare/spy/2192756

I am forced to seriously question how recent the drivers in those leaked TimeSpy scores are, only because they have the air cooled Vega 64 at ~7,200 which basically identical to the 7,126 score achievable by an air cooled Vega FE. That doesn't seem to make much sense, as it would mean that the Vega uarch features not enabled in FE (AVFS and DSBR, in particular) would be worth a literal ~0% performance improvement. It would also strongly suggest that RTG has been unable to fix the memory bandwidth issues in Vega FE, but if that is the case then the rumours of much better ETH hashrate performance should not also be able to be correct.

 

[Tup3x]

I am forced to seriously question how recent the drivers in those leaked TimeSpy scores are, only because they have the air cooled Vega 64 at ~7,200 which basically identical to the 7,126 score achievable by an air cooled Vega FE. That doesn't seem to make much sense, as it would mean that the Vega uarch features not enabled in FE (AVFS and DSBR, in particular) would be worth a literal ~0% performance improvement. It would also strongly suggest that RTG has been unable to fix the memory bandwidth issues in Vega FE, but if that is the case then the rumours of much better ETH hashrate performance should not also be able to be correct.

Apparently the drivers are new (I read from other forums) but anyone familiar with AMD driver version numbers should check it out.

 

[antihelten]

Because it is ONE benchmark where Fiji already does quite well (so its bottlenecks are not affecting it so much, where other architectures's bottlenecks can affect them, i.e. if there is a memory bandwidth bottleneck in an application you will have lower gains going from Fiji to Vega, if at all, where if the shader or polygon throughput is the bottleneck you'll see probably larger gains, due mainly to clock speed increse I'm afraid)

Yes it's only one benchmark, but it's one chosen by AMD themselves. Unless their marketing department is incompetent or sandbagging, they wouldn't pick one where Vega potentially suffers from bottlenecks compared to Fiji, since their goal is to show off Vega in the best possible light.

and there are other benchmarks if you look in the endnotes of the presentation deck where you have gains (on minimum framerates) where gains are already larger than that (and other show also less gain, but as told before it must be seen, where the bottleneck is),

Which other benchmarks are you talking about? I didn't see any in which Vega has a significantly bigger gain of Fiji than the 35% shown in BF1 4K. The only other high one I see is the min FPS for CoD:IW where Vega is 36% faster.

Other than that all of the others are lower, once again indicating that BF1 (and CoD:IW) are best case scenarios.

and because in other non-gaming tasks there are increases (i.e. Polygon, compute) much larger than that.

Wake me up when this actually translates to better game performance.

For the average consumer, these things will probably only result in one thing, and that's higher prices, since miners will be buying up Vega cards, due to the high compute capabilities.

Not saying that it will become a wonderful products, but at least one must not see only the black side of things.

I'm not seeing the black side of things, I'm literally quoting AMD marketing slides.

You on the other hand, appear to believe that Vega will actually turn out better than what AMD is indicating in their marketing material. Which seems optimistic to put it mildly.

 

[Muhammed]

I am forced to seriously question how recent the drivers in those leaked TimeSpy scores are

This last leak has the newest driver.

Latest RX Vega Leak: 22.19.666.1
Previous RX Vega leak : 22.19.640.2
Older RX Vega leak: 22.19.653.0
Vega FE driver : [URL='http://www.3dmark.com/fs/12987144']22.19.384.2[/URL]

 

[Rasterizer]

This last leak has the newest driver.

Latest RX Vega Leak: 22.19.666.1
Previous RX Vega leak : 22.19.640.2
Older RX Vega leak: 22.19.653.0
Vega FE driver : 22.19.384.2

If RX Vega turns out to not be any faster at all than Vega FE, then pretty much the entirety of RTG needs to get fired. The scale of failure implied would be so enormous that I'm going to hold off judgement until I see actual independent benchmarks. This would also mean that the multiple reports of an improvement in hashrate are also totally false and that RX Vega will end up no better at mining than FE is.

 

[Malogeek]

That doesn't seem to make much sense, as it would mean that the Vega uarch features not enabled in FE (AVFS and DSBR, in particular) would be worth a literal ~0% performance improvement. It would also strongly suggest that RTG has been unable to fix the memory bandwidth issues in Vega FE, but if that is the case then the rumours of much better ETH hashrate performance should not also be able to be correct.

That's assuming there are any gains to be had with DSBR or that bandwidth is an issue in this particular benchmark. I'm not saying it isn't, just that there are many factors that can cause bottlenecks that these particular parts you mention simply don't alleviate. With the amount of geometry present, it could simply be setup limited with the 4 geometry engines.

 

[Ryan Koller]

Also recall that TimeSpy uses the minimum DX12 featureset and not the maximum. When it was released, NVidia didn't have anything but the most barebones DX12 implementation and the TimeSpy devs decided to include only the ubiquitous features in the benchmark. It literally doesn't use any of the new features in Vega that aren't in Fiji. It is not representative of what Vega can do with optimization. Believing that those optimizations will or won't occur in more recent games is something of an unknown, but we do know the new Wolfenstein supports rapid packed math. 3DMark needs to come out with a newer benchmark that supports ALL of DX12's features, including rapid packed math as that is a defining feature of Volta. At least with only two players in the graphics arena we know specialized features, such as DSBR, will make it into at least a few games and engines. I'm hoping Tim Sweeney showing up on stage at Siggraph has some far reaching implications for UE and that it wasn't a minor marketing stunt.

It's a chicken and the egg problem, but using old software to benchmark new hardware has always been a problem. Unforunately, AMD doesn't have the manpower to rewrite portions of the rendering pipelines in their drivers to the same degree that NVidia does. NVidia had most of the benefits of DX12 tucked away into DX11, AMD will simply never compete in DX11 games. All in all I'm going to get a Vega card simply because I can get a really nice IPS 1440p 144hz Freesync monitor (https://www.newegg.com/Product/Prod...m_re=nixeus_edg_27-_-9SIA0ZW5F71608-_-Product) for nearly $300 less than an equivalent Gsync monitor (https://www.newegg.com/Product/Product.aspx?Item=N82E16824106004). BUT if Nvidia would support open standards and VESA adaptive sync, I would probably have grabbed a 1070 or a 1080 a while ago.

 

[Rasterizer]

That's assuming there are any gains to be had with DSBR or that bandwidth is an issue in this particular benchmark. I'm not saying it isn't, just that there are many factors that can cause bottlenecks that these particular parts you mention simply don't alleviate. With the amount of geometry present, it could simply be setup limited with the 4 geometry engines.

From my--admittedly limited--understanding, the fact that Vega FE trades blows with a Quadro P6000 in Creo, Solidworks, and 3ds Max (according to Tom's Hardware) would seem to indicate that Vega shouldn't be geometry performance bound. Gamer's Nexus clock-for-clock Vega FE vs Fury testing also seems to show that geometry performance is the one area in which Vega FE had a clear, measureable clock-for-clock advantage in the most geometry heavy 3DMark tests

GN's clock-for-clock testing seems to show Vega FE performing about even with a Fury X at the same clocks, but when you consider that Fury did zero primitive culling and had no TBR of any kind, them performing the same clock-for-clock in games implies that either there was a significant performance regression in some other area, or primitive discard and DSBR are worth a grand total of zero and every hour put into their development by RTG was thrown away.

As it turns out, we already know based on B3D Suite testing of Vega FE that it does show 20-30% regressions vs. Fiji in raw memory bandwidth, effective texture bandwidth and texture fillrate. This, in turn, leads me to infer that Vega is largely memory bound in gaming and game mimicking synthetic testing (as well as ETH mining hashrate). Of course, I'm still a total novice at understanding GPUs, so I could be totally misreading things.

 

[leoneazzurro]

Yes it's only one benchmark, but it's one chosen by AMD themselves. Unless their marketing department is incompetent or sandbagging, they wouldn't pick one where Vega potentially suffers from bottlenecks compared to Fiji, since their goal is to show off Vega in the best possible light.

Maybe they chose it because it is a benchmark because it is one where already the FE is doing better than the 1080.

Which other benchmarks are you talking about? I didn't see any in which Vega has a significantly bigger gain of Fiji than the 35% shown in BF1 4K. The only other high one I see is the min FPS for CoD:IW where Vega is 36% faster.

Other than that all of the others are lower, once again indicating that BF1 (and CoD:IW) are best case scenarios.

My bad, I miscalculated. But here is a 46% more in RottR for the Vega FE from an indipendent review, with early drivers (no tile based rasterization until August 14th, it seems).
https://www.pcper.com/reviews/Graph...ition-16GB-Air-Cooled-Review/Rise-Tomb-Raider
So it depends on where the bottleneck is.

Wake me up when this actually translates to better game performance.

For the average consumer, these things will probably only result in one thing, and that's higher prices, since miners will be buying up Vega cards, due to the high compute capabilities.

How so? Nvidia fans always complain about the green cards being so vastly superior because their geometry throughput...
Anyway, depends on the game and features turned on, but in any case having less bottleneck on geometry is preferrable to having one.

I'm not seeing the black side of things, I'm literally quoting AMD marketing slides.
You on the other hand, appear to believe that Vega will actually turn out better than what AMD is indicating in their marketing material. Which seems optimistic to put it mildly.

I am not saying anything before reviews come out. I already said that power consumption is a negative point. Many people, instead, are very fast to judge Vega as a failure before that, an this more due to their "brand loyalty". than facts. Facts that must take in account the novelty of the architecture and that the actual Vega chip will address much more than the gaming market.
I own an Nvidia card, but being a tech enthusiast I will be objective until I see the actual results.

 

[Rifter]

Then it doesn't matter if AMD prices them right does it?

So what was your point?

AMD still needs to play the game is my point, Due to mining most cards go for $50-150 over MSRP right now, so if you really want your card sold at $400 street price you should be setting MSRP at $300 or so. Then if mining ever does crash like they have been saying it will for the last year but it never happens then raise the prices to keep it selling at $400 street price

 

[crisium]

That's not how it works. The mining market determines GPU prices based on a combination of hashrate efficiency and raw hashrate. MSRP if anything sets a floor: it's why GTX 1080 is not cheaper than GTX 1070.

GTX 1070 prices would not change if the MSRP were $100 suddenly, except by any retailer sorry enough to blindly follow MSRP.

$400 MSRP for Vega 56 is based on a scenario where MSRP was actually followed. They want to price it +$50 over GTX 1070. If mining no longer impacted prices like in 2016 this is what AMD wants.

 

[EXCellR8]

AMD along w lots of retailers stating that stock should be good despite the inevitable surge of miners wanting the new cards. Damn invalids are gonna be swarming over them though... I cringe at all that wasted stock; I hope the gaming crowd at least gets a nice piece of Vega at launch.

https://videocardz.com/newz/biostar-teasing-motherboard-with-104-usb-risers-support-for-mining

NO.

 

[Glo.]

http://developer.amd.com/wordpress/media/2013/12/Vega_Shader_ISA_28July2017.pdf

In general ISA of Vega is focused on Graphics improvements, and FP16.

Also what is apparent is that it is not exactly the same architecture as previous versions of GCN. Just Instruction Set of this arch is more robust, and should allow for higher capabilities of the hardware.

 

[tviceman]

Ahh yes you are. You literally just did it again. Anytime you post a speculative it's AMD is losing more than it wins. In this last chuckle post your scenario is a clear loss scenario that you are trying to present as trading blows. That's what we call bias.

I see that you only want to argue for the sake of arguing. Jusy go grab yourself a couple of Vega 64s and enjoy the ride.

 

[Elixer]

That's not how it works. The mining market determines GPU prices based on a combination of hashrate efficiency and raw hashrate. MSRP if anything sets a floor: it's why GTX 1080 is not cheaper than GTX 1070.

No, supply & demand determines prices.
If AMD would use more fabs to makes Vega & Polaris, then there wouldn't be an issue at meeting the MSRP (Assuming ample supple of HBM2).
They could be doing that, but, we have no idea.

I said it before, and I'll say it again, GloFlo just isn't big enough to produce everything for AMD that AMD needs right now. They are churning out Ryzen, ThreadRipper, Epyc, Polaris & Vega.
The other issue is low supply of HBM, so, even if AMD did move production to a new fab, without ample supply of HBM2, it wouldn't do them any good.

 

[raghu78]

Ahh yes you are. You literally just did it again. Anytime you post a speculative it's AMD is losing more than it wins. In this last chuckle post your scenario is a clear loss scenario that you are trying to present as trading blows. That's what we call bias.

Don't waste your time arguing with these people. You are spot on.

 

[Tweak155]

I see that you only want to argue for the sake of arguing. Jusy go grab yourself a couple of Vega 64s and enjoy the ride.

Actually it's a valid gripe. Based on your assumption that "trading blows" == "losing majority", even if AMD showed them winning in 9/10 test cases and showed the games, seems likely you'd just nitpick the games they decided to show and now all of a sudden they should be downgraded to "trading blows".

In other words, since there are not more details / evidence, bias is the likely case for your reduction of their statement of "trading blows" being equal to "losing majority" to the consumer.

You need to phrase your post more clearly as opinion rather than trying to pass it as fact. Once more details are released, bias will not be a consideration as you can then link the evidence.

That's just my take, anyway.

 

[tviceman]

Actually it's a valid gripe. Based on your assumption that "trading blows" == "losing majority", even if AMD showed them winning in 9/10 test cases and showed the games, seems likely you'd just nitpick the games they decided to show and now all of a sudden they should be downgraded to "trading blows".

In other words, since there are not more details / evidence, bias is the likely case for your reduction of their statement of "trading blows" being equal to "losing majority" to the consumer.

You need to phrase your post more clearly as opinion rather than trying to pass it as fact. Once more details are released, bias will not be a consideration as you can then link the evidence.

That's just my take, anyway.

Read PR speak. This happens with every new hardware release. We would see way more evidence from AMD if "trading blows" was the norm. If Vega 64 won 90% of the time, AMD wouldn't say trade blows. When Nvidia released the 1070, they said it was faster than Titan X maxwell, not "usually faster" or "trade blows" as it sometimes did. It' as simple as that. Reviews will prove me right or wrong.