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

[tential]

I'm pretty well exhausted from following all the 'polite discussion' around this launch, and while I'm not expecting miracles from RX Vega, it's worth bearing in mind how dangerous it can be to jump to conclusions. Exhibit A:

Vega Crashes during blender benchmark Hundreds of people piling in with their incredibly witty and original observations about 'poor volta, wait for navi, another fail for AMD' etc etc

After this particular 'truth' has gotten the whole way around the world, this appears:

Crashes are caused by a bug in the benchmark, not the gpu/drivers, fix available

RE: The die size: Seems to be some confusion over this, last I heard it is indeed <500mm^2, and the measurements suggesting otherwise are in error (accidentally measured extra area outside the actual die, even a tiny error in the linear dimension measure makes a big difference to the area. I'll see if I can find the post)

Good post but this goes to what many of us were saying. Putting this into the wild without any guidelines or real comments wasn't good. Amd should have given the card to reviewers and told them how to review it in their reviewer guide. They only brought all of this speculation on themselves

 

[Cloudfire777]

If that were the case, AMD would have released RX Vega at the same time. The gaming drivers aren't ready. I am not just taking their word for it, I know this to be fact.

Did you know that both AMD and Nvidia have optimized code paths for the top games? They may even use shader specific optimizations. None of that stuff is used in pro applications, but it IS used in gaming. All the stuff in the pro drivers is also used in the gaming drivers so which would you build first?

Did you know that Quadro P6000 is a little bit faster in games than GTX Titan X? I know its hard to accept, but Pro cards isnt really that much slower in gaming than Geforce.

Here is 3840 core Pascal P6000 vs Titan X 3584 core:

 

[nathanddrews]

Right, professional cards used to be slower in some games than their gaming variants, but that is not typically the case anymore. Clock speeds are the usually the reason why pro cards suffer in games.

 

[ZGR]

Right, professional cards used to be slower in some games than their gaming variants, but that is not typically the case anymore. Clock speeds are the usually the reason why pro cards suffer in games.

Not to mention the lack of open air variants to keep them from achieving higher clocks.

 

[Trumpstyle]

I expect Rx vega to have even worse gaming performance than vega FE (despite Rajas claim). Rx vega is most likely a cut-down version with less gpu cores and 8gb hbm. Drivers will probably increase performance by 0-2% not 20-40% some people are guessing.

Amd has had working vega cards for over 6 months now, their drivers should be good as they ever get.

 

[ZGR]

I expect Rx vega to have even worse gaming performance than vega FE (despite Rajas claim). Rx vega is most likely a cut-down version with less gpu cores and 8gb hbm. Drivers will probably increase performance by 0-2% not 20-40% some people are guessing.

Amd has had working vega cards for over 6 months now, their drivers should be good as they ever get.

I definitely don't expect a 40% increase, but 0-2%??? AMD GPU's that are cut down are barely cut. Look at the 7850, 7950, 290, Fury, and 470. These cards are barely cut down and can achieve the full performance of the next tier above just by a simple overclock.

Combined with open air coolers to achieve 1600 MHz+, and more mature drivers, I expect GTX 1080 OC performance.

 

[exquisitechar]

I expect Rx vega to have even worse gaming performance than vega FE (despite Rajas claim). Rx vega is most likely a cut-down version with less gpu cores and 8gb hbm. Drivers will probably increase performance by 0-2% not 20-40% some people are guessing.

Amd has had working vega cards for over 6 months now, their drivers should be good as they ever get.

The flagship RX Vega will almost certainly not be cut down, and the fact that it has less memory will make a negligible to no difference in games.

 

[parvadomus]

The current game performance is just... pathetic. In my eyes the last good cards from AMD where all GCN 1 + Hawaii. The rest of the cards were just a mess: barely an attempt to copy Nvidia technologies + the unknown need to go HBM asap.
First stupid card was Tonga, it basically incorporated delta color compression from Kepler, adding I dont know how many transistors over Tahiti, to barely keeping up with the original Tahiti, and getting further away from Kepler efficiency (in terms of performance/transistors).
After this Fiji comes, why the fuck we need a 4Gb HBM if we already learned DCC!.
Then it came Polaris to continue the same route, more transistors for less perf/xtors (remember Polaris 10 is almost as big as Hawaii in terms of transistors). Basically a die shrunk Hawaii would have been better, maybe even reach 1070 performance levels or better...
Finally Vega comes, trying to mimic nvidia´s tile based rasterization and adding the HBC thing, with both things adding maybe 5 or 6 billion more transistors??
If nvidia gets a new technology and AMD copies it again, it will need like 3x the transistors for the same performance, and maybe HBM 3 or 4...

 

[OatisCampbell]

If, as Raja said, creating new drivers for Vega is extremely difficult, then it's not surprising that the drivers that the FE shipped with are slower than the RX drivers will be. In professional applications, stability is king, where with gaming speed is what you focus on. Which would professionals rather have faster, more unstable drivers or slower stable ones.

My take is that what we have here is a mix of all of the discussed problems. A hardware issue that required a new stepping that pushed back the launch date. An absolute need to meet the 1H release promised to investors. Incredibly difficult to code for drivers for all of the new additions to Vega, and possibly a need to rework all of the driver work previously done for the new stepping.

I expect that the FE drivers will significantly increase in speed in a little bit, but that will take time to make sure that they are absolutely stable for the pro users. I also expect that the RX drivers will make use of all or most of the new parts of Vega at release and that they will be significantly better than what we have seen to date with the FE. It's also possible that a lot of the new power management systems aren't being used yet, so what we are seeing is pretty much a raw, stable and functional Vega, but not what it will be in the near future once it's running on all cylinders.

I don't know that Raja said it's incredibly difficult (implying that much optimization needs be done for new arch). Didn't he just say they wish it was Polaris 2.0? Which implies new arch requires new drivers from scratch, a lot of work, but par for course.
I don't think we will see radical performance difference in RX. I think we will see a 1080 competitor that uses a lot more power, and people will point that out way more than anyone who is not a miner needs to. And there will be the inevitable bulldozer comparisons that don't make any sense at all.
For AMD to come as close as they do with Ryzen and Vega when you consider the budgets and staff drain they have vs the other two companies, it's pretty amazing. What they're doing is the equivalent of RCA coming out of nowhere with competitive HDR TVs.

 

[HurleyBird]

My feeling is that the disappointing performance is mostly down to drivers and we'll probably eventually see a radical performance increase once drivers catch up. But who knows if that will be ready for RX Vega launch date. And in the chance that they are ready in terms of performance when RX Vega comes out, will they be stable in their seemingly rushed state?

 

[tential]

From reddit:

AMD also delayed very quietly the desktop Raven Ridge, that I was told by AMD representatives here in the forum in a thread was coming in H2 2017 and it is for H1 2018 now. A lot of things must have gone wrong with the driver's team. I still want to upgrade my A8 7600 for a Raven Ridge processor with whatever the full GPU cores enabled are (probably 704 stream processors). I have no intention to go to desktop cards at today's prices.

https://www.reddit.com/r/Amd/comments/6kpdzo/not_to_add_to_the_doomsday_mood_of_ramd_but_the/

Not confirmed

 

[tential]

New Vega benches!!!!!
https://youtu.be/YflRl9np70k

Game mode pro mode and game mode oc tested!
All vs the gtx 1080
Spoiler : the performance looks way better with the oc and looks promising that there is still things amd can do to get more performance

 

[Magee_MC]

I don't know that Raja said it's incredibly difficult (implying that much optimization needs be done for new arch). Didn't he just say they wish it was Polaris 2.0? Which implies new arch requires new drivers from scratch, a lot of work, but par for course.
HDR TVs.

This was the Raja quote I was thinking about.
https://www.reddit.com/r/Amd/commen...n_technologies_group_at_amd_and_were/dhqnpo5/

"Developing drivers for new architecture is one of the most complex and difficult engineering tasks for a GPU company...In fact this is one of the reasons why there are only so few GPU companies."

 

[Bouowmx]

From reddit:


https://www.reddit.com/r/Amd/comments/6kpdzo/not_to_add_to_the_doomsday_mood_of_ramd_but_the/

Not confirmed

The one previously leaked roadmap put AMD Raven Ridge PGA in 2018, so not officially a delay: https://forums.anandtech.com/threads/raven-ridge-news-release-date.2503930/

 

[Elixer]

This was the Raja quote I was thinking about.
https://www.reddit.com/r/Amd/commen...n_technologies_group_at_amd_and_were/dhqnpo5/

"Developing drivers for new architecture is one of the most complex and difficult engineering tasks for a GPU company...In fact this is one of the reasons why there are only so few GPU companies."

While that is somewhat true, it is also true that they have been working on drivers ever since they started with the simulations.
This is how they get their baseline performance level, and then they decide what needs to be optimized, rinse & repeat.
Then, after first silicon is back, they keep optimizing the drivers, run it through test suites, and compare that to the baseline they had in the simulations.
If they don't match, they see what went wrong, and depending on the problem, could result in a respin of the silicon.

It would be pretty hard to not know the ballpark performance level after a few iterations of the above, unless the simulator was incredibly wrong, and nobody would reasonably think that basic driver optimizations could get you a ton of extra performance.
Game specific optimizations are a different animal, there, you pretty much know what needs to be improved, and you can work on those things individually.

 

[.vodka]

https://twitter.com/ryanshrout/status/881317151277285376
https://www.pcper.com/news/Graphics-Cards/Radeon-Vega-Frontier-Edition-GPU-and-PCB-Exposed

(Editor's Update: we have updated the die measurements after doing a remeasure. I think my first was a bit loose as I didn't want to impact the GPU directly.)

• Die size: 25.90mm x 19.80mm (GPU only, not including memory stacks)
  • Area: 512.82mm2
• Package size: 47.3mm x 47.3mm
  • Area: 2,237mm2

So, that's much closer to Raja's statement that Vega is smaller than 500mm². It is probably around that mark... It's much better than the wrong 564mm² measurement we had.

At least it's less shocking now when compared to GP104 and GP102. 500mm² is still much bigger than what a simple Fiji shrink to 14nm would have ended up at... Still, we'll see how RX Vega does at gaming in a month. At least on the pro side of things, FE Vega is a great product for what it costs.

Vega in general should also get a nice boost over its useful life thanks to driver tweaks and optimizations, same as GCN1-4 did these past six years. Who knows, RX Vega maybe gets some of that driver magic at launch... I don't really think it'll be much better, but any improvement is welcome.

What worries me more is that rumor of Raven Ridge being delayed because of Vega...

 

[Elfear]

New Vega benches!!!!!
https://youtu.be/YflRl9np70k

Game mode pro mode and game mode oc tested!
All vs the gtx 1080
Spoiler : the performance looks way better with the oc and looks promising that there is still things amd can do to get more performance

Something is off in those benchmarks. Vega is faster in Pro Mode (PM) than in Game Mode (GM) which seems contrary to how it should be. Either there are no game optimizations in GM and/or it's running at a lower clock for some reason. On average, PM was 1.1% faster than GM (2.7% faster if you ignore Unigine Valley and Cinebench).

Also, Vega seems to be memory bandwidth bottlenecked as there was a large uptick in performance with only a 6% oc on the core (16% on memory). Average performance increase when overclocked was almost 13%.

 

[iBoMbY]

The one previously leaked roadmap put AMD Raven Ridge PGA in 2018, so not officially a delay: https://forums.anandtech.com/threads/raven-ridge-news-release-date.2503930/

Raven Ridge for Desktop/AM4 was 2018 for a very long time. And Raven Ridge for mobile is still H2/2017.

 

[tential]

Something is off in those benchmarks. Vega is faster in Pro Mode (PM) than in Game Mode (GM) which seems contrary to how it should be. Either there are no game optimizations in GM and/or it's running at a lower clock for some reason. On average, PM was 1.1% faster than GM (2.7% faster if you ignore Unigine Valley and Cinebench).

Also, Vega seems to be memory bandwidth bottlenecked as there was a large uptick in performance with only a 6% oc on the core (16% on memory). Average performance increase when overclocked was almost 13%.

Did you see my conclusion? I think there is a more amd can do. Obviously I didn't spell it out much I think the numbers and video speaks for itself. If you clock this card right with good cooling, and if amd truly is running this card on junk drivers currently, we may not have seen all this card has too offer.

Amd is known for massive screw ups and that's a good and bad thing. It's a bad thing for the perception of the company, but for those of us in the know, we know the underlying potential.

From what I've seen, this card isn't good. Period. It's so bad I don't even know if I can justify getting it. But from my gut, there is no way amd did worse than Fiji. There has to be more and amd is the type of company to completely screw this up at launch and have it working correctly later down the line.

Either way, once you see those oveclocked benches and factor in the pro mode better than game mode... There is hope. It's not much, but it's there.

Personally, Vega 10 will be far worse competition wise than Vega 11.
My own speculation is that Vega 11 with the power savings that the 11 nomenclature seems to designate will actually allow the card to compete with the Volta card.
Ive always expected Vega to end up between a 1080 and 1080ti and the analyst forecaster in me says that I couldn't be wrong. It won't be a good card, but once you factor in Freesync, which was always my gambit from the beginning, the price and performance of the card becomes almost irrelevant(I do worry whether I'll get adequate enough performance at 4k to hit minimums).

 

[Oddzz]

Regarding the possible driver issues. Is it possible that AMD put the vast majority of its coder into the ROCm development (link to German article about it) so they get it ready for Radeon Instinct? So my theory is that they left the driver development alone when they had a working but unoptimized driver for Vega and focused all their resources into their deep learning framework.
Are those two fields (driver and framework development) are even correlated in a way that they share resources? Please don't stone me if what I've wrote doesn't makes any sense at all I know it's very speculative and might be far from the truth.

 

[Cloudfire777]

GamersNexus tested the FE Vega. Basically this is what they discovered

- Gaming mode adds 15% extra performance. Which means it will indeed run the same driver as any Vega gamer card.

- Power consumption is about the same as Titan Xp. Performance between the two is worlds apart though.

- Vega FE is 18% slower than GTX 1080 Reference in the games they tested.

 

[OatisCampbell]

GamersNexus tested the FE Vega. Basically this is what they discovered

- Gaming mode adds 15% extra performance. Which means it will indeed run the same driver as any Vega gamer card.

- Power consumption is about the same as Titan Xp. Performance between the two is worlds apart though.

- Vega FE is 18% slower than GTX 1080 Reference in the games they tested.

I don't think that means anything in regard to the RX,

It could just as easily mean it's a preliminary gaming driver and they needed time to QA a newer, better driver.

The hardware is almost certainly the same, and it would seem an AIO or open HSF is in order for Vega.

 

[french toast]

Raja has confirmed the die size is 484mm2.
If we can get 25℅ with Rx drivers disaster will be averted I feel, still won't make it a great card but it won't be a contender for worst gpu ever.

 

[IEC]

Raja says die size is closest perfect square to 500mm² = 484mm² die size (22²):
https://twitter.com/GFXChipTweeter/status/881370308694822913

PCPer and GamersNexus both overestimated the die size...

For comparison:
Polaris 10 = 232mm² (14nm)
Fiji = 596mm² (28nm)
Hawaii = 438mm² (28nm)

 

[majord]

Doesn't seem to be enough chatter around the The 4K results, which are particularly startling, and nonsensical to me..

at 1080p, 1440p even, The GTX1080 (or slightly below) levels of performance @ the 1400-or-so Mhz it seems to be running at during these work loads make perfect sense.. albeit still problematic, since it really needs to clock higher, but still - Firestrike, timespy, scores all check out.

@ 4K though it's falling flat on its face for no known reason. This should be it's strength, not it's weakness. There are instances where the card in its current forum is barely faster than a Fury X @4K that are head scratching. What is going on here?