AMD Radeon Vega GPU Architecture

[RejZoR]

If it can't compete with Titan X or 1080 Ti, why would they name it RX Vega and not RX 580/590? And Volta should have x70 running as fast 1080 Ti! Making fun of it with a chip twice as large would be just bad marketing. Also 1080 Ti price surprises me greatly. It could be $800 (and have a $900 FE) if big Vega is really not a threat, and people would still buy it. I doubt we will see top Vega under $600 ... it's big and expensive, so it won't be able to undercut nVidia cards by that much.

I'm still cautiously optimistic, because I know most of their R&D cash was probably spent on Ryzen, but I think another 980 Ti vs Fury X is perfectly doable.

People were saying how Fury X is inferior. But if you look at it today, sure, it can't beat GTX 1080, but there are quite a lot of games where Fury X beats GTX 1070. And where it doesn't, it has a lot higher minimal framerate compared to GTX 980Ti. And I'm not talking about DX12/Vulkan games only. If it's roughly a GTX 1080Ti competetor, that's already job done well.

 

[Captain_Tom]

People were saying how Fury X is inferior. But if you look at it today, sure, it can't beat GTX 1080, but there are quite a lot of games where Fury X beats GTX 1070. And where it doesn't, it has a lot higher minimal framerate compared to GTX 980Ti. And I'm not talking about DX12/Vulkan games only. If it's roughly a GTX 1080Ti competetor, that's already job done well.

All AMD needs to do is accomplish 60% more performance than the Fury X. We already know that it is essentially a Fury X clocked 50% faster, and with 50%+ better memory compression. LOL if the arch is even 10% better IPC (Which Polaris already is), then they will match the bloody 1080 Ti.

 

[the54thvoid]

The optimism is strong for some, which is perfectly fine. I'm not waiting too long before buying my next card but what some are missing here are the signs that give Vega away.
Also, some are missing just how far ahead of the 980ti the Pascal version is. I have been saying for ages it needs to match 1080ti in performance but now that card is out, I don't think it will.
Much like Ryzen, i think Vega will be a far better chip with efficiency. That's the driver for them. I think all the time spent from reveal (Doom gaming demo) to release isn't about clocking higher but clocking more efficiently. It will surely beat the 1080 but just by enough.
And as reassuring to some in team red that think GP102 is a laughable Maxwell refresh, if I were AMD I'd be depressed. That simple refresh is blazingly fast.

 

[kruk]

It will surely beat the 1080 but just by enough.

Specs wise it's 2 x 470 so it probably will be 1080 level even if they wouldn't do any optimizations driver or hardware wise. If that is true, they only need ~20% more to catch up to 1080 Ti at 1080p - I don't think that is a lot. Custom 1080 Ti's and 4K might be another story though ...

Anyway, WhyCry from VCZ just dug up another device in Sisoftware DB with ID 687F:C3.

 

[the54thvoid]

Specs wise it's 2 x 470 so it probably will be 1080 level even if they wouldn't do any optimizations driver or hardware wise. If that is true, they only need ~20% more to catch up to 1080 Ti at 1080p - I don't think that is a lot. Custom 1080 Ti's and 4K might be another story though ...

Anyway, WhyCry from VCZ just dug up another device in Sisoftware DB with ID 687F:C3.

Better not use a Ryzen CPU if that 1080p level is to mean anything. We're already seeing 7700k being the bottleneck to a 1080ti at 1440p. Can't recall which review but it was mentioned out there.
On that point, we saw truck loads of Ryzen benches, from AMD themselves, a month or so before release. That we don't see a sniff of Vega compared to an Nvidia card directly could be telling. With 1080ti out now, in stores and sold out, with more coming, AMD really ought to show fence sitters like me to hold off for Vega. If they don't, it's a hint it's not as fast.

 

[Kanan]

Vega isn't directly comparable to Polaris because it's a new architecture with brand new features like Primitive Shaders etc.

Also the card "isn't out" that's just a pre release model and on top only a rumour. It's a bit fast to say how fast the card will be based on that.

We will have to wait, but as I said it will be most likely a tad under 1080 ti performance wise and will have futuristic features that will set it apart from Pascal on the long run. Whoever wants a gpu for the long run would better be off buying Vega.

I also don't see why it should be named RX 590, it's fast enough to be comparable to 1080 Ti and Titan. I think they reserve RX 590 for a dual Polaris chip, but that's just a guess.

 

[EarthDog]

Lol, it's all speculation. We have no idea.

MY guess is it will fall between 1080 and 1080ti performance wise with it being closer to (guessing 10% behind) the ti. As far as future things and what is better in the long run, seems incredibly premature to even speculate...

 

[Captain_Tom]

The optimism is strong for some, which is perfectly fine. I'm not waiting too long before buying my next card but what some are missing here are the signs that give Vega away.
Also, some are missing just how far ahead of the 980ti the Pascal version is. I have been saying for ages it needs to match 1080ti in performance but now that card is out, I don't think it will.
Much like Ryzen, i think Vega will be a far better chip with efficiency. That's the driver for them. I think all the time spent from reveal (Doom gaming demo) to release isn't about clocking higher but clocking more efficiently. It will surely beat the 1080 but just by enough.
And as reassuring to some in team red that think GP102 is a laughable Maxwell refresh, if I were AMD I'd be depressed. That simple refresh is blazingly fast.

Actually the Ryzen parallel is funny because in reality Ryzen is only like 10% weaker at gaming, and better than Intel at everything else. Not to mention it is being held back by a couple of Windows 10 HT and cache scheduling bugs. So yeah if Vega is like Ryzen, it will skullcrush Pascal.

 

[Captain_Tom]

Lol, it's all speculation. We have no idea.

MY guess is it will fall between 1080 and 1080ti performance wise with it being closer to (guessing 10% behind) the ti. As far as future things and what is better in the long run, seems incredibly premature to even speculate...

Again, if that's the case - Vega is the biggest screw up in AMD's history. Look at the specs lol. It would mean negative IPC compared to Polaris.

 

[EarthDog]

Edit your posts to add man..

And.. as far as specs/performamce goes: https://www.google.com/amp/s/www.pcgamesn.com/amd/amd-vega-gpu-specifications?amp

 

[dalekdukesboy]

Better not use a Ryzen CPU if that 1080p level is to mean anything. We're already seeing 7700k being the bottleneck to a 1080ti at 1440p. Can't recall which review but it was mentioned out there.
On that point, we saw truck loads of Ryzen benches, from AMD themselves, a month or so before release. That we don't see a sniff of Vega compared to an Nvidia card directly could be telling. With 1080ti out now, in stores and sold out, with more coming, AMD really ought to show fence sitters like me to hold off for Vega. If they don't, it's a hint it's not as fast.

Maybe this is true, but maybe not on why compared vs not compared to Nvidia...

Vega isn't directly comparable to Polaris because it's a new architecture with brand new features like Primitive Shaders etc.

Also the card "isn't out" that's just a pre release model and on top only a rumour. It's a bit fast to say how fast the card will be based on that.

We will have to wait, but as I said it will be most likely a tad under 1080 ti performance wise and will have futuristic features that will set it apart from Pascal on the long run. Whoever wants a gpu for the long run would better be off buying Vega.

I also don't see why it should be named RX 590, it's fast enough to be comparable to 1080 Ti and Titan. I think they reserve RX 590 for a dual Polaris chip, but that's just a guess.

No kidding...people are practically counting their Vega chickens before the egg is laid never mind hatched. Until we get a review, we have an idea of what it can do, but far too early to say it "will be" a 1080, or 1080ti, or somewhere inbetween. Yeah, one or any of those are possibly right or close depending on clocks and drivers etc but that's the point, too much variance of likely possibilities right now and very little hard proof of actual performance.

 

[FordGT90Concept]

And as reassuring to some in team red that think GP102 is a laughable Maxwell refresh, if I were AMD I'd be depressed. That simple refresh is blazingly fast.

Because TSMC's 16nm process is actually good. A lot of the surprise with Pascal came from the obscenely high clockspeeds. GloFo's 14nm process has improved since making RX 480 which is why they can now re-release and rebrand them to RX 580. It also likely means Vega will be better too. I think 1200 MHz is reasonable for Vega, 1400 MHz for Polaris.

 

[Captain_Tom]

Because TSMC's 16nm process is actually good. A lot of the surprise with Pascal came from the obscenely high clockspeeds. GloFo's 14nm process has improved since making RX 480 which is why they can now re-release and rebrand them to RX 580. It also likely means Vega will be better too. I think 1200 MHz is reasonable for Vega, 1400 MHz for Polaris.

1200MHz would be pathetic. 1500 MHz+ or it is a failure.

 

[FordGT90Concept]

No way can GloFo hit that with such a large chip. I will be shocked if it reaches 1300 MHz.

 

[Captain_Tom]

No way can GloFo hit that with such a large chip. I will be shocked if it reaches 1300 MHz.

Buddy we already know that it's 12.5 TFLOPS. So I guess I should clarify that 12.5 TF or GTFO. If it takes 5000 SP's to get there, who cares?!


But I think it is more likely that it is 4096 @ 1500....

 

[FordGT90Concept]

12.5 TFLOP doing FP32 compute work (e.g. deep learning), not GPU work. For comparison, Fury X is allegedly 8.6 TFLOP in the same situation.

 

[kruk]

We're already seeing 7700k being the bottleneck to a 1080ti at 1440p.

Great catch. I thought it was another Fury like problem, but now it makes sense.

After reading the comments, analyzing some slides (see one of them below) and thoroughly thinking about the situation - here are some of my thoughts:
- single Vega 10 will be kept at clocks high enough to beat the 1080 and low enough to have great efficiency (max 1200 MHz boost). Pushing it to 1500 or more would probably consume a lot of power, tanking it's efficiency
- by putting only 8 GB of HBM2 on the card, they can keep the cost lower than I initially thought, and price the card between 1080 and the Ti.
- 1080 Ti and Volta will be attacked by dual Vega 10 chip instead. Does anyone think it's possible to glue two Vega 10 chips together like they did with Ryzen and have better performance than Crossfire?

 

[Captain_Tom]

12.5 TFLOP doing FP32 compute work (e.g. deep learning), not GPU work. For comparison, Fury X is allegedly 8.6 TFLOP in the same situation.

Yes I am well aware of that. The 480 is also 5.8 TF btw, and the Titan X is around 200% the performance of a 480 . 12.5/5.8 = 215% performance of a 480...

 

[Kanan]

No kidding...people are practically counting their Vega chickens before the egg is laid never mind hatched. Until we get a review, we have an idea of what it can do, but far too early to say it "will be" a 1080, or 1080ti, or somewhere inbetween. Yeah, one or any of those are possibly right or close depending on clocks and drivers etc but that's the point, too much variance of likely possibilities right now and very little hard proof of actual performance.

What I did was a educated guess, nothing more or less, based on facts that happened through the years. AMD is always futuristic with its features, so it's not a wild guess that it will have the more future oriented features compared to Pascal and could be very well better on the long run (let's say 1-2 years or more). It's also not a wild guess to say its performance will most likely be under 1080 Ti, seeing it has "only" 4096 shaders most likely not clocked high enough to beat GP102, the typical problem AMD has since years as well: low clocks.

People here are fast to say what I say is just a guess. It's easily more than just that, and everyone can happily call me later when I'm probably right. I was right with Ryzen too. That all said I'd happily buy a Vega instead of a 1080 Ti, as I'm more forward thinking and not buying GPUs for 1-2 years, rather longer than that. I'm always into supporting the more forward thinking companys, that tank less on the same shit to make the most $$ out of it. Nvidia is Intel 2.0, why in hell should I like that. Since Pascal Nvidia has grown into something that's not really good for the end-user. Now AMD is coming with Vega and suddenly they stopped their "Founder's Edition" crap with big price premiums for mediocre coolers that can't even shut down their fans, suddenly the shit is priced for 699 for the big chip, without any big price premiums.

 

[FordGT90Concept]

- 1080 Ti and Volta will be attacked by dual Vega 10 chip instead. Does anyone think it's possible to glue two Vega 10 chips together like they did with Ryzen and have better performance than Crossfire?

Radeon Pro Duo 2, anyone? If you look at the power consumption, those GPUs have low clocks. It definitely isn't an R9 295 X2. These are the cards that virtualization and deep learning companies want and that they were talking about at the capsaicin event. It's also the card they would have used to render the VR video if they had one at the same event.

Yes I am well aware of that. The 480 is also 5.8 TF btw, and the Titan X is around 200% the performance of a 480 . 12.5/5.8 = 215% performance of a 480...

Could suggest similar performance then in GPU loads. Explains why NVIDIA introduced the Ti at a lower than expected price.

 

[Captain_Tom]

What I did was a educated guess, nothing more or less, based on facts that happened through the years. AMD is always futuristic with its features, so it's not a wild guess that it will have the more future oriented features compared to Pascal and could be very well better on the long run (let's say 1-2 years or more). It's also not a wild guess to say its performance will most likely be under 1080 Ti, seeing it has "only" 4096 shaders most likely not clocked high enough to beat GP102, the typical problem AMD has since years as well: low clocks.

Frankly I think the most correct part of what you said is that "We don't know what Vega will be". Seriously that's the truth.


I am inclined to say though that at the very least it will be able to touch the 1080 Ti (Like the Fury X easily "touched" the 980 Ti). If it couldn't at least do that it would mean negative IPC compared to Polaris lol.


More likely I do think it will match or beat it though.

 

[medi01]

So, shader count ("cuda cores") / chip size in mm2
1060 - 1280 / 200
1070 - 1920 / 314 (harvested 1080)
1080 - 2560 / 314
1080Ti - 3584 /471 (harvested Titan, although, number of shaders is the same) => 2.8 times 1060 shaders,, 2.35 times 1060 chip size

RX 480 - 2304 / 232
RX Vega - 4096 / 500 (?) => 1.8 times 480 shaders, 2.15 times 480 chip size

So:
1) math is simply... a tad weird, when scaling up their chips, nvidia/AMD focus on different things
2) Vega is simply too big a chip with superior memory to simply beat 1080
3) There must be a reason why nVidia dropped pricing so drastically, with Volta being at least 6 month away

 

[Captain_Tom]

So, shader count ("cuda cores") / chip size in mm2
1060 - 1280 / 200
1070 - 1920 / 314 (harvested 1080)
1080 - 2560 / 314
1080Ti - 3584 /471 (harvested Titan, although, number of shaders is the same) => 2.8 times 1060 shaders,, 2.35 times 1060 chip size

RX 480 - 2304 / 232
RX Vega - 4096 / 500 (?) => 1.8 times 480 shaders, 2.15 times 480 chip size

So:
1) math is simply... a tad weird, when scaling up their chips, nvidia/AMD focus on different things
2) Vega is simply too big a chip with superior memory to simply beat 1080
3) There must be a reason why nVidia dropped pricing so drastically, with Volta being at least 6 month away

Also consider that the 480 made real progress on how efficient AMD is with TFLOPS/FPS. After all the 480 tends to perform as well or a bit better than the 390X while having a less TFLOPS.

Considering Vega is a SUBSTANTIALLY bigger upgrade to the architecture compared to Polaris, I have no reason to not think AMD could close the TFLOP/FPS gap with Nvidia. In fact the bigger die size per SP hints this is what AMD is in fact trying to do.

 

[Kanan]

It could also be that Vega is produced on a superior node at Samsung or TSMC, which would possibly increase the clocks. /speculation

 

[Captain_Tom]

It could also be that Vega is produced on a superior node at Samsung or TSMC, which would possibly increase the clocks. /speculation

This is could be true, but it wouldn't explain why AMD's Die-size / SP is so much larger than before. Spacing out SP's almost always leads to higher sustainable clocks (Less density = less concentrated heat).