[Rumor, Tweaktown] AMD to launch next-gen Navi graphics cards at E3

[Stuka87]

I'm just happy that it'll be the last GCN architecture. And since it also goes in consoles, it'll be more gaming oriented, unlike Vega.

The design of the chip doesn't really have anything to do with it being "GCN". GCN is the instruction set, not the chip design. Its like saying AMD should have dumped x86 because Bulldozer sucked.

 

[ozzy702]

Nvidia's naming each and every tweak of their architecture like some completely new and radically different thing, has clouded a lot of ppl's view on gpus.

No, it's just that architecture improvements execution are visible on NVIDIA's side and ultimately what that means for performance, not so much with AMD. Hopefully that changes with NAVI.

 

[railven]

When Vega came out to disappoint, it should not surprise anyone that is pragmatic. This forum in particularly was hit hard by the hype train crash because of the generally pro AMD attitude and the inability to notice the warning signs that Vega was going to disappoint(even super pro AMD Adoredtv notice this).

Gonna have to agree with this part. I was still strong for AMD during Vega. But this forum took new posters with "inside scoops" as messiahs. Suddenly Vega was going to destroy NV, we just needed that Magic Driver. Fine Wine, you should all remember this. Even the lead up to Vega launch, the "can you tell a difference" campaign - it was terrible. And when that hype train crashed these new posters who "I worked with Vega, it's amazing" disappeared.

Why I started to watch AdoredTV. Then he lost his angle of being direct and honest and turned used car salesman. I hope AMD can deliver, more so because the market is terrible right now, but I won't buy a ticket for this train, I'm on the Intel train now! (And I already feel it's gonna derail at the first bend haha).

 

[tajoh111]

Nvidia's naming each and every tweak of their architecture like some completely new and radically different thing, has clouded a lot of ppl's view on gpus.

People are not really mistaken.

People are aware when Nvidia does a tick and a tock.

Nvidia's bigger redesigns such as those brought with Kepler, Maxwell and Turing show bigger gains in performance per tflop(the performance in game vs a tflop in computational power).

While Kepler bought a regression, Maxwell and turing brought them up significantly. More so than any single transition from AMD.That is at the same tflop, Nvidia gains are stronger than AMD which is why people keep on harping on AMD to change their architecture. It is running into bottlenecks which slows down their performance in games which is not representative of their computational power. Vega brought a regression vs Polaris.

This is a comparison that shows performance equalizes the tflops between cards and compares the performance.

https://www.computerbase.de/2016-08/amd-radeon-polaris-architektur-performance/2/

Between 1st gen Gen(7970) and 4th gen(polaris), there was a 18% gain in performance per tflop. Between Tonga and Polaris, there was only a 7% gain in performance per tflop. Between Tahiti and Tonga, there was only a 10% gain. The later of which is not bad but nothing like the gains on Nvidia's big changes that are stated to be architecture changes for nvidia.

Between Kepler and Maxwell, there was around a 25-30% gain in performance per tflop, if we look at 20% tflop advantage of the gtx 780 ti over gtx 980 and the 5-10% performance advantage of the GTX 980 in games(results from techpowerup, which are very similar to computerbase.de).

If we compare the GTX 1080 ti and the GTX 2080, the former has a 15% computational advantage, while the later has a 5-8% performance advantage, which translate into a 23% gain in IPC. Note in all these cases for Nvidia, there was a regression in bandwidth unlike the comparisons in IPC for computerbase. Where the Memory bandwidth was equalized in the computerbase.de comparison for AMD architecture, the GTX 780 ti has a significant 50% bandwidth advantage over the GTX 980 and the GTX 1080 ti 18% bandwidth advantage over the RTX 2080. Equalize the bandwidths and the gains in the Nvidia architectures are even greater.

This is why people want AMD to get off GCN. The larger changes between architectures in Nvidia(most people know maxwell and pascal is pure clocks), which denote a new architecture speak for themselves in terms of performance. When the gains in performance per tflop in a Nvidia architectures changes resemble the entire incremental improvements of GCN 1 to GCN 5 or greater, you can't blame marketing and people to think Nvidia is doing an architectural change while AMD is simply revising GCN.

 

[GodisanAtheist]

Yes, lets please not board the NAVI hype train.

RTX 2070 performance (aka GTX 1080/Vega 64 performance) has been the expectation for a long time and it sounds about right given the release of VII earlier this year at roughly 2080/1080TI performance levels.

As for the Shanghai team, I think they will really prove themselves with their first post-NAVI clean slate design. Improving someone else's design (of anything: hardware, software, a simple business process, etc) can often times be much messier and difficult than just building something of your own from the ground up.

If the post-Navi design is well received, then we know the issues with the last couple launches probably stemmed more from this batch of engineers having to pick-up a product mid-stream and see it to completion. if the successor design is mediocre or sub-par, then that still doesn't mean the design team is necessarily bad, but that is a much more real possibility along side other issues such as cultural issues, language barriers, mis-management, etc... (If hardware is China, software is India, and executive management is America... jesus)

 

[Mopetar]

Did you see Adored’s latest video? It’s pretty anti-hype train. He seems pretty honest to me if he’ll report that the most recent rumors paint Navi in a bad light. Whether he’s actually getting leaks or it’s mainly just someone in marketing feeding him what they want him to hear is another matter, but as I’ve said before he’s at least open about how he comes to his conclusions.

Personally I think he should focus more on actual products since his analysis is quite good, but I can see why he does rumor videos since those are going to get more clicks.

Hopefully AMD releases a product that isn’t pushed to the extremes. Even if Navi isn’t as powerful as initially hoped, it can still be a good product for the right price. I think AMD has more to gain from releasing something more power efficient than trying to eek out every last drop of performance and making another batch of power hungry cards that need to be undervolted.

 

[fleshconsumed]

Yeah, my expectations are rather modest after Vega. I hope AMD will deliver something reasonably fast, reasonably cool, and reasonably priced. If not then I guess I'm keeping my RX480 8GB until post-Navi.

 

[Mockingbird]

Did you see Adored’s latest video? It’s pretty anti-hype train. He seems pretty honest to me if he’ll report that the most recent rumors paint Navi in a bad light. Whether he’s actually getting leaks or it’s mainly just someone in marketing feeding him what they want him to hear is another matter, but as I’ve said before he’s at least open about how he comes to his conclusions.

Personally I think he should focus more on actual products since his analysis is quite good, but I can see why he does rumor videos since those are going to get more clicks.

Hopefully AMD releases a product that isn’t pushed to the extremes. Even if Navi isn’t as powerful as initially hoped, it can still be a good product for the right price. I think AMD has more to gain from releasing something more power efficient than trying to eek out every last drop of performance and making another batch of power hungry cards that need to be undervolted.

...or he's just putting out as many controversial videos as possible to get the most views

 

[Leadbox]

People are not really mistaken.

People are aware when Nvidia does a tick and a tock.

Nvidia's bigger redesigns such as those brought with Kepler, Maxwell and Turing show bigger gains in performance per tflop(the performance in game vs a tflop in computational power).

While Kepler bought a regression, Maxwell and turing brought them up significantly. More so than any single transition from AMD.That is at the same tflop, Nvidia gains are stronger than AMD which is why people keep on harping on AMD to change their architecture. It is running into bottlenecks which slows down their performance in games which is not representative of their computational power. Vega brought a regression vs Polaris.

This is a comparison that shows performance equalizes the tflops between cards and compares the performance.

https://www.computerbase.de/2016-08/amd-radeon-polaris-architektur-performance/2/

Between 1st gen Gen(7970) and 4th gen(polaris), there was a 18% gain in performance per tflop. Between Tonga and Polaris, there was only a 7% gain in performance per tflop. Between Tahiti and Tonga, there was only a 10% gain. The later of which is not bad but nothing like the gains on Nvidia's big changes that are stated to be architecture changes for nvidia.

Between Kepler and Maxwell, there was around a 25-30% gain in performance per tflop, if we look at 20% tflop advantage of the gtx 780 ti over gtx 980 and the 5-10% performance advantage of the GTX 980 in games(results from techpowerup, which are very similar to computerbase.de).

If we compare the GTX 1080 ti and the GTX 2080, the former has a 15% computational advantage, while the later has a 5-8% performance advantage, which translate into a 23% gain in IPC. Note in all these cases for Nvidia, there was a regression in bandwidth unlike the comparisons in IPC for computerbase. Where the Memory bandwidth was equalized in the computerbase.de comparison for AMD architecture, the GTX 780 ti has a significant 50% bandwidth advantage over the GTX 980 and the GTX 1080 ti 18% bandwidth advantage over the RTX 2080. Equalize the bandwidths and the gains in the Nvidia architectures are even greater.

This is why people want AMD to get off GCN. The larger changes between architectures in Nvidia(most people know maxwell and pascal is pure clocks), which denote a new architecture speak for themselves in terms of performance. When the gains in performance per tflop in a Nvidia architectures changes resemble the entire incremental improvements of GCN 1 to GCN 5 or greater, you can't blame marketing and people to think Nvidia is doing an architectural change while AMD is simply revising GCN.

Tflops #s have never been representantive of gaming performance and no, they don't need to get off GCN. The gaming industry just needs to get off dated APIs

 

[Stuka87]

Yeah, my expectations are rather modest after Vega. I hope AMD will deliver something reasonably fast, reasonably cool, and reasonably priced. If not then I guess I'm keeping my RX480 8GB until post-Navi.

Odds are, we will get two of those three things.

 

[prtskg]

The design of the chip doesn't really have anything to do with it being "GCN". GCN is the instruction set, not the chip design. Its like saying AMD should have dumped x86 because Bulldozer sucked.

For me, GCN for AMD's gpu is what Bulldozer was for its cpu.
If Navi is good, I'll be very happy. Probably buy one of the gpus. But I get the feeling that it'll just be good enough. Nothing spectacular, nor bad; like polaris, another gpu for console. That's my expectation. The reason I think that is I don't expect AMD to invest heavily on an architecture which is on its last iteration, if it has to choose between next gen and Navi.

 

[Glo.]

For me, GCN for AMD's gpu is what Bulldozer was for its cpu.
If Navi is good, I'll be very happy. Probably buy one of the gpus. But I get the feeling that it'll just be good enough. Nothing spectacular, nor bad; like polaris, another gpu for console. That's my expectation. The reason I think that is I don't expect AMD to invest heavily on an architecture which is on its last iteration, if it has to choose between next gen and Navi.

You guys do realize that GCN is in latest Exascale Supercomputer GPUs?

GCN IS BLOODY ISA! Its like saying Skylake is last iteration of x86 architecture/ISA.

 

[JDG1980]

Nvidia's naming each and every tweak of their architecture like some completely new and radically different thing, has clouded a lot of ppl's view on gpus.

Yes, Nvidia's architectures since Fermi have been iterative rather than clean-sheet designs. But Kepler->Maxwell alone was a bigger jump than everything AMD has managed with GCN since at least Hawaii, if not Tahiti.

 

[JDG1980]

Tflops #s have never been representantive of gaming performance

Yes, that's the whole point - Nvidia does a better job at converting raw TFlops of computing power into actual gaming performance.

and no, they don't need to get off GCN. The gaming industry just needs to get off dated APIs

I hate to break it to you, but DX11 is here to stay forever. AMD can either accept it, or endlessly remain behind.

 

[JDG1980]

You guys do realize that GCN is in latest Exascale Supercomputer GPUs?

GCN IS BLOODY ISA! Its like saying Skylake is last iteration of x86 architecture/ISA.

And it might be possible to create a good, performant GPU that uses the GCN ISA if it was a clean-sheet design. But trying to fix AMD's *current architecture*, with its many bottlenecks and limitations, is almost certainly a dead end since the current team didn't invent it and doesn't understand it.

 

[Glo.]

And it might be possible to create a good, performant GPU that uses the GCN ISA if it was a clean-sheet design. But trying to fix AMD's *current architecture*, with its many bottlenecks and limitations, is almost certainly a dead end since the current team didn't invent it and doesn't understand it.

No, it is not. The only bottlenecks for AMD GPUs are always clocks, and Memory bandwidth.

If AMD can solve this, and clock their GPUs properly they will have equally good GPUs as Nvidia's. And that is BEFORE even architectural changes to their GPUs. If they could add architectural changes: more Geometry registered/culled, SIMDs doing more work with each cycle, etc the performance increase can be more than just simple nodeshrink/more ALUs approach.

 

[Glo.]

I hate to break it to you, but DX11 is here to stay forever. AMD can either accept it, or endlessly remain behind.

Nobody appears to care. Future of the APIs is Vulkan, and nobody is right now trying to downplay this.

 

[Head1985]

No, it is not. The only bottlenecks for AMD GPUs are always clocks, and Memory bandwidth.

If AMD can solve this, and clock their GPUs properly they will have equally good GPUs as Nvidia's. And that is BEFORE even architectural changes to their GPUs. If they could add architectural changes: more Geometry registered/culled, SIMDs doing more work with each cycle, etc the performance increase can be more than just simple nodeshrink/more ALUs approach.

Radeon 7 clock similar to 1080TI FE around 1700mhz under load and its still slower average despite 1TB bandwidth.
https://www.computerbase.de/thema/grafikkarte/rangliste/#diagramm-performancerating-2560-1440

The bottlenecks with GCN are:
4xShader engines which 1x geometry engine per shader engine.So only 4x geometry engines total.GCN is bad with geometry since beginning.They needed switch to 6x shader engines after hawaii.With 6x engines they will also gain more shader utilization because each SE will have less shaders to take "care".With 6x SE they will gain +50% geometry performance.

4x shader engines with max 16x rops each SE so only 64rops total.Reducing pixel fillrate.Again with 6x SE its +50% pixel fillrate.

And last memory bandwidth bottleneck.AMDs delta color compression is so much worse than NVs its not even funny.Polaris was last time they did something to DCC and it was still worse than old maxwell DCC!!!RX480 have 256GB/s and GTX980 224GB/s

Since then Nv introduces pascall DCC with +20% better compression over maxwell and turing with another +20% over pascall.So i think AMD is behind like for 50% vs turing(polaris is worse than maxwell and vega didnt do anything with DCC)
Radeon 7 with 1TB/s bandwidth is still bottleneck by memory bandwidth(insane)

So for NAVI they need move to 6x Shader engines if they want use more than 2560SP.Also they need better DCC like alot better.Atleast +40% vs polaris.Some tweaks to geometry engines itself so gains can be even better than +50% in geometry(best will be if they can just double geometry engines per SE along with 6x shader engines so from 4x geometry engines to 12x total)..With 12x they will probably be very close to Nvidia geometry performance.
Some cache tuning.

 

[ozzy702]

Since then Nv introduces pascall DCC with +20% better compression over maxwell and turing with another +20% over pascall.So i think AMD is behind like for 50% vs turing(polaris is worse than maxwell and vega didnt do anything with DCC)
Radeon 7 with 1TB/s bandwidth is still bottleneck by memory bandwidth(insane)

Wow... I knew things were bad, but that's impressively bad, especially at only 1440p. A 2080TI would have a mountain of headroom with 1TB/s. That's a huge gap for .1% lows as well. Thanks for posting.

 

[Evleos]

There's 87 job postings with AMD in Markham, 74 in Shanghai. Both locations have a mix of software & hardware engineering, and seem to focus on GPUs and APUs. The rumours about Markham's death within RTG are exaggerated.

https://jobs.amd.com/go/Engineering/2566900/?q=&q2=&alertId=&locationsearch=&title=&location=markham
https://jobs.amd.com/go/Engineering...tId=&locationsearch=&title=&location=shanghai

 

[Evleos]

Going through the LLVM changes the list of HW bugs for GFX10.1 is a lot



https://github.com/llvm-mirror/llvm...939#diff-983f40a891aaf5604e5f0b955e4051d2R733

Probably GFX10_2 might fix a bunch of them, but that is a lot of HW bugs which needs SW workaround/mitigation
But this GFX10_1 has none of the new ML specific instructions added for Vega20 which would probably make it a purely gaming chip it seems.
Lacks ECC, new Vega20 dot instructions (for ML), lacks 1/2 DPFP, some multiply add features etc

Has these new features though not present in GFX9_0_6 (Vega 20), ( I might have missed some )

There is one feature defined but not added to GFX10_1, (FeatureCuMode)
Someone whol could explain these new Features would be great. (Only some of those are self explanatory for me, like the 8 bit dot functions for ML example)
But it seems, GFX10 has a lot more changes from GFX9 than GFX9 had from GFX8 or any other GCN uArch major revision transition/update


I guess the silicon needing a rework could be real... But apparently the bugs are deemed not critical and can be worked around in SW.

Do you guys think there will be another revision resolving some of these outstanding bugs prior to launch? I guess it's unlikely, with launch being a month away.

 

[DrMrLordX]

Radeon 7 clock similar to 1080TI FE around 1700mhz under load

What load? Games? The only applications I can run to make my Radeon VII clock anywhere near 1700 MHz are synthetics like Superposition. Mine boosts well over the max clock limit even with an undervolt in some of the games I've played. With a 1940 MHz OC, mine boosts to over 2050 MHz in The Division 2 on a regular basis, even with hotspot temps hitting 95C and higher. That's with no increase in power limit. And memclocks got as high as 1300+ MHz with a 1200 MHz mem OC. Which really surprised me.

 

[Mopetar]

...or he's just putting out as many controversial videos as possible to get the most views

If that were true, he'd post them more frequently than every other week or once a month. Other channels that deal primarily in rumors are pushing out videos every day. Even other more established sites like Gamer's Nexus that will cover or touch on big rumors as part of a larger weekly news video tend to publish more often than AdoredTV.

No, it is not. The only bottlenecks for AMD GPUs are always clocks, and Memory bandwidth.

Clock speed doesn't make sense because it speeds everything up, including the parts where the bottleneck occurs. It doesn't solve the underlying issue. AMD has apparently tried to push for higher clocks because making some of the changes that @Head1985 suggested in his post are not necessarily straightforward themselves. It's the kind of thing you only want to do if every other approach requires massive overhauls or you can't make small changes without necessitating a lot of other changes to coincide with them. It's even a bad idea as historically we've seen time and time again that chasing after high clock speeds as a solution tends not to work out well.

 

[tajoh111]

No, it is not. The only bottlenecks for AMD GPUs are always clocks, and Memory bandwidth.

If AMD can solve this, and clock their GPUs properly they will have equally good GPUs as Nvidia's. And that is BEFORE even architectural changes to their GPUs. If they could add architectural changes: more Geometry registered/culled, SIMDs doing more work with each cycle, etc the performance increase can be more than just simple nodeshrink/more ALUs approach.

This is incredible over simplification of a problem.

Increasing memory bandwidth and clock speeds would improve near any architecture linearly.

By this logic Pentium 4 could have become a good architecture if Intel hit their original clocking goals.

The problem that AMD and Intel both encountered is power consumption and an exponential increase for the increasing clocks further. This on top of having less headroom than Nvidia in terms of clocks.

What made Maxwell such a big deal was that it increased clocks while decreasing power consumption, on the same node while also increasing core occupancy which increased its performance per tflop. It also improved efficiency in terms of bandwidth usage.

Producing an all encompassing improvement is easier said than done. AMD has had 3 shots at producing such a change with FIJI, Polaris and Vega but they have not.

What is more important than noticing obvious problems(or in this case obvious improvements that can improve architectures), it's developing solutions that solve these problems. This is where Nvidia has delivered consistently in recent years.

E.g Fermi was inefficient and had high power consumption.

Kepler fixed the inefficiency. New problem recognized, bad performance in games which used hardware accelerated global illumination(these are the games where kepler losses to Maxwell the most).

Maxwell employs hardware global illumination, while changing the cache structure, SM layout, adds memory compression to improve performance, efficiency. New problem is Nvidia lacks async, and can fall behind in some directx 12 title.

Pascal doesn't fix the async issue but because maxwell is so ahead in the first place, it simply brings its tflops to near parity with AMD because of efficiency advantage. Not a long term solution but good enough for now since most games uses directx 11 and AMD efficiency woes due to stagnation with GCN. Async still an issue and FP16 may be valuable for future gaming performance. Another potential weakness is AMD freesync and the cost advantage it brings.

Turing addresses async and FP16 issues, curing the directx 12 woes. Develops drivers to deliver freesync on Nvidia cards. Current weaknesses, GPU are too large, RTX benefit and die cost are questionable.

Post turing, address die size with 7nm, potentially produce more cards like GTX 1660 ti without RTX if adoption of RTX does not pan out, etc.

This is what has allowed Nvidia to widen the gap against AMD. That is they recognize problems and are able to deliver long term solutions to problems. This in terms of business is a matter of execution strength that AMD lacks perhaps because of budget or leadership.

AMD current problem AMD needs to fix.

Improve memory compression to prevent reliance on HBM2.
HBM2 costs make it only practical in professional market as any price drops on GPU kill margins.
Day 1 driver performance.
Inefficiency Vs Nvidia due to having to clock cards closer to limits to have competitive performance.
Inconsistent silicon quality which causes AMD to apply higher voltage to increase yields.
Loud coolers.

How many of these problems existed since day 1 of GCN and still have not been fixed?

This is the difference between Nvidia and AMD. Nvidia is able to recognize a weakness and correct it(Directx 12 performance, efficiency, freesync). AMD on the other hand still has problems with silicon consistency which results in overvolts, problems with day 1 driver performance, loud coolers on their flagships, having to clock cards closer to their end limit which hurts efficiency. AMD solutions have mostly been short terms like HBM2(margins + die yield effects) or getting to a node first in the case of Vega 20. They need to employ more dramatic architectural changes that deliver universal improvements like Nvidia, unlike the GCN improvement that show a specialized improvement in games here and there(I.e not cherry picked scenarios like world war Z). Maxwell improvements across the board, Turing improved directx 12 performance across the board.

If AMD only solution to increase performance across the board is increasing clocks and memory bandwidth when their cards show weakness in terms of clocks and efficiency and HBM2 cost woes are getting worse(HBM2 is much more expensive than HBM1), it's time for a new solution.

 

[Stuka87]

For me, GCN for AMD's gpu is what Bulldozer was for its cpu.
If Navi is good, I'll be very happy. Probably buy one of the gpus. But I get the feeling that it'll just be good enough. Nothing spectacular, nor bad; like polaris, another gpu for console. That's my expectation. The reason I think that is I don't expect AMD to invest heavily on an architecture which is on its last iteration, if it has to choose between next gen and Navi.

Seriously? The HD7970 is quite commonly referred to as the best card of that generation. Its life greatly exceeded that of the GTX 680. Bulldozer was never considered good. Piledriver made that architecture adequate, but nobody considered Bulldozer to be good. While lots of people considered Tahiti to be amazing.