Vega/Navi Rumors (Updated)

[Krteq]

Previous GCN iterations? Fiji? We don't know :/

//nevermind

 

[Glo.]

Tile Based Rasterization. Primitive Shaders. Its completely new architecture. Or more factually, completely new iteration of the architecture.

 

[AtenRa]

And polaris is a $299 980Ti.

No, Polaris 10 is a $199 GTX 980 😉

 

[TimCh]

The question is here ... compared to what?

Has to be Fiji.

Sendt fra min SM-G928F med Tapatalk

 

[Flapdrol1337]

No, Polaris 10 is a $199 GTX 980 😉

Not really, anyway, I got a used 290 instead.

 

[xpea]

The question is here ... compared to what?

^^This

"4x power efficiency" 🙄
we all know its not compared to Polaris. And talking about Polaris..... this Vega bullshit list remains me a picture:

 

[RussianSensation]

Tile Based Rasterization. Primitive Shaders. Its completely new architecture. Or more factually, completely new iteration of the architecture.

It's looking promising. Based on IP8 vs. IP9, AMD's road-map showing Vega with higher perf/watt than Polaris, and rumours up to now that Vega wasn't just a scaled up Polaris, I never understood some naysayers who claimed that Vega would barely approach AIB 1070 cards and lose to the 1080.

I have my fingers crossed for a 1525mhz 4096 shader 12.5Tflops FP32 part. The amazing part is this 'just' AMD's 512GB/sec card so just an R9 290X Hawaii spiritual successor [I expect die size in the 400-450mm2 range]. The future is bright as Volta and Navi should give us 0.75-1TB/sec memory bandwidth and maybe 600mm2 die sizes before the shift to 7nm.

I'll be on the lookout for performance delta over Fury X in DX12 titles (to gauge shader IPC improvements and shader utilization), and improvements in shader intensive games such as The Witcher 3, Deus Ex Mankind Divided, The Division, BF1 4K, etc.

GCN ages like Fine Wine. Fury X is now almost 10% faster at 4K than a 980TI while 1080's lead has shrunk to 27%.
https://www.techpowerup.com/reviews/Zotac/GeForce_GTX_1080_Amp_Extreme/29.html

If Vega has 1.5Ghz clocks and 4096 shaders, 96 ROPs, 256TMUs, and doesn't have any major architectural bottleneck, it should easily surpass the 1080. Finally we may see some real progress on the high-end.

4X efficiency improvement sounds like marketing to me. The fastest card right now has 3x over Hawaii and 2x over Fury X. I am going to call BS on Vega having 4X efficiency improvement.

https://www.techpowerup.com/reviews/NVIDIA/Titan_X_Pascal/25.html

 

[RussianSensation]

^^This

"4x power efficiency" 🙄
we all know its not compared to Polaris. And talking about Polaris..... this bullshit list remains me a picture:

Everything on that chart is correct. Polaris 10 has higher IPC over Hawaii and Tonga. IPC over Tahiti scales from 15%-40% faster depending on the game. Computerbase already tested this. Polaris 10's biggest weaknesses are 32 ROPs and low memory bandwidth.

Unfortunately for GPU designers, games don't not scale linearly with Tflops/shaders. 1080 is barely 21-23% faster than 1070 is despite a 37% increase in Tflops/shader power.

Either way, even if Vega has 1350mhz clocks, but only a 10% increase in IPC, it should be up to 40% faster than Fury X in the most shader intensive demanding titles. I am hoping for 1500mhz+ GPU clocks and 45-50% faster at 4K.

 

[iBoMbY]

Everything on that chart is correct. Polaris 10 has higher IPC over Hawaii and Tonga. IPC over Tahiti scales from 15%-40% faster depending on the game. Computerbase already tested this.

The problem is some people made all kinds of stuff up based on the Polaris Architecture Preview (like the updated Command Processor was supposed to magically solve all DX11 issues), and many more believed them.

 

[crisium]

4x power efficiency compared to VLIW?

I know it's not compared to Polaris or Fiji.

 

[.vodka]

Well, they're mentioning 2x bandwidth per pin. That's one of the benefits of HBM2 vs HBM1.. which is used by Fiji.

Who knows apart from that. It does seem like a major reworking of the architecture, Polaris seemed mostly like evolutionary changes. CES is 2 days away. It'll be clarified soon.

 

[TimCh]

4x power efficiency compared to VLIW?

I know it's not compared to Polaris or Fiji.

Most likely fp16 vs Fiji(Fury non X), that is possible with tdp of 230W(higher with a higher than expected clock speed).

Sendt fra min SM-G928F med Tapatalk

 

[Glo.]

Polaris architecture was 2.5x better in efficiency compared to 2015y released GPUs. Vega will also be compared to those.

 

[raghu78]

AMD's marketing talk in the past with Fiji and Polaris was basically hype with reality turning out different. So I am going to just wait for Vega to launch and be judged on its actual perf and perf/watt. Anyway as long as AMD has made a significant architectural improvement everything will be fine. We need AMD Vega to compete with Pascal refresh of their current GP102 and GP104 based GPUs. I think 2017 could end up very competitive in both CPU and GPU markets. I think the last time that happened was more than a decade back. Lets all hope AMD brings back competition.

 

[Bacon1]

The problem is some people made all kinds of stuff up based on the Polaris Architecture Preview (like the updated Command Processor was supposed to magically solve all DX11 issues), and many more believed them.

Except it does help a ton with tessellation and other cpu related bottlenecks compared to a 290/390.

Look at games like The Division or GTA V:

https://www.computerbase.de/2016-06/radeon-rx-480-test/5/#diagramm-the-division-1920-1080_2

480 is 22% faster than 390 in The Divison and 15% in GTA V

 

[AtenRa]

Not really, anyway, I got a used 290 instead.

Well yea, most of the time its faster than 980 and equal in many other latest games. Exceptional performance for a $199 card.
But the jewel is the RX 470 4GB, some times faster and other times extremely close to GTX 980 4Gb for sub $170, amazing $/perf and Watt/Perf at 120W TDP.

 

[.vodka]

Found an interesting comment over at videocardz on the binning renderer.

Some more interesting info here on binned rendering, which likely applys to binned rasterisating.

http://www.trustedreviews.com/...

Essentially it helps greatly reduce the memory required per frame, reducing the bandwidth required to render a frame. Or in this case to rasterise the frame. So it should greatly increase rasterisation performance.

https://software.intel.com/en-...

Further reading here just reveals that Binned Rendering is another name for Tiled rendering. So AMD has moved Vega to a tiled Rasteriser.

'As noted earlier, Larrabee uses chunked (also known as binned or tiled)
rendering, where the target is divided into multiple rectangles, called
tiles. The rendering commands are sorted according to the tiles they
touch and stored in the corresponding bins, and then the contents of
each bin are rendered separately to the corresponding tile. It's a bit
complex, but it considerably improves cache coherence and
parallelization.'

Interesting. One of Maxwell's efficiency tricks was the move to tile based rendering.

This is a pretty fundamental reworking of GCN, then. Valid of the new year, new architecture tweet.

 

[Head1985]

Found an interesting comment over at videocardz on the binning renderer.




Interesting. One of Maxwell's efficiency tricks was the move to tile based rendering.

This is a pretty fundamental reworking of GCN, then. Valid of the new year, new architecture tweet.

good and along with "2x Peak Throughput per Clock" - probably 8x SE engines(finally!) vega should be decent upgrade over Furyx.
I will be upgrading to vega from GTX1070 which i dont like.

 

[Glo.]

Found an interesting comment over at videocardz on the binning renderer.




Interesting. One of Maxwell's efficiency tricks was the move to tile based rendering.

This is a pretty fundamental reworking of GCN, then. Valid of the new year, new architecture tweet.

good and along with "2x Peak Throughput per Clock" - probably 8x SE engines(finally!) vega should be decent upgrade over Furyx.
I will be upgrading to vega from GTX1070 which i dont like.

AMD was presenting Vega architecture on the New Horizon event, with Fiji drivers, because it has the same 64CU design. However, that means that Vega 10 is 10% faster than OC'ed GTX 1080, but WITHOUT all of the improvements, and new features available to the game to use. So it is absolutely worst case scenario.

 

[RussianSensation]

4x power efficiency compared to VLIW?

I know it's not compared to Polaris or Fiji.

I think they are talking about FP16 performance. Notice how AMD advertised the Instinct cards with 25Tflops of compute but that's actually 1/2 precision. For PC gaming this metric is meaningless.

AMD's marketing talk in the past with Fiji and Polaris was basically hype with reality turning out different. So I am going to just wait for Vega to launch and be judged on its actual perf and perf/watt. Anyway as long as AMD has made a significant architectural improvement everything will be fine. We need AMD Vega to compete with Pascal refresh of their current GP102 and GP104 based GPUs. I think 2017 could end up very competitive in both CPU and GPU markets. I think the last time that happened was more than a decade back. Lets all hope AMD brings back competition.

Fury X is hated too much. The main reason 980Ti looks good is overclocking.

In modern games, at 4K $649 Fury X > 980Ti and = Titan X.
https://www.techpowerup.com/reviews/Zotac/GeForce_GTX_1080_Amp_Extreme/29.html

RX480 suffered the biggest online pre-launch smear campaign in history. When everyone and their mother kept saying RX480 was a 7850-7870 replacement that would have R9 390/390X performance, we had HardOCP and various posters online claim that RX480 was a design failure that was meant to compete against 1070/1080. Pathetic level of bias & unprofessionalism that to this day goes unpunished.

Once again in modern games:

RX 480 > 1060 in DX11
https://m.youtube.com/watch?v=s12S74umruY

and DX12
https://m.youtube.com/watch?v=bSxGPCwTIlM

Thr competition's marketing will ensure that even if AMD has better products in every price level, they will not outsell the other team.

The amazing part is billions of dollars later and 2-3 decades, and AMD's GPU competitor still design a solid modern CPU, cries about console design wins only to later shove gimped versions of its failed smartphone/tablet APU no one wanted. AMD's CPU competitor has all but given up in the discrete GPU space and is being eaten for lunch in the GPU deep learning/AI space.

But I guess we'll still hear how AMD is a giant failure isn't their $649 card isn't 10-15% slower than a $1200 card since apparently all the 1% income earners have come out of the woodwork since 2013 when $1000 GPUs became the "enthusiast choice."

It's amazing to see the competitors to AMD do everything and anything to discredit price/performance as the most important metric 97% of PC users can relate to. Looking at the tens of millions of CPUs and peasant 1080p gaming and lower Steam PCs, > 80% of Steam gamers would wish they had even an RX 480 level card. Forums such as ours tend to call anything a failure unless it's the fastest even if the win is 1%. We've already learned from 9700/9800Pro, X800/850XT PE, X1900/1950XTX, HD5870, HD7970Ghz, R9 290X/295X2 eras that AMD's performance crown means nothing to the very segment of the population that ALWAYS seems to complain about AMD's products. It's good to see AMD finally stopping to care for this group of consumers. Maybe someone at AMD finally looked up the definition of loyalty in the dictionary. Most of PC gamers would much rather want the days of HD4850/4870/5850/5870/6950/7950/R9 290 back than today's market of $700 mid-range cards and $1200 flagships.

AMD doesn't have to beat the best. They just have to shake up the price/performance across key high-end segments as plenty of PC gamers have no interest paying $800-1200 for 2016-2017 high-end cards that will be easily matched by a $450 2018 mid-range. The opportunity for market share gains in the $299-549 space is absolutely enormous right now.

 

[Saylick]

Found an interesting comment over at videocardz on the binning renderer.




Interesting. One of Maxwell's efficiency tricks was the move to tile based rendering.

This is a pretty fundamental reworking of GCN, then. Valid of the new year, new architecture tweet.

Agreed. A tile based renderer would be a noteworthy improvement for GCN, not the mention the other improvements and features which hopefully remove some of the deficiencies of Fiji.

I do have to wonder how much of the variable width CU patent is incorporated into Vega. It was discussed heavily in one of the earlier Polaris threads. Perhaps a 0-20% IPC uplift depending on workload?
https://forums.anandtech.com/thread...-11-sku-spotted-has-16-compute-units.2469674/

 

[Glo.]

I think they are talking about FP16 performance. Notice how AMD advertised the Instinct cards with 25Tflops of compute but that's actually 1/2 precision.

4x Perf./watt is apparent in their GPU roadmap:

And Polaris was supposed to be 2.5x perf/watt over 2015 release GPUs. The same thing goes for Vega.

And it is obvious that it will be UP TO 4x perf/watt.

 

[Thedarkchild]

ATi cards were always looked at as inferior, no matter of their performance. Best selling cards in ATi 9xxx and Nvidia FX generation were actually FX cards. They were much bigger, louder, hotter and slower yet they where cards to get here, when in performance ring, it was no contest. They were also late for months and came out factory overclocked with huge fans to keep them from burning your house. Meanwhile 9800pro was beating it with ease, and it had especially huge advantage in AA/AF aswell as newer titles using Shader 2.0.

When I was younger (6-7 years ago) I worked at retail computer shop part time as a student. Basically, 3 out of 5 of my coworkers always recommended Nvidia cards, other one would recommend based on their performance. These 3 would literally say "Just go with Nvidia...". That was a time of GTX400 and GTX500 series. They sold better then much better AMD equivalent cards, I was flabbergasted. I even asked them why are they recommending worse cards for same money and all they said was "Nvidia and Intel are a way to go if you are gaming, thats no brainer".

Even today you have bunch of people on forums saying "AMD always disappoints, they have been losing for more then a decade" "When was the last time they had good chip?". That is rewriting of a history, up until Maxwell, AMD cards were every bit as good as Nvidia counterparts (if not better, considering their design innovations like 360 GPU and unified shaders). Only last 2 gens have they really fallen back, primary on perf per watt. And who wouldn't? With what money would they bankroll Zen R&D, which is make it or brake it for entire company, if not from putting GPU design updates aside? AMD has repackaged Tahiti chips for 2 generations, only Polaris getting some needed updates. Yet, even with all that their cards have held up admirably well. 7970 is beating GTX680 with ease now (was similar back in a day), while 290x/390x are faster cards then GTX970 is. RX480 had ~10% deficit 6 months ago, now its faster card. The newer the title and API, the better RX480 is.

Yes, they have worse TDP. Yes, they run hotter. But AMD has agreement with GF. They are factually using worse fabs then Nvidia because of their contract. Their design and DX11 drivers are probably not as efficient as Nvidias (especially latter), but that tends to happen when one company makes tons of money on slower chips (GTX400 and 500) and other loses on great ones (HD4xxx-7xxx). I will not write off AMD. Their hardware team is still sound and they know how to design good chips. They are working on both high powered next gen consoles (Scorpio and Pro), so they ought to know a thing or two where the things are headed. Zen money is now gone, and Vega probably got biggest R&D budget in years. Who knows what future holds as far as performance goes? AMD is used to making high density chips with hardware that new APIs need. Will Nvidia easily add all that without touching efficiency of their cards? Will they be able to run at such high clocks with new uarch and changed CUDA cores (assuming they change them with Volta?)? Who knows, only thing that is likely to remain the same is the fact that Nvidia, no matter of their performance, will remain "go to" company until AMD marketing team starts doing much better job then they currently are doing.

 

[CatMerc]

http://www.freepatentsonline.com/y2016/0371873.html

Patent for the Binning rasterizer probably.

 

[ConsoleLover]

Review the rules for the AMD subforum, and stop posting inflammatory garbage here.
-- stahlhart