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

[ThatBuzzkiller]

And you know the thing that frustrates me the most? We're probably never going to know what happened to poor Vega. Oh, if this is as bad as it's looking, we'll have a full Bulldozer style postmortem going on with people decompiling the drivers bit by bit and doing god knows what to the hardware to try and figure out how on earth they regressed from Fiji. I look forward to the endless theorycrafting and the same questions being asked as Bulldozer faced: Was this doomed before it even left the drawing board? Good design screwed over by a bad process? Hardware good but the drivers proved too much to handle? Did they get 99% of the way there and something unexpected came up and completely derailed them and they could never recover? Is Raja a cackling demon drinking our salty tears or is he currently cowering under a table smoking his last cigar before Lisa Su comes along to drown him in a barrel of FineWine(tm)?

The issue with AMD is that they focus too much on the future to go in their way. The main lesson for AMD should be is 'build for the present and not upon the future'. The problem with predicting the future is that you don't know which scenario will play out so you're product could either end up as a dead end or a disruptive innovation in the marketplace ...

The biggest difference between Bulldozer and Zen is that the former tried so hard to be future proof whereas the latter just did enough in it's limits and we can see which one truly made the 'Epyc' comeback here (pun intended) ... (Bulldozer's module design didn't make any inroads with the future but Zen on the other hand struck gold and fears into it's competitor with better perf/watt, similar or better perf/socket in HEDT/Servers at the highest end, and lastly better perf/$ while Intel was focusing so hard on getting AVX-512 in their Skylake-SP/X platforms whereas Zen only had half rate AVX/AVX2 and they got caught off guard by the infamous 'glue' which was arguably built to take advantage of the present! )

With Vega, AMD is banking on shader model 6 or SPIR-V to shine for them but the problem is there are absolutely no games (aside from maybe the recent Doom) using the new shading languages! (Even then we don't know if SM6 or SPIR-V will give AMD the definitive advantage in the future!)

How can you hope to possible dominate the future if you can't make an impact in the present ? If Raja had never considered that thought for even a moment when working with AMD for years then I think he needs to step down and AMD needs to do everything in their power to get Eric Demers back so he can put the graphics division back in order just like how AMD hired Jim Keller to put their CPU division back on the right track ...

AMD really needs to give a lot of free ISV support if their really going to build their future around DX12 or Vulkan otherwise their going to experience some hardships and tons of roadblocks ...

 

[Zstream]

The issue with AMD is that they focus too much on the future to go in their way. The main lesson for AMD should be is 'build for the present and not upon the future'. The problem with predicting the future is that you don't know which scenario will play out so you're product could either end up as a dead end or a disruptive innovation in the marketplace ...

The biggest difference between Bulldozer and Zen is that the former tried so hard to be future proof whereas the latter just did enough in it's limits and we can see which one truly made the 'Epyc' comeback here (pun intended) ... (Bulldozer's module design didn't make any inroads with the future but Zen on the other hand struck gold and fears into it's competitor with better perf/watt, similar or better perf/socket in HEDT/Servers at the highest end, and lastly better perf/$ while Intel was focusing so hard on getting AVX-512 in their Skylake-SP/X platforms whereas Zen only had half rate AVX/AVX2 and they got caught off guard by the infamous 'glue' which was arguably built to take advantage of the present! )

With Vega, AMD is banking on shader model 6 or SPIR-V to shine for them but the problem is there are absolutely no games (aside from maybe the recent Doom) using the new shading languages! (Even then we don't know if SM6 or SPIR-V will give AMD the definitive advantage in the future!)

How can you hope to possible dominate the future if you can't make an impact in the present ? If Raja had never considered that thought for even a moment when working with AMD for years then I think he needs to step down and AMD needs to do everything in their power to get Eric Demers back so he can put the graphics division back in order just like how AMD hired Jim Keller to put their CPU division back on the right track ...

AMD really needs to give a lot of free ISV support if their really going to build their future around DX12 or Vulkan otherwise their going to experience some hardships and tons of roadblocks ...

I agree with many other posters, but the biggest issue, and one often overlooked, is that AMD controls all of the gaming sectors on the console front. I mean, why are people even using Nvidia cards to work on console games? I will never understand how you can control the gaming market and still suck at optimizations and efficiency on dedicated GPU cards.

This is definitely a problem with their partner program. It clearly sucks and needs to be destroyed and built up properly.

 

[Paratus]

That is the whole conundrum here.

That is one big chip, on a smaller node, yet, it is barely beating Fury on some workloads.
As a prosumer card, it also is losing on workloads it shouldn't.

Worst case is, that the first iteration of Vega's plumbing is fubared, and they had to fuse it off, and write workarounds in the driver which makes it really, really slow for certain workloads. This makes the most sense right now considering what we have seen.

If that is the case, and since it takes so long to do a re-spin, the next iteration would be back from the fabs by the end of August. So, if all of the sudden they release a Vega Ti a few months from now, we know why.

That is all conjecture though.

Best case is, the drivers will bring a 30% performance boost to some things.

10 days to go, and I think this whole thread has pretty much covered every single possible explanation there could be for why Vega turned out the way it turned out.
Can't think of anything that anyone missed.

Except it's not as big as we were thinking, when the rumors were 500+ mm^2.

• 1080 TI - 471mm^2 & 12B xtors
• Vega FE - 484 mm^2 & ?? xtors

I'd assume Vega would have anywhere between 11.9B and 13.7B xtors based on Polaris at the low end and a 16% xtor density increase similar to what AMD pulled off on 28nm going from gen 1 to 2.

So Vega is basically within a few percent size wise with GP102. Despite the theoretical improvements saying otherwise, AMD doesn't usually beat NV for outright performance with similar sized or smaller chips.

Game performance doesn't make sense. So I'm left with bounding Vega RX performance at worst case as Vega FE air which is. ~ 90% the speed of a 1080 @ ++ power.

Best case for water cooled or high end air RX Vega:

• ~1630 boost clock and it hits it regularly
• Smaller HBM stack for 8GB - bumps speed back to 512GB/s and allows small increase in core performance due to less heat and power from small stack
• Driver optimizations fix certain currently unknown bottlenecks (ROPs? Mem?)

Beats factory OC'd 1080's by ~10% @ ++power.

Damn I want a post mortem on this.

 

[Armsdealer]

These analyses overlook the 800 lb gorilla in the room: AMD WAS ALMOST BANKRUPT AND HAD NO ACCESS TO CAPITAL WHILE VEGA WAS BEING DEVELOPED.

Also, as raja mentioned at a recent financial analyst day, vega was developed in an environment in which nobody thought gpu was the place amd to be investing.

No money + perception that gpu has no future = mediocre gpu, focus on cpu

Now amd has access to plenty of money. They raised a billion dollars from the equity and debt markets when wall street got sexed up on Zen, and they can still raise more if they need to given their share price remains elevated.

Perceptions in the stock market have *completely* changed for the positive on the gpu outlook. It's probably been the biggest shift in market perceptions of growth for any industry in the global economy. It's very logical for them to invest heavily there for the next generation. Want to make a smart bet? AMD is going to do better with gpu in the future and will try to find a way to maintain mindshare and relevance in the space this generation.

 

[leoneazzurro]

Personally, I think the major concern is the power draw. Die size is a concern for AMD, not the consumer, if the card price is right. And it is true that feature wise, Vega is ahead of current Nvidia offerings (Primitive shader, support for tier 3 DX12 resources, and so on). So it is bound to be bigger anyway: if someone thinks that features come for free, he is highly mistaken. Performance is not stellar, but with better cooling and some driver optimization it is completely possiible we could see RX Vega in the average a bit above the 1080 (standard) but I think that there are practically no chance to see it in a striking distance of 1080Ti. But this could be fine, depending on price, and the fact that quite probably we will not see a comparable (price bracket wise) Volta part before March 2018 (and anyway I don't think we will see very high performance uplift in gaming with it, Volta seems a quite compute oriented part).
That power draw, anyway, is something to be addressed. AMD is out of the mobile market, period. And mobile market is an HUGE slice of the whole tart. I have a GTX1070 laptop (no place for a desktop at home) and quite happy with it. It would have been nice if AMD put some competition there, but simply there was not.
Also, it does not bode well for future iterations with bigger resources pool. And I think it's quite clear that is not simply a matter of production process, because Ryzen already demonstrated GF can produce power optimized parts even if it hits a barrier in frequency. This is mainly an issue with the underlying architecture. Granted, we will se an handful of watts being saved by the 8GBytes less of HBM2, but if there was no respin on the die, or if we have not seen yet any working power optimization, I'd say that performance/power efficiency so far is quite disappointing.

 

[stahlhart]

Thread cleaned.
-- stahlhart

 

[PhonakV30]

Vega FE Is based on Revision 1?

 

[iwulff]

Vega FE Is based on Revision 1?

What do you mean?

 

[Cloudfire777]

What do you mean?

Some wet fantasy that Vega RX will have magical properties (New revision or stepping)

 

[Cloudfire777]

So we are 1 week from launch and no tests/leaks out yet?
Shouldnt retailers etc have these cards by now?

 

[amenx]

Last time I recall them being so tight lipped about about a products performance before launch was Phenom.

 

[Head1985]

So we are 1 week from launch and no tests/leaks out yet?
Shouldnt retailers etc have these cards by now?

Its 100% paper launch because shops dont have cards yet.It will be similar to GTX1080-reveal at 30.7 and in shops month later.
Also non reference are still not in mass production.So they will be avalaible in november/december...

 

[Cloudfire777]

Its 100% paper launch because shops dont have cards yet.It will be similar to GTX1080-reveal at 30.7 and in shops month later.
Also non reference are still not in mass production.So they will be avalaible in november/december...

So people get these products first in like late August or September? Not only have they delayed launch til July which pissed people off, but they didnt even supply retailers so they are ready when the launch finally happens?

Can AMD get any worse? When I saw the AIB news that they would be late, I atleast expected AMD to have reference ready.

Usually we have leaks weeks before retailers starting selling them. Always someone that get a hold of one or two through retailers not giving a crap about dates

 

[Veradun]

So people get these products first in like late August or September? Not only have they delayed launch til July which pissed people off, but they didnt even supply retailers so they are ready when the launch finally happens?

Can AMD get any worse? When I saw the AIB news that they would be late, I atleast expected AMD to have reference ready.

Usually we have leaks weeks before retailers starting selling them. Always someone that get a hold of one or two through retailers not giving a crap about dates

This is happening a lot lately, being Intel. AMD or NVidia. They all do a preview they call launch and then it's reviewed at least one month later. No surprises here.

 

[Cloudfire777]

This is happening a lot lately, being Intel. AMD or NVidia. They all do a preview they call launch and then it's reviewed at least one month later. No surprises here.

The difference here though is that Nvidia have one presentation and then you get to buy them after a while.

With Vega it have been 100 presentations, one official date for launch and you wont even get to buy it after?

 

[Qwertilot]

Non ref as late as Nov/December would surely make those totally pointless?

 

[Veradun]

The difference here though is that Nvidia have one presentation and then you get to buy them after a while.

With Vega it have been 100 presentations, one official date for launch and you wont even get to buy it after?

And this changes what?

Do you think Volta will be on shelves without any other presentations?

How cute :>

 

[Guru]

AMD should have just released Vega in Q1 2017, yes the 1080ti took their thunder, but who cares, they would have still be competitive with the 1080 and would have taken at least some market share.

In fact if they actually released the card before March, they would have had a good one month before 1080ti and of course good 2 months before full 1080ti availability.

The fact that this **** performs so bad, yet its more transistors than the 1080ti, its bigger die size than the 1080ti and its over a year late compared to the 1080, over 3 months late over the 1080ti and still performs so bad is game over for AMD in the high end segment.

Didn't Raja join like 3 years ago? He would have been full on on this project, not on the polaris project, but on the Vega project. I think he had full control, even if he wasn't there from the start, in 3 years he should have been able to 100% move the project in his direction fully.

So he should be fired ASAP! AMD should just do things more simply, stop with the futuristic designs that don't have effect 5 years into the future and build graphics for now.

Nvidia is going to come out with Volta in 2018 and it will be DX12 and Vulkan optimized and 100% beat AMD in that sphere as well. Now they came out with DX11 optimized graphics and beat AMD in 80% of the titles, because most are DX11 titles.

Sure there are some DX11 titles that are optimized for AMD, but that is it, Nvidia wins in 80% of the DX11 titles, or at least it did. Now RX 580 is overall a better purchase, but it took a whole year and a new revision, higher stock clocks and rebranding to do it.

Harsh language and swearing is not allowed in the tech section. -Shmee

 

[lobz]

AMD should have just released Vega in Q1 2017, yes the 1080ti took their thunder, but who cares, they would have still be competitive with the 1080 and would have taken at least some market share.

In fact if they actually released the card before March, they would have had a good one month before 1080ti and of course good 2 months before full 1080ti availability.

The fact that this **** performs so bad, yet its more transistors than the 1080ti, its bigger die size than the 1080ti and its over a year late compared to the 1080, over 3 months late over the 1080ti and still performs so bad is game over for AMD in the high end segment.

Didn't Raja join like 3 years ago? He would have been full on on this project, not on the polaris project, but on the Vega project. I think he had full control, even if he wasn't there from the start, in 3 years he should have been able to 100% move the project in his direction fully.

So he should be fired ASAP! AMD should just do things more simply, stop with the futuristic designs that don't have effect 5 years into the future and build graphics for now.

Nvidia is going to come out with Volta in 2018 and it will be DX12 and Vulkan optimized and 100% beat AMD in that sphere as well. Now they came out with DX11 optimized graphics and beat AMD in 80% of the titles, because most are DX11 titles.

Sure there are some DX11 titles that are optimized for AMD, but that is it, Nvidia wins in 80% of the DX11 titles, or at least it did. Now RX 580 is overall a better purchase, but it took a whole year and a new revision, higher stock clocks and rebranding to do it.

Harsh language and swearing is not allowed in the tech section. -Shmee

your comment - and it is not the first here - sounds like your wife cheated on you with AMD and then AMD killed your whole family

Keep your posts on topic and respectful towards other forum participants, or you're going to get kicked out of here.
-- stahlhart

 

[Elixer]

Didn't Raja join like 3 years ago? He would have been full on on this project, not on the polaris project, but on the Vega project. I think he had full control, even if he wasn't there from the start, in 3 years he should have been able to 100% move the project in his direction fully.

You think he had full control, but, the design team was working on both (Polaris & Vega) at the same time IIRC. They are really close in terms of tech, with only a few additions in Vega to speed some stuff up.
Navi is the only tech that is fully under Raja's belt, so, not sure why you think he should be canned right now.
If Navi is a disaster, then sure, that is what is most likely going to happen to him.

 

[Magic Hate Ball]

You think he had full control, but, the design team was working on both (Polaris & Vega) at the same time IIRC. They are really close in terms of tech, with only a few additions in Vega to speed some stuff up.
Navi is the only tech that is fully under Raja's belt, so, not sure why you think he should be canned right now.
If Navi is a disaster, then sure, that is what is most likely going to happen to him.

It's not just that he helmed the projects, but also consider that AMD considered dGPUs a failing market several years ago and I bet R&D budget went more towards Zen and recovering in CPU market share.

Also consider that Vega probably had Infinity Fabric foisted upon them to work with Zen in the RR APU.

I don't think Raja is going anywhere as long as Raven Ridge works as needed for the low end desktop and laptop markets which are huge volume.

 

[GodisanAtheist]

If you assume rational players here, GCN/Fiji and Vega make a lot of sense from the perspective of a CPU company looking for a solid, specialized co-processor that works well in computationally intensive workloads where GPUs are strong and CPUs are relatively weak + with low memory bandwidth limitations.

HBM + the GPU makes perfect sense when a company wants to sell their distributors a SOC (CPU/GPU/RAM) on one package and keep all the margins to themselves instead of spreading it around with different vendors.

AMD is slowly merging their various products into a SOC they can sell to high margin markets and they're trying to do it with Sony and Microsoft's money.

The fact that we're getting a stand alone GPU for gaming out of the whole thing was probably an afterthought of an afterthought. No one is losing their job if the APU side of this endeavor goes off with minimal hiccups.

 

[Rifter]

so when is the official launch date for this? this week or next week?

 

[Elixer]

so when is the official launch date for this? this week or next week?

Vega FE already launched.
Vega RX will "launch" on July 31st, but, 99.999% chance of it being paper launch that day.

 

[Guru]

Don't get it how an architecture a bit bigger than the 1080ti, more transistors, 1 year more development time and it can barely even compete with the GTX 1080 and consumes more power!