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

[Guru]

They need to be competitive, otherwise their GPU division will be up in flames. Good thing is they have the GCN 4 architecture RX 500 series and it seems to be a really good architecture. If they scaled that architecture up, by adding 2300 more shaders, increasing clocks to 1600-1700MHz they can easily gain double in performance.

Hopefully with driver updates and game optimizations they can increase performance up to 15%, but yeah it doesn't look too good.

 

[Veradun]

They need to be competitive, otherwise their GPU division will be up in flames. Good thing is they have the GCN 4 architecture RX 500 series and it seems to be a really good architecture. If they scaled that architecture up, by adding 2300 more shaders, increasing clocks to 1600-1700MHz they can easily gain double in performance.

And they can melt the floor, the crust and easily reach the stiffer mantle.

 

[beginner99]

They've done exactly that. It was Raja speaking on his reddit AMA.

I disagree. "Vega FE is not for gamers" and "RX Vega will have significant higher gaming performance" are completely different in their meaning.

The first is "manager speak" which leads so much room for interpretation and leaves an a way out to backpedal. Could mean as much that it is a lot cheaper at same performance. It does not indicate anything about it having significantly (at least 10% rather 30%) higher performance. And that is why this launch was fail. Clear statement would have prevented any issue. But I suspect this is all intentional as in "there is no bad publicity" or Vega really sucks badly.

 

[wanderica]

Guess what? You have to wait for its launch. How strange :>

The architecture is the same. Is it not? History tells us not to expect too much. All of this really feels like a lot of people are waiting on that "miracle patch" to arrive in the last hour before launch. I've never known a game studio to pull that off, and I seriously doubt AMD will do it either with hardware, no less. It might be drivers. It might even be the respin talked about earlier. Regardless, I agree we'll see at least some uptick in performance. I really hope I'm wrong, but I remain doubtful that AMD will somehow pull a rabbit out of their hat in a month's time when the last 6 haven't been enough. Everything AMD has shown us up until the launch of Vega FE has failed to live up to the hype. Maybe that's on us for expecting too much, but AMD certainly hasn't helped by comparing their products to Titan XP, and even unreleased next gen products.

To be clear, I don't have an issue with the performance. Priced correctly, RX Vega will make a lot of people happy even if it doesn't improve, but Vega FE has been marketed poorly. Why is "wait for RX" not more bad AMD marketing hype?

 

[Veradun]

I disagree. "Vega FE is not for gamers" and "RX Vega will have significant higher gaming performance" are completely different in their meaning.

The first is "manager speak" which leads so much room for interpretation and leaves an a way out to backpedal. Could mean as much that it is a lot cheaper at same performance. It does not indicate anything about it having significantly (at least 10% rather 30%) higher performance. And that is why this launch was fail. Clear statement would have prevented any issue. But I suspect this is all intentional as in "there is no bad publicity" or Vega really sucks badly.

"Consumer RX will be much better optimized for all the top gaming titles and flavors of RX Vega will actually be faster than Frontier version!"
https://www.reddit.com/r/Amd/commen...n_technologies_group_at_amd_and_were/dhqmwfj/

"Can you play games on Frontier Edition? Yes, absolutely. It supports the RX driver and will deliver smooth 4K gaming. But because it is optimized for professional use cases (and priced accordingly), if gaming is your primary reason for buying a GPU, I’d suggest waiting just a little while longer for the lower-priced, gaming-optimized Radeon RX Vega graphics card"
https://www.reddit.com/r/Amd/commen...n_technologies_group_at_amd_and_were/dhqnsgo/

"RX will be fully optimized gaming drivers, as well as a few other goodies that I can't tell you about just yet"
https://www.reddit.com/r/Amd/commen...n_technologies_group_at_amd_and_were/dhqot92/

Enjoy.

 

[Gideon]

The architecture is the same. Is it not? History tells us not to expect too much. All of this really feels like a lot of people are waiting on that "miracle patch" to arrive in the last hour before launch. I've never known a game studio to pull that off, and I seriously doubt AMD will do it either with hardware, no less. It might be drivers. It might even be the respin talked about earlier. Regardless, I agree we'll see at least some uptick in performance. I really hope I'm wrong, but I remain doubtful that AMD will somehow pull a rabbit out of their hat in a month's time when the last 6 haven't been enough. Everything AMD has shown us up until the launch of Vega FE has failed to live up to the hype. Maybe that's on us for expecting too much, but AMD certainly hasn't helped by comparing their products to Titan XP, and even unreleased next gen products.

To be clear, I don't have an issue with the performance. Priced correctly, RX Vega will make a lot of people happy even if it doesn't improve, but Vega FE has been marketed poorly. Why is "wait for RX" not more bad AMD marketing hype?

Agreed. Based on the vagueness of the PR statements, I'm more inclined to believe that the main performance benefits of RX Vega will be:

1. Higher clocks, as the top model will probably be watercooled and can sustain the ~1600mhz clock
2. More bandwidth. Hopefully 4-hi stacks of 8GB and 4GB can reach the previosuly mentioned 512GB/s bandwidth, giving a significant boost on it's own.

The drivers will obviously improve as well, but I'd be really hesitant, to think it will be more than 5-10%, with more in really bad examples such as GTA. I wouId be really realy happy to be proven wrong though!

Some context (before someone starts to blame me for a shill): The last components of my Ryzen 1700X rig will arrive this Saturday. What do you think is the only recycled component in that build? That's right, my GTX 970 from 2014! I really wanted to build an all-AMD rig for a change, but it looks less and less likely that I can justify buying Vega for it. On top of that, I've been withholding buying a new monitor for the same question.

Maybe this puts my aggravation @AMDs marketing in a bit better context ...

 

[Gideon]

"Consumer RX will be much better optimized for all the top gaming titles and flavors of RX Vega will actually be faster than Frontier version!"
https://www.reddit.com/r/Amd/commen...n_technologies_group_at_amd_and_were/dhqmwfj/

"Can you play games on Frontier Edition? Yes, absolutely. It supports the RX driver and will deliver smooth 4K gaming. But because it is optimized for professional use cases (and priced accordingly), if gaming is your primary reason for buying a GPU, I’d suggest waiting just a little while longer for the lower-priced, gaming-optimized Radeon RX Vega graphics card"
https://www.reddit.com/r/Amd/commen...n_technologies_group_at_amd_and_were/dhqnsgo/

"RX will be fully optimized gaming drivers, as well as a few other goodies that I can't tell you about just yet"
https://www.reddit.com/r/Amd/commen...n_technologies_group_at_amd_and_were/dhqot92/

Enjoy.

Let's wait and see ... if the driver alone can indeed improve performance more than say, 15%, on top of other improvements (higher clocks, hopefully more BW), I would be a very happy puppy.

So far AMD has said that some flavors of RX Vegas will be faster (obviously, at the very least the watercooled ones). RX Vega will be better for gaming (this is true because of price alone). And that it will have a better optimized driver, yes ... but nowhere have they even hinted, that the driver will make a huge difference.

I hope it does, I really do, but I still remember getting burned so badly by Bulldozer, despite being more pessimistic than most and having a very low lower-bound for my expectation*, I'm not jumping to the conclusion of a super-mega-driver as many of you seem to take 100% for granted already.

*IPC about comparable to Thuban @ same process & clocks (lose some, win some), instead they wasted their last good node shrink for years (45nm -> 32nm), being hardly competitive to their own old turboless (Thuban turbo was a joke) 6 core @ similar or slightly worse power-usage, despite having higher clocks, a die-shrink and a totally new arch. An 8-core shrink with turbo (Llano cores) would have absolutely decimated Bulldozer. Yet many clinged to a marketing guys "20% more throuhput promise", as a holy grail for about as much more IPC.

 

[Veradun]

Let's wait and see ... if the driver alone can indeed improve performance more than say, 15%, on top of other improvements (higher clocks, hopefully more BW), I would be a very happy puppy.

So far AMD has said that some flavors of RX Vegas will be faster (obviously, at the very least the watercooled ones). RX Vega will be better for gaming (this is true because of price alone). And that it will have a better optimized driver, yes ... but nowhere have they even hinted, that the driver will make a huge difference.

I hope it does, I really do, but I still remember getting burned so badly by Bulldozer, despite being more pessimistic than most and having a very low lower-bound for my expectation*, I'm not jumping to the conclusion of a super-mega-driver as many of you seem to take 100% for granted already.

*IPC about comparable to Thuban @ same process & clocks (lose some, win some), instead they wasted their last good node shrink for years (45nm -> 32nm), being hardly competitive to their own old turboless (Thuban turbo was a joke) 6 core @ similar or slightly worse power-usage, despite having higher clocks, a die-shrink and a totally new arch. An 8-core shrink with turbo (Llano cores) would have absolutely decimated Bulldozer. Yet many clinged to a marketing guys "20% more throuhput promise", as a holy grail for about as much more IPC.

Well, actually he said "flavours of RX" not "some flavours of RX" but who cares. The only thing I care about is price/performance. If they deliver there I'm in as I'm still on a HD6870 with an ASUS MG279Q and stuck to mobas

 

[96Firebird]

Idiots reject what they see. Wise man reject what they think and believe. - Buddhist proverb.

Yet you yourself seem to reject all the benchmarks and data provided for Vega FE and can come up with a million and a half excuses (that you think and believe) to justify those results.

Please, heed your own advice.

 

[parkerface]

This thread sucks. The sheer amount of information cherry picking being made almost defies belief. Then pile on top of that the same tired strawman arguments and obtuse rhetoric from the usual suspects... It's almost as bad as the cable news churn.

What happened to maintaining an objective evaluation of the facts?

Why is it so hard for some to just wait and see without jumping to a half-baked conclusion?

And to think there is two more weeks of this to go.

 

[leoneazzurro]

Three, in reality, but I agree with you.

 

[Glo.]

Yet you yourself seem to reject all the benchmarks and data provided for Vega FE and can come up with a million and a half excuses (that you think and believe) to justify those results.

Please, heed your own advice.

Do I reject what I see? Do I reject that I see in tests made specifically to test Tile-Based Rasterization it is non apparent, despite the fact it is the technique Vega uses?

Do I reject the fact that tests from Doom made by Gamers Nexus, contradict what AMD has been demoing with this GPU?

I know it is easy for people to jump to conclusions, but that is the base of what I have written.

P.S. There is a review site which tests Doom, in 4K with TSAA on.
Im sure you will know which one site is it:

Based on the IDTech 6 engine by ID Software, the game takes advantage of OpenGL 4.5 and Vulkan. We tested the game with Vulkan as this is the first AAA title that works reliably on the new cross-platform 3D API. Also, TSSAA was active to enable async compute.

 

[stahlhart]

This thread sucks. The sheer amount of information cherry picking being made almost defies belief. Then pile on top of that the same tired strawman arguments and obtuse rhetoric from the usual suspects... It's almost as bad as the cable news churn.

What happened to maintaining an objective evaluation of the facts?

Why is it so hard for some to just wait and see without jumping to a half-baked conclusion?

And to think there is two more weeks of this to go.

Three, in reality, but I agree with you.

Then stop derailing the thread and contribute some of the objective evaluation you believe is lacking here.
-- stahlhart

 

[tential]

https://youtu.be/zeoiPVVUnl8

Saw this on reddit didn't see it posted here yet.

Looks like this board is "overbuilt" in typical amd fashion judging from comments.
Amd is known for a robust design while Nvidia is known for "skimping" in this regard.

Anyway, interested to see what y'all think of the analysis..
From my skim, looks like the card is built very well unsurprisingly given the past.

If there is a magic driver, then you can throw a ton of power at it with confidence from the looks of this.

Anyway look at the video and see for yourselves any info you can find.

 

[Peicy]

While the old vs. new Doom results are odd indeed, its been quite a long time since we saw that video.
It is actually possible that parts of the new hardware features may not work and would require major reworking which is not feasable for AMD for a variety of reasons. What if their entire tile-based rasterization approach doesnt work as expected for example due to major bugs?
Of course we dont know that, but we dont have any facts pointing to a major increase in performance which would be necessary to make Vega worthwhile. A 10% increase wouldn´t be enough and AMDs statements are so ambigious in light of the current controversy. Vega is bad PR at the moment, and if they knew about a major performance increase (which they should by now), they better communicate that.

What i find strange though is that major features, such as the new rasterization method do not seem to work as of now. You can´t debug what does not work @Glo.

Nvidia gradually went from a good amount of hardware scheduling to a software-based method over the course of years, and they have a lot more ressources than AMD to implement new features such as tile-based rasterization. Apparently, AMD has not yet produced a generalized software package that makes some significant features work in a stable fashion, which is worrying to say the least.
Maybe, and just maybe, it needs major work for each game title on a driver level (and maybe from the game devs themselves) which their wording could imply ("...much better optimized for all the top gaming titles..."), but that would not be good either since AMDs ressources are limited, meaning that a host of games could be in slow "fall back" mode for ages.
Or...some hardware features are bugged, forcing AMD to disable them. Which would be another kind of disaster.


Something else, did anyone try to test the same scene from Prey that they showed with dual Vega on a 1080TI? That was a more recent showing involving games.

 

[Tee9000]

HBM2 on Vega FE is running at over spec at 1.35V even though it's not reaching its intended frequency. Standard is 1.2V. Fiji's HBM1 was 1.3V over spec as well, but it reached its target frequency.

Source.

 

[Cloudfire777]

Agreed. Based on the vagueness of the PR statements, I'm more inclined to believe that the main performance benefits of RX Vega will be:

1. Higher clocks, as the top model will probably be watercooled and can sustain the ~1600mhz clock
2. More bandwidth. Hopefully 4-hi stacks of 8GB and 4GB can reach the previosuly mentioned 512GB/s bandwidth, giving a significant boost on it's own.

The drivers will obviously improve as well, but I'd be really hesitant, to think it will be more than 5-10%, with more in really bad examples such as GTA. I wouId be really realy happy to be proven wrong though!

Some context (before someone starts to blame me for a shill): The last components of my Ryzen 1700X rig will arrive this Saturday. What do you think is the only recycled component in that build? That's right, my GTX 970 from 2014! I really wanted to build an all-AMD rig for a change, but it looks less and less likely that I can justify buying Vega for it. On top of that, I've been withholding buying a new monitor for the same question.

Maybe this puts my aggravation @AMDs marketing in a bit better context ...

Vega FX biggest problem will be sustained clocks because of power I think.
Vega FE already alone consumes 320W max. Compare that with Fury X which consumed about 230W max. Vega FE touched these 320W tops rarely, but was most of the time at 280W during testing.
I think the problem is that the silicon and cooling might have an issue running 320W and 1600MHz all the time.
Higher clocks on the RX than the sustained ~1350MHz for Vega FE, probably, but not locked 1600MHz.

I think most gain over FE will be from the driver. 10-15% over Vega FE perhaps? Should put it close to GTX 1080

 

[french toast]

HBM2 on Vega FE is running at over spec at 1.35V even though it's not reaching its intended frequency. Standard is 1.2V. Fiji's HBM1 was 1.3V over spec as well, but it reached its target frequency.

Source.

Interesting, probably had to buy low binned chips to meet volume? May explain some of the power consumption exceeding expectations if the quality of the first batch of hbm 2 is not meeting yield/ production/ quality targets.

I Do wonder what would have been the result had they just gone with 512bit bus and gddr5x? Probably be seeing this 6 months earlier for similar power consumption, higher performance?
Wouldn't help the software side though.

Edit; I would just like to add if they had followed the nvidia route (and the limitless resources ) and gone with TSMC +gddr5x I believe this chip would be much more competitive out of the gate without any driver optimisation which I'm convinced will arrive at some point.
Hindsight is a wonderful thing and so is nvidia R&D budget

 

[Peicy]

Vega FX biggest problem will be sustained clocks because of power I think.
Vega FE already alone consumes 320W max. Compare that with Fury X which consumed about 230W max. Vega FE touched these 320W tops rarely, but was most of the time at 280W during testing.
I think the problem is that the silicon and cooling might have an issue running 320W and 1600MHz all the time.
Higher clocks on the RX than the sustained ~1350MHz for Vega FE, probably, but not locked 1600MHz.

I think most gain over FE will be from the driver. 10-15% over Vega FE perhaps? Should put it close to GTX 1080

Considering that AMD said that the NCUs are designed for higher clockspeeds, they did not get very far as of yet. A clock speed of 1600 is not a huge jump compared to Polars-based cards.

 

[Tee9000]

Fury X which consumed about 230W max

You mean average game load, not max that was measured at 450W during gaming loops & stress test.

Source.

Vega FE's max power consumption looks really really good compared to Fiji's for now.

 

[crisium]

"Max" power consumption results when achieved via power virus (same as TPU "max") is not even worth the 1 second it takes to view it with your eyes. Useless.

 

[Tee9000]

Open the source link and watch carefully. The max power consumption was achieved while gaming not during the stress test. There are many spikes that go above 300W.

 

[crisium]

And Tom also got ~430W for 980 Ti reference while gaming.

http://www.tomshardware.com/reviews/nvidia-geforce-gtx-980-ti,4164-7.html

Their method of measuring gets results like that.

You can only compare Tom's to Tom's. Or is Vega also much better than 980 Ti at max power consumption? We won't know until Tom measures.

 

[Tee9000]

You can only compare Tom's to Tom's

Fair point.

 

[swilli89]

HBM2 on Vega FE is running at over spec at 1.35V even though it's not reaching its intended frequency. Standard is 1.2V. Fiji's HBM1 was 1.3V over spec as well, but it reached its target frequency.

Source.

Remember it wasn't until recently that AMD shocked the world and said that they would be offering 16GB from 2 stacks of HBM2. I expect the 4GB stacks on the RX Vega to clock higher, on less voltage.