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

[HurleyBird]

FuryX doesn't have pro (or semi-pro) drivers, so it's not a valid comparison.

That said, my money is on drivers rather than architecture causing the ipc regression.

 

[raghu78]

That graph is only relevant to those applications/tests which is nothing to do with games. The other graphs indicate a zero architecture improvement for games at the same clock.

The specviewperf comparisons show huge compute performance improvements across a suite of workloads. This proves that the Vega architecture is performing very well for professional workloads which confirms compute performance and IPC has gone up significantly. For gaming we have to wait for RX Vega and launch drivers to draw conclusions.

 

[Head1985]

Told you that compute performance per clock is higher in Vega than it is in Fiji.

In games it shows reduction per clock versus Fiji.

http://www.gamersnexus.net/guides/2977-vega-fe-vs-fury-x-at-same-clocks-ipc
Food for thought.

Yeah vega is just memory bandwidth bottleneck in games.They should use 3stacks with 768gb/s atleast.I wonder how much tile rasterizer will increase memory bandwidth after they turn it on.Probably around 30%.AMD is way way behind NV in delta color compression.Vega at 99% still using old polaris compression whitch is bad compare to pascal one.

 

[Glo.]

Yeah vega is just memory bandwidth bottleneck in games.They should use 3stacks with 768gb/s atleast.I wonder how much tile rasterizer will increase memory bandwidth after they turn it on.Probably around 30%.AMD is way way behind NV in delta color compression.Vega at 99% still using old polaris compression whitch is bad compare to pascal one.

Three things. 512 GB/s is enough to feed the cores. There is tile-based rasterization which also should lower the requirements for memory bandwith, and Vega has HBCC.

I am afraid that what you describing, and what we see in games has zero to do together.

 

[lobz]

If the 1080ti makes the titan xp irrelevant, then swap the 1080ti out to use instead and the 1080ti is even better vs the fe.....

except you're wrong. AMD tried to show that while Titan is irrelevant, there IS a way to make a great prosumer card without outrageous pricing. And let's all just wait till RX driver gets released and see if it helps the FE's gaming performance.

Sent from my VTR-L09 using Tapatalk

 

[lobz]

Yeah vega is just memory bandwidth bottleneck in games.They should use 3stacks with 768gb/s atleast.I wonder how much tile rasterizer will increase memory bandwidth after they turn it on.Probably around 30%.AMD is way way behind NV in delta color compression.Vega at 99% still using old polaris compression whitch is bad compare to pascal one.

How could you POSSIBLY know how much Vega's color compression is behind Nvidia's? What are you, a wizard?

Sent from my VTR-L09 using Tapatalk

 

[Paratus]

Told you that compute performance per clock is higher in Vega than it is in Fiji.

In games it shows reduction per clock versus Fiji.

http://www.gamersnexus.net/guides/2977-vega-fe-vs-fury-x-at-same-clocks-ipc
Food for thought.

That shows some interesting things.

Throwing out the snx-02 outlier we get the following comparisons

• Vega@1050 vs Fury X: 66% faster
• Vega@1600 vs Vega@1050: 25%faster
• Vega@1600 vs Fury X: 106% faster

So it appears that Vegas architecture is significantly faster than Fury.

Depending on our assumptions Vega has:

• 11.9B xtors (33% more) over Furys 8.9B assuming Polaris densities
• Upto 14.3B (60% more) if we allow for a 20% increase in density (same increase AMD had from Gen 1 to Gen 2 on 28nm)

This suggests the transistor budget provides a greater than 1:1 performance improvements over Fury.

What's also interesting is the clockspeed comparison.

• 1600/1050 = 152%

So Vega only gets about 1% of performance increase per 2% increase in clock speed compared to itself, (25% increase from 1050 to 1600)

Compared to Fury the 52% clock speed increase equals another 40% performance over Fury.

If the 106% performance increase over Fury in calculations transferred over to gaming then RX Vega would be faster than the 1080Ti

Somehow that doesn't seem to be happening with the FE. We've seen what, a 25%-40% increase over Fury?


So what's the deal? My only thoughts are

• Gaming Driver issues
• Significant thermal throttle
• Significant memory bandwidth bottlenecks
• ????

 

[Bacon1]

How could you POSSIBLY know how much Vega's color compression is behind Nvidia's? What are you, a wizard?

Sent from my VTR-L09 using Tapatalk

It would make sense for Vega to still be behind, but they should have at least improved over Fiji / Polaris.

http://techreport.com/review/31562/nvidia-geforce-gtx-1080-ti-graphics-card-reviewed/3

http://techreport.com/review/30328/amd-radeon-rx-480-graphics-card-reviewed/5

Random = not much compression and just raw bandwidth, black = easily compressed.

Not sure why the 980 goes from 286 -> 407 from the 480 to 1080 Ti review though...

But Nvidia has been ahead for a while now and AMD has made strides with Polaris and hopefully again with Vega. Not sure if anyone has tested it yet but hopefully someone will and again after RX launches.

 

[Det0x]

That shows some interesting things.

Throwing out the snx-02 outlier we get the following comparisons

• Vega@1050 vs Fury X: 66% faster
• Vega@1600 vs Vega@1050: 25%faster
• Vega@1600 vs Fury X: 106% faster

So it appears that Vegas architecture is significantly faster than Fury.

Depending on our assumptions Vega has:

• 11.9B xtors (33% more) over Furys 8.9B assuming Polaris densities
• Upto 14.3B (60% more) if we allow for a 20% increase in density (same increase AMD had from Gen 1 to Gen 2 on 28nm)

This suggests the transistor budget provides a greater than 1:1 performance improvements over Fury.

What's also interesting is the clockspeed comparison.

• 1600/1050 = 152%

So Vega only gets about 1% of performance increase per 2% increase in clock speed compared to itself, (25% increase from 1050 to 1600)

Compared to Fury the 52% clock speed increase equals another 40% performance over Fury.

If the 106% performance increase over Fury in calculations transferred over to gaming then RX Vega would be faster than the 1080Ti

Somehow that doesn't seem to be happening with the FE. We've seen what, a 25%-40% increase over Fury?


So what's the deal? My only thoughts are

• Gaming Driver issues
• Significant thermal throttle
• Significant memory bandwidth bottlenecks
• ????

Well it's not 550MHz because the card actually just stays at around 1400MHz most of the time, so it's more of a 350MHz / 33% difference.

I vote number 1 regarding whats the issue is.. Seems like the current Vega FE driver is optimized for "professional workloads", while its running in "Fiji fallback driver path" for gaming workloads, however that's possible..

I really hope for a major performance increase in gameing with the RX driver, otherwise iam forced to bite the bullet and wait for gv102

 

[Snarf Snarf]

Less than a month to go, here's hoping that one of the AIB partner's employee's give us a juicy leak on one of the premium cards like MSI Gaming X or Zotac AMP! Extreme.

The throughput comparison to Fiji is really interesting though because in compute it's showing massive gains over Fury X. This is starting to feel a bit like the Ryzen launch, software is going to have to be patched to take full advantage of the massive change in architecture going on. This actually worries me the most, they have barely any market share and until XBO-X is out there isn't any other Vega in the wild to help push the incentive to patch optimizations into games. A lot of the features they had touted looked to be transparent to developers but after the AMA with Raja it's looking more and more like this uarch is going to take 4-5 months to reach peak performance which puts it squarely against Volta.

RTG better have one of their close developers working on a patch for a AAA title within a couple of days of launch so they can push the narrative that once engines utilize the features that Vega is a good architecture. I'm thinking DICE or Bethesda here, maybe BF1 or Prey patch on launch day showing DX12 performance equal to or greater than 1080ti at 1440p and 4k with minimums far above. If they can show consumers that level of performance I think they'll have a real winner on their hands

 

[Head1985]

It would make sense for Vega to still be behind, but they should have at least improved over Fiji / Polaris.

http://techreport.com/review/31562/nvidia-geforce-gtx-1080-ti-graphics-card-reviewed/3

http://techreport.com/review/30328/amd-radeon-rx-480-graphics-card-reviewed/5

Random = not much compression and just raw bandwidth, black = easily compressed.

Not sure why the 980 goes from 286 -> 407 from the 480 to 1080 Ti review though...

But Nvidia has been ahead for a while now and AMD has made strides with Polaris and hopefully again with Vega. Not sure if anyone has tested it yet but hopefully someone will and again after RX launches.

AMD didnt say a word about better delta color compresion.So i think they still using polaris one whitch is worse than maxwell one and far worse than pascal one.

 

[Bacon1]

AMD didnt say a word about better delta color compresion.So i think they still using polaris one whitch is worse than maxwell one and far worse than pascal one.

I thought I'd seen something in one of the slides from many months ago but yeah, they have been behind a lot on compression which has hurt them as their cards have all been memory bandwidth bottlenecked for a while now, even Polaris is often. So I'm assuming they would have been working on it for Vega with all the other changes, but guess we won't find out til after RX launches and TechReport runs their tests. Not sure if any other sites run them as well which is sad, as they are good to see how archs change.

 

[Cloudfire777]

So according to videocardz:
RX Vega scores about what stock GTX 1080 does.
Overclocked RX Vega equals an overclocked GTX 1080.

Pretty much what many here thought was realistic. Not a 30-40% gain because of some vague fantasy excuse about unused feature that is some secret Vega RX sauce. But a 15% ish gain from FE Vega, mostly from higher clocks.

AMD better be prepared to sell these for $400 if they are gonna have a choice of moving units.

Most have moved on from the GTX 1070/1080 tier a looong time ago and have bought and played with those GPUs so many months now and the only reason to get a Vega was a GTX 1080Ti performance for a reasonable price.
Since AMD is looking to screw the pooch bigtime with Vega, I think waiting for Volta maybe late this year is the next logical step.

Or Navi, in like 2020. Im sure Raja and the team got plans for future amazing reveals about Navi for dedicated people though. First up will be an event about an event about a Navi event 1 year from now. Followed by the amazing reveal that they plan to make, not just 1 but 2 (!!!) Navi cards. Followed by an event that will reveal the names of these cards 1 year from that. The actual launch will be far from the expectations people had, and they will be stuck waiting for years while Nvidia buyers get to play with their Volta cards all that time.
But it will be worth it. After all its AMD. (tongue in cheek)

 

[Qwertilot]

A lot of the features they had touted looked to be transparent to developers but after the AMA with Raja it's looking more and more like this uarch is going to take 4-5 months to reach peak performance which puts it squarely against Volta.

They're utterly doomed in that comparison though. Vs Pascal there's the odd niche to aim at, esp if the Rx does improve quite a bit.

If Volta is anything like expectation it'll stamp on those very firmly.

 

[Guru]

No one will buy either for pro tasks, they will buy a quadro because certified drivers are much more important then a little extra performance. The whole Vega FE for pro is a complete red herring fired out their by AMD marketing to make Vega look better at something. Both the Titan and the Vega FE are gaming cards. The titan sells because it is the absolute fastest, Vega FE won't because it isn't.

This is false. The Vega FE offers great performance at applications, at computing, etc... for those who do NOT need certified drivers, who do not dabble in designing engines that lives depend on or designing vaccine or stuff like that, that do need a proper GPU for development, computing, VR, etc.... but again don't want to pay $2000 or $5000 or whatever for a professional card with certified drivers that probably performs around the same level, this is the card.

Vega FE wins in MOST of the application benchmarks, it does lose some to the Titan XP, but again wins out in most of them. Its also $200 cheaper than the Titan XP.

Again this card CAN be used for gaming, it has driver support for it, so games won't crash or won't start even playing on it, but its a card optimized for workstations that again do not need certified software and can't pay $3000 for a card with the same performance, but way more expensive because of the certified drivers.

Do I think AMD has some magic card that performs 50% faster than the Vega FE and beats the 1080ti easily? NO, that is not realistic, but add in 5% from driver optimizations, add in 5% due to sustaining 1600MHz clock over the 1350Mhz clock the FE is doing, add in lets call it will of faith that AMD has a feature like rasterization that is fully and properly implemented with the RX Vega and say 5% performance gain from that and we are looking at 15% better performance which would put it above the 1080, though behind the 1080ti, but with the right pricing it can compete against the 1080 and cut down versions compete against the 1070.

Again not the 1080ti killer we all were expecting, but still a competitor in some of the high end segment against Nvidia, more options is always better.

Still with a competitive field like this, Raja should be given the door after Vega. Clearly he couldn't deliver and even a realistic 15% gain over Vega FE is not good enough, not when the belief was a competitor to the 1080ti.

No one thought it would be greatly faster than the 1080ti, I think realistically everyone knew that 1080ti level is what it seems like, but we are ending up with a 1080 competitor, which makes this card a failure because they are using HBM2, so more expensive, less available product compared to GDDR5x, but ultimately worthless.

 

[Glo.]

That shows some interesting things.

Throwing out the snx-02 outlier we get the following comparisons

• Vega@1050 vs Fury X: 66% faster
• Vega@1600 vs Vega@1050: 25%faster
• Vega@1600 vs Fury X: 106% faster

So it appears that Vegas architecture is significantly faster than Fury.

Depending on our assumptions Vega has:

• 11.9B xtors (33% more) over Furys 8.9B assuming Polaris densities
• Upto 14.3B (60% more) if we allow for a 20% increase in density (same increase AMD had from Gen 1 to Gen 2 on 28nm)

This suggests the transistor budget provides a greater than 1:1 performance improvements over Fury.

What's also interesting is the clockspeed comparison.

• 1600/1050 = 152%

So Vega only gets about 1% of performance increase per 2% increase in clock speed compared to itself, (25% increase from 1050 to 1600)

Compared to Fury the 52% clock speed increase equals another 40% performance over Fury.

If the 106% performance increase over Fury in calculations transferred over to gaming then RX Vega would be faster than the 1080Ti

Somehow that doesn't seem to be happening with the FE. We've seen what, a 25%-40% increase over Fury?


So what's the deal? My only thoughts are

• Gaming Driver issues
• Significant thermal throttle
• Significant memory bandwidth bottlenecks
• ????

We do not know, whether Vega reached the theoretical maximum clock speed. Most likely, as tests done by PCPer shown - it did not. Most likely it hovered between 1430 and 1480 MHz, during the tests.

IMO, if anyone will ask me what is happening.

Drivers are properly reporting the features, however the software has to be rewritten by developers to utilize it. Similar situation to Ryzen, where non-compute oriented software had to be rewritten to increase performance on Ryzen platform.

 

[Phynaz]

We do not know, whether Vega reached the theoretical maximum clock speed. Most likely, as tests done by PCPer shown - it did not. Most likely it hovered between 1430 and 1480 MHz, during the tests.

IMO, if anyone will ask me what is happening.

Drivers are properly reporting the features, however the software has to be rewritten by developers to utilize it. Similar situation to Ryzen, where non-compute oriented software had to be rewritten to increase performance on Ryzen platform.

So my entire Steam library of 1,700 games needs to be re-written to take advantage of Vega?

I think you are in for a lot of disappointment if you are expecting that to happen.

 

[Glo.]

So my entire Steam library of 1,700 games needs to be re-written to take advantage of Vega?

I think you are in for a lot of disappointment if you are expecting that to happen.

I think the best answer we would get about the optimizations is if someone would test Prey on Vega Frontier Edition.

And I think you can know why...

 

[Bacon1]

So according to videocardz:
RX Vega scores about what stock GTX 1080 does.
Overclocked RX Vega equals an overclocked GTX 1080.

Pretty much what many here thought was realistic. Not a 30-40% gain because of some vague fantasy excuse about unused feature that is some secret Vega RX sauce. But a 15% ish gain from FE Vega, mostly from higher clocks.

Where are these 15% higher clocks? Its 1630mhz vs 1600mhz(maybe).

http://www.3dmark.com/compare/3dm11/12256250/3dm11/12256169

The GPU shows as the same, the difference is the CPU.

Oh wait... 3dMark doesn't properly show clocks... which is why people are comparing Vega FE @ 1400ish but stating its 1600 and like these we have no idea what actual clocks were used as 3dMark only shows the driver clocks not actual testing clock rates.

 

[Cloudfire777]

Where are these 15% higher clocks? Its 1630mhz vs 1600mhz(maybe).

http://www.3dmark.com/compare/3dm11/12256250/3dm11/12256169

The GPU shows as the same, the difference is the CPU.

Oh wait... 3dMark doesn't properly show clocks... which is why people are comparing Vega FE @ 1400ish but stating its 1600 and like these we have no idea what actual clocks were used as 3dMark only shows the driver clocks not actual testing clock rates.

Who said Vega FE runs 1600MHz all the time? Nobody because it doesnt.
Here comes Vega RX. Told you it would have higher sustained clocks. Doubt it runs 1600MHz locked though.

Its AMDs only choice. Cant go much higher because of power/heat/silicon etc.

 

[Det0x]

Rys (AMD RTG) said:

The "Fiji fallback driver" or "Fiji drivers" meme needs to stop. That's not how it works or should be described. Otherwise you should start calling Volta drivers Pascal drivers Maxwell drivers. There is obviously commonality -- that's just how software engineering needs to work for a GPU -- but calling it a Fiji driver or Fiji fallback is wrong.

https://forum.beyond3d.com/threads/...rs-and-discussion.59649/page-138#post-1990723

 

[iwulff]

I thought I'd seen something in one of the slides from many months ago but yeah, they have been behind a lot on compression which has hurt them as their cards have all been memory bandwidth bottlenecked for a while now, even Polaris is often. So I'm assuming they would have been working on it for Vega with all the other changes, but guess we won't find out til after RX launches and TechReport runs their tests. Not sure if any other sites run them as well which is sad, as they are good to see how archs change.

Polaris showed good improvements over previous versions of GCN in regards to DCC. I also recall saw them showing further improved for compression. I don't think Vega is as important for gaming to AMD compared to workstation/APU. Sure it would be good if it would compete for the first place, but there are other places where it is possible to gain much more financially. I'm sure Ryzen mobile will bring a new era, where you can play most games decently and with great efficiency without a dgpu. Can't wait for 7nm and what this means for APU development.

 

[Cloudfire777]

Rys (AMD RTG) said:

The "Fiji fallback driver" or "Fiji drivers" meme needs to stop. That's not how it works or should be described. Otherwise you should start calling Volta drivers Pascal drivers Maxwell drivers. There is obviously commonality -- that's just how software engineering needs to work for a GPU -- but calling it a Fiji driver or Fiji fallback is wrong.

https://forum.beyond3d.com/threads/...rs-and-discussion.59649/page-138#post-1990723

I love how everytime someone that have actually worked on or knows the products explains how clueless all the average joe's are with their funny out-of-this-world theories

 

[maddie]

It shows that throughput of the cores is actually higher than in Fiji, which also shows that registry file sizes, and cache size has been increased, which should affect gaming performance, also.

What we see in games is that performance per clock vs Fiji - decreases.

It should not be this way.

I agree wholeheartedly with this post.

Anyone looking at the increase vs Fiji at similar clocks and belittling it is either biased or unable to analyze properly.

We should always remember that it's impossible to get more than what an architecture can theoretically deliver, but quite possible to get less. Vega in this case, shows a large improvement at similar clocks and clearly indicates some POSSIBLE massive improvements as not being completely unrealistic.

 

[railven]

Rys (AMD RTG) said:

The "Fiji fallback driver" or "Fiji drivers" meme needs to stop. That's not how it works or should be described. Otherwise you should start calling Volta drivers Pascal drivers Maxwell drivers. There is obviously commonality -- that's just how software engineering needs to work for a GPU -- but calling it a Fiji driver or Fiji fallback is wrong.

https://forum.beyond3d.com/threads/...rs-and-discussion.59649/page-138#post-1990723

Well now I'm even more confused. Because the resident AMD guys here are saying Vega FE is performing poorly because January drivers, inactive features, and being read/performing like a Fiji card.

But AMD is saying the drivers aren't gimped (I interpreted as features aren't disabled/missing), the drivers are old, but AdoreTV recent video shows they aren't January old, and this is not falling back to Fiji mode.

This can of worms is getting just uglier as the days go by.