Vega/Navi Rumors (Updated)

[Valantar]

*snip*

Aware of those leaks/rumors, yep. That still doesn't make them any more than leaks/rumors. We had leaks/rumors about Polaris easily clocking past 1600MHz just a month or two before release. That sure didn't happen. On the other hand, others were right. I prefer taking it all with a pinch of salt.

Oh, and Videocardz is barely more credible than WCCFtech. Barely.

I don't doubt that the biggest Vega will have 64CUs. It would be very weird for them to have fewer CUs than Fiji. Clock speeds? Anyone's guess. IPC improvements? The same. If the demos we've seen matching/beating the 1080 are ES samples running at 1000-1200MHz, and final clocks end up around 1500? That would be fantastic, especially if they peg it at 225W total board power. That'd be a win, for sure. Do I feel certain that that's where they'll end up? No. We simply don't know. Do I hope they do? Heck yes. But we don't know. Which is what @w3rd was saying.

 

[french toast]

The reality that you are ignoring is that what matters in the end is actual performance. Sure, one could fairly argue that Vega is more "Advanced" than Pascal, simply because it does things differently, simply because it has a greater theoretical TFlop advantage, but in the end--none of that matters if those super cool internal silicon whatsitgizmos are grossly underutilized by the intended software, and thus represents a real performance loss on the table. What good is 12.3TFlop potential, if the end result from application to application is only something in the neighborhood of 10.5 TFlop? This is the ongoing story with AMD, and you do yourself a credibility disservice by ignoring this historical truth.

Yes, I would like to see AMD come out of nowhere and show with a Vega 10 @, say, $500-600 that actually outperforms the latest and greatest 1080Ti from nVidia, but that is pure speculation to assume this happens, and is essentially based on ignoring the fundamental history of AMD's design with relation to actual developers. It is the same story over and over again, and I am not sure why you keep buying it.

I will say this: with the surprising pricing of 1080Ti, I do think nVidia has a pretty good idea of what Vega will bring, and I do see this as acknowledgement that Vega is probably a bit more legitimate than Fury and Fury X--so I see this as good news. I wouldn't hold out hope, though. All we actually know about Vega is that it has a 0-5% advantage over a 1 year-old card that, itself, now shows a ~20-30% disadvantage compared to its successor which is actually "available" now. Sure, you can say that those early Vega demos were unoptimized with crap drivers and crap systems...but does that represent a recoverable 20-25% gain in performance? I mean...I'd like for that to be the case, but my the rational side simply laughs at the proposition.

In the end, you have to look at price/performance. I'm going with the best-case scenario of Vega 10 coming it at a 10-15% disadvantage to 1080Ti @ 20-25% less cost. This leaves the expected potential for yet another "AMD ages much better" scenario where, at some point, that Vega will come very close to a 1080Ti at whatever future time, if not matching it, with a lower up-front cost. I feel that leaves enough room for a pleasant surprise, and far less disappointment.

This is EXACTLY what i have said/predicted, 15% slower, a bit less power 225w? And 25% cheaper. = fantastic product.

 

[french toast]

Aware of those leaks/rumors, yep. That still doesn't make them any more than leaks/rumors. We had leaks/rumors about Polaris easily clocking past 1600MHz just a month or two before release. That sure didn't happen. On the other hand, others were right. I prefer taking it all with a pinch of salt.

Oh, and Videocardz is barely more credible than WCCFtech. Barely.

I don't doubt that the biggest Vega will have 64CUs. It would be very weird for them to have fewer CUs than Fiji. Clock speeds? Anyone's guess. IPC improvements? The same. If the demos we've seen matching/beating the 1080 are ES samples running at 1000-1200MHz, and final clocks end up around 1500? That would be fantastic, especially if they peg it at 225W total board power. That'd be a win, for sure. Do I feel certain that that's where they'll end up? No. We simply don't know. Do I hope they do? Heck yes. But we don't know. Which is what @w3rd was saying.

I believe vega 10 (largest die) will be <500mm2, yes 64 compute units, optimised for higher clocks and potentially drastically higher ipc in optimal well developed games.
There will be no bigger dies or even the same size from amd ever again.
Vega 11 is most likely a dual vega 10, IF it even exists, there is no space for a smaller vega in the lineup, when you consider cut down vega 10 and rx 580 rebrand.

 

[piesquared]

Aware of those leaks/rumors, yep. That still doesn't make them any more than leaks/rumors. We had leaks/rumors about Polaris easily clocking past 1600MHz just a month or two before release. That sure didn't happen. On the other hand, others were right. I prefer taking it all with a pinch of salt.

Oh, and Videocardz is barely more credible than WCCFtech. Barely.

I don't doubt that the biggest Vega will have 64CUs. It would be very weird for them to have fewer CUs than Fiji. Clock speeds? Anyone's guess. IPC improvements? The same. If the demos we've seen matching/beating the 1080 are ES samples running at 1000-1200MHz, and final clocks end up around 1500? That would be fantastic, especially if they peg it at 225W total board power. That'd be a win, for sure. Do I feel certain that that's where they'll end up? No. We simply don't know. Do I hope they do? Heck yes. But we don't know. Which is what @w3rd was saying.

AMD's architectures have much higher IPC than Nvidia's for a long time now. Of course, we are told that IPC doesn't matter for graphics cards, since around the time of Keplers release.

 

[Headfoot]

I don't believe that Vega will hit 1500+ stock when Polaris couldn't get close to that, unless they really didnt hardly put any investment in polaris and banked 100% on Vega. I just dont see that much clock speed uplift happening this late in the game, since Vega has had to be in the pipe for quite some time now.

 

[Glo.]

This is EXACTLY what i have said/predicted, 15% slower, a bit less power 225w? And 25% cheaper. = fantastic product.

Well it will actually depend on clock speed of the GPU itself, however, when you consider how powerful culling technique AMD has implemented in the hardware of Vega, expect it to be very, very long living architecture. And on this front w3rd is correct, but not exactly based on faith, but based on simply technical analysis.

I will not speculate on the hardware performance, tho, even if I have my suspicions. Or educated guesses.

But let me put about Vega one thing here. AMD, is making Vega completely separate brand because they want to reestablish themselves in the minds of people. Intel already has extreme competition on the CPU front from AMD. It will be interesting in upcoming months to see how GPU market will be.

What I would want is powerful and efficient APU, tho...

 

[Det0x]

I don't believe that Vega will hit 1500+ stock when Polaris couldn't get close to that, unless they really didnt hardly put any investment in polaris and banked 100% on Vega. I just dont see that much clock speed uplift happening this late in the game, since Vega has had to be in the pipe for quite some time now.

Far more intriguing is MI25. Notice how these cards' names match up reasonably well with their TFLOPS throughput. If that naming convention holds true, the Vega-powered MI25 offers a staggering 25TFLOPS of compute. Yet before you salivate and think of what this could mean for gaming in, perhaps, RX 490 form, we can tell you that such a figure is based on half-precision throughput (16-bit). Vega, as you will learn, offers more mixed-precision ability than ever before, and specific to machine learning, enhanced floating-point accuracy is not always required.

http://hexus.net/tech/news/graphics/100255-amd-releases-radeon-instinct-gpus-machine-learning/

https://videocardz.com/65521/amd-vega-10-and-vega-20-slides-revealed

12TFLOPs
64 Next Generation Compute Unit (NCU)

FLOPs = Compute Units/Core Clusters x operations per clock (per CU) x GPU clock speeds
or
FLOPs = GPU Stream Processors x operations per clock (per core) x GPU clock speeds

Either the slide is fake, or consumer Vega10 will run atleast 1500mhz
And a few other quotes:

The improved and more efficient NCU also means that Vega cards can run at higher clock speeds – so that they can motor through their tasks at a faster pace. That's one of the key attributes in processing performance. Expect top-end Vega cards to run at around 1,500MHz or 1,600MHz – about level with the best Nvidia Pascal chips, and a huge improvement over the last generation, which topped out at around 1,000MHz.

http://www.trustedreviews.com/news/amd-vega-specs-performance-release-date-technology-explained

The MI25 professional card can deliver 12 teraflops of single precision computing. That would put it ahead of the Nvidia GTX Titan X which bats around the 11 teraflop mark, so it’s not unreasonably to think that could end up being a new AMD RX Fury.
.
.
The big-boy Vega 10 chip, the one that’s meant to be based on the Radeon Instinct professional MI25 card, could potentially hit 12.5 teraflops of single precision processing.

https://www.pcgamesn.com/amd/amd-vega-gpu-specifications

Though we can take a good guess of the Vega GPU's shader and speed composition by looking at the vital stats of the MI25 Instinct accelerator - which would provide 12.5 TFLOPS in standard single-precision performance - AMD has yet to divulge the finer details.

http://hexus.net/tech/news/graphics/101056-amd-vega-gpu-architecture-uncovered/?page=3

I could post more sites, but i think it wouldn't make any difference at this point.

 

[tential]

So are we guessing end of may early june?

Hmm... I wonder what investment I want to gamble on in the meantime with my Vega money.

I just don't believe OCed Vega will be within 10% of OCed GTX 1080Ti.

But if it is..... my wallet is ready.

Best part of AMD is CF at least scales.....
Those latest benches reminded me why I didn't want a 980Ti. I couldn't get 2 -.-

 

[Bacon1]

So are we guessing end of may early june?

Early May most likely, they are advertising it with Prey which launches May 5th.

 

[Headfoot]

images

I'll believe it when I see it. I suspect there is something more going on here, that the CU's themselves are different in a way that lends more FLOPS per clock. CU divided by clock assumes the CUs are the same as in previous versions of GCN. I just find it highly unlikely they'll go from Polaris clockspeeds to 1500mhz+ on architectures that are closely related in time, architecture and process.

 

[JDG1980]

Clock speeds? Anyone's guess. IPC improvements? The same.

AMD's architecture slides for Vega specifically mention that it is "optimized for higher clock speeds and higher IPC". It would make no sense to advertise higher clocks and then release at 1200 MHz, which is lower than the clocks on Polaris (RX 480 has a default of 1266 MHz). As for IPC, we know from official data that AMD is implementing tiled rendering ("Draw Stream Binning Rasterizer"), which was a major part of Maxwell's IPC gains over Kepler. The only question is how much.

 

[JDG1980]

I don't believe that Vega will hit 1500+ stock when Polaris couldn't get close to that, unless they really didnt hardly put any investment in polaris and banked 100% on Vega.

Vega is a far bigger architectural change than Polaris. Think of it as Polaris is "GCN 1.3" and Vega is "GCN 2.0".

And Polaris pretty much was a pipe-cleaner product with minimal investment. The reason why it was hyped as a major release is that it was pretty much all AMD had for 2016. It was meant to kick off the GloFo 14nm process and work out its bugs, while also providing something that would be at least moderately competitive on the low to mid range and in thin-and-light laptops.

 

[Valantar]

AMD's architectures have much higher IPC than Nvidia's for a long time now. Of course, we are told that IPC doesn't matter for graphics cards, since around the time of Keplers release.

Really? Remember, IPC is performance per clock per core. AMD has significantly more cores than comparable Nvidia models, even if clocks are lower. As an example: if a 1300MHz RX 480 and a 1709MHz GTX 1060 both perform 1 000 000 in some benchmark, that's .457 (1 000 000 / 1280 / 1709 = .457139) points per core per clock for the Nvidia card, and .33 (1 000 000 / 2304 / 1300 = .333868) for the AMD. Not that this is accurate (most 1060s boost past 1709, and they rarely score identically), but it's in the ballpark. Then again, AMD is able to have significantly more cores in the same die size and power budget.

AMD is implementing tiled rendering ("Draw Stream Binning Rasterizer"), which was a major part of Maxwell's IPC gains over Kepler. The only question is how much.

My understanding is that tiled rendering is a large part of Maxwell's efficiency gains over Kepler, but that IPC stayed roughly the same (and that performance increased due to increased core counts and significant clock boosts).

 

[Valantar]

AMD's architecture slides for Vega specifically mention that it is "optimized for higher clock speeds and higher IPC". It would make no sense to advertise higher clocks and then release at 1200 MHz, which is lower than the clocks on Polaris (RX 480 has a default of 1266 MHz). As for IPC, we know from official data that AMD is implementing tiled rendering ("Draw Stream Binning Rasterizer"), which was a major part of Maxwell's IPC gains over Kepler. The only question is how much.

Yes, and? That doesn't tell us anything other than "they should be better". That's what we're all assuming. Higher clock speeds could be 1300 base, 1500 base, 1700 base, whatever. IPC gains could be 5% or 50%. We don't know. Keeping a critical distance to this is essential if we're to avoid screaming fanboyism and actually discuss something interesting.

 

[Mopetar]

I just find it highly unlikely they'll go from Polaris clockspeeds to 1500mhz+ on architectures that are closely related in time, architecture and process.

They're somewhat separate I think. Polaris was an Optimization of GCN 3 (or GCN 1.2 if you use that system), kind of like how Pascal isn't a truly new design either. Vega does have a lot more architectural changes than Polaris did. Polaris was the more conservative design to get the 14 nm process at GF worked out.

I don't know if the architecture itself clocks better, though I expect its something AMD will have worked on since the previous versions of GCN weren't the best at scaling to higher speeds. However, I think that one reason that Vega is so late is that AMD realized that the GF LPP process wouldn't let them hit the higher clock speeds (this should be blatantly obvious after looking at how quickly Ryzen OC falls off. Also Apple made their A9 using both TSMC's 16 nm and Samsung's 14 nm process, but the A9X which was clocked higher was only made at TSMC which adds further support to that notion) necessary so they shifted to using Samsung's newer LPU, which is supposed to be targeted for high performance. My guess is wwe get both an architecture and fab process that will enable Vega to hit higher clock speeds.

 

[JDG1980]

Will vega be better than Polaris? I don't doubt that for a second. But will it perform more than 30% better for the same power? That would be nigh on miraculous.

GTX 980 (Maxwell) had the same ~180W TDP as GTX 680 (Kepler), on the same 28nm TSMC process node. It was over 56% faster at 1080p gaming. (source) This demonstrates how much of a difference architectural improvements can make.

An increase of 30% perf/watt for Vega over Polaris would not be at all a stretch, especially since one of its known improvements is the same thing that largely contributed to Maxwell's performance (tiled rendering). In fact, if AMD didn't do more than 30% perf/watt improvements for what is being touted as the biggest architectural change yet to GCN, I would consider it a disappointment.

 

[w3rd]

GTX 980 (Maxwell) had the same ~180W TDP as GTX 680 (Kepler), on the same 28nm TSMC process node. It was over 56% faster at 1080p gaming. (source) This demonstrates how much of a difference architectural improvements can make.

An increase of 30% perf/watt for Vega over Polaris would not be at all a stretch, especially since one of its known improvements is the same thing that largely contributed to Maxwell's performance (tiled rendering). In fact, if AMD didn't do more than 30% perf/watt improvements for what is being touted as the biggest architectural change yet to GCN, I would consider it a disappointment.

I don't think you have to explain things to that guy. He is completely ignorant of technology & His line of reasoning is the same in every post..
AMD sucks and can't do anything right. Nvidia Yo...

Too many people have had to explain things to him.


Ironically, he doesn't even know that Polaris isn't Vega & to him, they are the same uarch. Nvidia is the extent of Valantar's knowledge.

 

[zinfamous]

I don't think you have to explain things to that guy. He is completely ignorant of technology & His line of reasoning is the same in every post..
AMD sucks and can't do anything right. Nvidia Yo...

Too many people have had to explain things to him.


Ironically, he doesn't even know that Polaris isn't Vega & to him, they are the same uarch. Nvidia is the extent of Valantar's knowledge.

He said none of these things and it is clear that you are blind to everyone's comments. You are a pure troll, and it astonishes me that you keep posting this nonsense.

 

[tamz_msc]

You cannot estimate gaming performance from an architecture preview. Did AMD give any performance numbers? Where is the equivalent of +40% IPC that was targeted with Zen in the case of Vega?

Until then, it's all speculation.

Leaked compubench scores don't count.

 

[Glo.]

My understanding is that tiled rendering is a large part of Maxwell's efficiency gains over Kepler, but that IPC stayed roughly the same (and that performance increased due to increased core counts and significant clock boosts).

How come then GTX 980 was faster than GTX 780 Ti, despite having less cores, and lower TFLOPs output?

Differences between Kepler and Maxwell. 128 Cores in each SM, vs 192. 4 times bigger cache size, and the size of registry files, available to each SM.

What this means is that lower amount of cores in Maxwell was doing the same job as 192 cores in Kepler. That is the reason why Maxwell outperformed Kepler both in Gaming and Compute scenarios. Tile Based Rasterization brings in the first place efficiency. It does not increase the throughput essentially but allows for very efficient culling techniques, to be effective. Nvidia basically have said that those 128 Maxwell cores have 90% of performance of 192 Kepler cores.

Vega has everything: HBM2 - efficiency, TBR - Efficiency, Increased Geometry throughput, next generation scheduling, CUs that are capable of doing more stuff, and larger registry files(TBC), extremely enhanced Culling...

Like I have previously said. Vega will gain performance with time. Games will have to be updated to use Primitive Shaders, so that is postponed into sometime in the future. But let me give you an example. Lets say, we have 1024 GCN Polaris GPU, clocked at 1200 MHz, and we have 1024 GCN core Vega GPU clocked at 1200 MHz. Vega will be in for example Overwatch 25% faster than Polaris GPU, and in game like Doom, it will be up to 75% faster, without using Primitive Shaders. Overall gains can be much higher than this.

 

[Valantar]

But let me give you an example. Lets say, we have 1024 GCN Polaris GPU, clocked at 1200 MHz, and we have 1024 GCN core Vega GPU clocked at 1200 MHz. Vega will be in for example Overwatch 25% faster than Polaris GPU, and in game like Doom, it will be up to 75% faster, without using Primitive Shaders. Overall gains can be much higher than this.

(emphasis added)
What do you base this on? Do you have some sort of inside knowledge that you're not allowed to share? Because otherwise, you're just pulling numbers out of your a**, and expressing it as fact. Remember, word choice matters. What you think, what you know, what might happen and what will happen are all different things.

Now, I'm not saying that's impossible. I'm not even saying it's unlikely. I'm just saying that you're talking like this is absolute certainty - which it isn't. There can't be until we actually have independent testing verifying this.

Again: Will Vega be faster than Polaris? Of course it will. AMD engineers aren't incompetent. But how much faster, and in what ways? We don't know. We can't know. And that's okay. AMD keeps giving us little tidbits, and if Ryzen is anything to go by, they're reasonably accurate. Until we know more, I'll be cautiously optimistic. But do I go around spouting speculation that the fastest Vega will beat the 1080Ti at 225W? No. Because not only would that be utterly baseless speculation, it would be building up stupidly high expectations - which, should they not be met, would cause a PR sh*tstorm against AMD. Let's keep hopes high (heck, I want AMD to kick Nvidia's a** again, it's about time!), but our expectations as realistic as possible. Anything else is unwise in many, many ways.

Edit: spelling, grammar, that kind of stuff. My oh my.

 

[JDG1980]

You cannot estimate gaming performance from an architecture preview. Did AMD give any performance numbers? Where is the equivalent of +40% IPC that was targeted with Zen in the case of Vega?

Until then, it's all speculation.

This whole thread is devoted to speculation. We can make reasonable inferences from AMD's slides and public statements, and from the gains that Nvidia obtained by implementing similar technology.

AMD told Anandtech that the MI25 Vega accelerator is going to have ~12.5 TFlops of single-precision performance. We know, on every GPU ever released in the modern era from both AMD and Nvidia, the formula for calculating TFlops. It's a simple function of (shaders x MHz x 2) / 1,000,000. Thus, if the flagship Vega chip has 4096 shaders (and all the information we've seen from multiple leaks so far indicate it does), then it has to run at ~1525 MHz to hit the TFlop goal. We also know based on past experience that professional cards almost never run higher clocks than consumer-grade cards; if anything, they run lower. From all this we can reasonably infer that the flagship consumer Vega card will run no slower than 1500 MHz, and quite possibly faster.

As for "IPC" - as used here, most people really mean something like "DX11 gaming performance per TFlop". Whether that is what AMD means in its architectural slides is not clear. And you're right that they didn't give specific numbers. However, again, we can make inferences from what we do know. AMD is implementing tiled rendering, so how much will that gain them? The closest comparison I can find is Nvidia's GTX 650 Ti (Kepler) to GTX 750 Ti (1st-generation Maxwell). Delta color compression (which both AMD and Nvidia now have), was, as far as I know, introduced only with 2nd-generation Maxwell (900 series), so using 1st-generation Maxwell as a baseline should mostly isolate the gains from tiled rendering. On 1080p gaming, GTX 750 Ti did about 22% better than GTX 650 Ti. This was despite the fact that the Maxwell card had only about 91.5% the raw TFlops of the Kepler card. In other words, Maxwell 1st generation had about 33% better DX11 gaming performance per TFlop than Kepler. (1.22 / 0.915) We can therefore reasonably infer that tiled rendering should provide significant gains for AMD in Vega, likely over 30% better DX11 gaming performance per TFlop. This is assuming no other significant bottlenecks - and we know from the slides that Vega does remove at least one major bottleneck, the 4-shader-engine limitation, which was one of the primary reasons that Fiji was so underwhelming.

These, then, are my predictions based on what we do know: clock speed for the flagship card of no lower than 1500 MHz (except perhaps in Furmark and other extreme TDP-constrained scenarios), and improvements in DX11 gaming performance per TFlop of no less than 30% over Polaris. We will see in a couple months whether I am right or wrong.

 

[tamz_msc]

I don't think 12.5 TFLOPs of peak throughput was ever hinted at being available in consumer Vega. People make that comparison because the Titan X Pascal is 11 TFLOPs.

 

[Det0x]

I don't think 12.5 TFLOPs of peak throughput was ever hinted at being available in consumer Vega.

I would recommend tightening your glasses

 

[Krteq]

I don't think 12.5 TFLOPs of peak throughput was ever hinted at being available in consumer Vega. People make that comparison because the Titan X Pascal is 11 TFLOPs.

It's FP32 theoretical peak throughput, you can't cut it artificially in any way to differentiate products.