Vega/Navi Rumors (Updated)

[CatMerc]

Freesync 2 is a monitor only deal, all freesync capable GPUs support Freesync 2.

Freesync 2 just requires LFC and HDR

And in general stricter standards similar to GSync.

 

[Armsdealer]

13tflops vega 10 with 512 GBps memory bandwidth vs 11.3 tflops / 484GBps on 1080ti...

5.8 tflop rx 480 / 256GBps is generally slower / comparable to 3.8 tflop / 192GBps gtx 1060

Unlikely vega 10 beats gp102 unless there's some architectural miracle. Should beat 9 tflop / 320-352GBps gp104, but probably not by a lot.

Given the extravagant pricing of hbm2, how is vega not a disaster? I've been away what's the pro vega thesis here?

 

[Bouowmx]

If users want to use FLOPS, can they please remember NVIDIA GPU Boost: Pascal operates at 1.7-1.9 GHz, and up to 2.1 GHz.

 

[beginner99]

Unlikely vega 10 beats gp102 unless there's some architectural miracle

There will be similar as from Kepler to Maxwell which saved a lot of power and better utilization especially in dx11 games.

 

[CatMerc]

13tflops vega 10 with 512 GBps memory bandwidth vs 11.3 tflops / 484GBps on 1080ti...

5.8 tflop rx 480 / 256GBps is generally slower / comparable to 3.8 tflop / 192GBps gtx 1060

Unlikely vega 10 beats gp102 unless there's some architectural miracle. Should beat 9 tflop / 320-352GBps gp104, but probably not by a lot.

Given the extravagant pricing of hbm2, how is vega not a disaster? I've been away what's the pro vega thesis here?

Eh, Vega is the first time really major and deep changes came to GCN since Tahiti. I'd give it the benefit of the doubt in theoretical FLOPS vs gaming performance.

While the devil is in the details, much of the features they discussed for Vega are similar to what we've seen going from Kepler to Maxwell.

Hilariously if Vega has Hawaii's perf/TFLOP, it would be a good 20%-30% ahead of 1080 Ti. Hawaii is by far the most robust and balanced GCN chip in terms of ALU's vs Fixed Function Hardware.

 

[Karnak]

13tflops vega 10 with 512 GBps memory bandwidth vs 11.3 tflops / 484GBps on 1080ti...

5.8 tflop rx 480 / 256GBps is generally slower / comparable to 3.8 tflop / 192GBps gtx 1060

Unlikely vega 10 beats gp102 unless there's some architectural miracle. Should beat 9 tflop / 320-352GBps gp104, but probably not by a lot.

Given the extravagant pricing of hbm2, how is vega not a disaster? I've been away what's the pro vega thesis here?

Comparing Vega to Polaris is stupid. Comparing Vega to any GCN GPU is stupid.

Why? Because Vega is a new GPU architecture.

 

[CatMerc]

Comparing Vega to Polaris is stupid. Comparing Vega to any GCN GPU is stupid.

Why? Because Vega is a new GPU architecture.

I wouldn't call it stupid. Vega's architecture is built on the basis of GCN, nothing gets made from scratch. Even Zen is a mish mash of the best of Cat cores, Construction cores, and new ideas and tech built inside, around, and on top.

Seeing what problems they faced with GCN, and what steps they took to solve them, will tell us a lot about Vega.

That's not to say I expect 1:1 performance TFLOP to TFLOP between Vega and previous GCN. Even GCN now has major differences between chips. But comparing Vega to GCN is a valid and understandable practice.

 

[exquisitechar]

Hilariously if Vega has Hawaii's perf/TFLOP, it would be a good 20%-30% ahead of 1080 Ti. Hawaii is by far the most robust and balanced GCN chip in terms of ALU's vs Fixed Function Hardware.

True, it would be great if Vega is as well-designed as Hawaii was. I'm not expecting performance at that level but looking at Vega on paper, I will be shocked if it falls short of GP102.

 

[Karnak]

I wouldn't call it stupid. Vega's architecture is built on the basis of GCN, nothing gets made from scratch. Even Zen is a mish mash of the best of Cat cores, Construction cores, and new ideas and tech built inside, around, and on top.

Seeing what problems they faced with GCN, and what steps they took to solve them, will tell us a lot about Vega.

That's not to say I expect 1:1 performance TFLOP to TFLOP between Vega and previous GCN. Even GCN now has major differences between chips. But comparing Vega to GCN is a valid and understandable practice.

- NCU
- Primitive Shaders
- Intelligent Workgroup Distributor
- 2x improvement in peak geometry throughput per clock
- Draw Stream Binning Rasterizer
- ROPs now clients of the L2 cache
- High Bandwidth Cache Controller

to be continued. You can do whatever you want, that's up to you. But comparing Vega to previous GCN is... it's not that simple and IMO it's just not possible. Especially in terms of efficiency (e.g. Draw Stream Binning Rasterizer) and performance (e.g. Intelligent Workgroup Distributor, and even that has something to do with efficiency).

 

[Veradun]

- NCU
- Primitive Shaders
- Intelligent Workgroup Distributor
- more than 2x improvement in peak geometry throughput per clock
- Draw Stream Binning Rasterizer
- ROPs now clients of the L2 cache
- High Bandwidth Cache Controller

to be continued. You can do whatever you want, that's up to you. But comparing Vega to previous GCN is... it's not that simple and IMO it's just not possible.

fixed :>

I agree, it's impossible, at least atm, to extrapolate PPF figures for Vega based on previous GCN designs.

Whether you are on the optimistic or pessimistic side you can imagine PPF to go down/up due to improvements made mostly for computing or mostly for gaming.

Add to the equation that some features cost silicon while needing dev support to be effective in improving PPF (Primitive Shaders, I'm looking at you) and the guesswork gets even more complicated.

I guess I'll stick to "let's wait and see" as the best approach to the new arch

 

[davide445]

http://wccftech.com/amd-gpu-supply-exhausted-by-cryptocurrency/
With this status I continue to think didn’t make sense for AMD to release their consumer Vega.
Let's wait all Ethereum coins are found, first deplete Polaris stock, and only next release Vega. For Vega limit the purchase to max 2, limit mining in the firmware so that you need to hack it and go out of warranty for mining, push out a stripped down dual Polaris for the mining market.
Also with high Polaris price and lower availability the mining pressure will be possibly balanced also on Nvidia cards.

 

[CatMerc]

- NCU
- Primitive Shaders
- Intelligent Workgroup Distributor
- 2x improvement in peak geometry throughput per clock
- Draw Stream Binning Rasterizer
- ROPs now clients of the L2 cache
- High Bandwidth Cache Controller

to be continued. You can do whatever you want, that's up to you. But comparing Vega to previous GCN is... it's not that simple and IMO it's just not possible. Especially in terms of efficiency (e.g. Draw Stream Binning Rasterizer) and performance (e.g. Intelligent Workgroup Distributor, and even that has something to do with efficiency).

I agree that it's a big change, I've been saying that myself. It does not contradict my statement.

 

[Malogeek]

Yes it's still an evolution of the GCN architecture but Vega is a significant evolution compared to some of the previous iterations. So let's say you're both right

 

[raghu78]

13tflops vega 10 with 512 GBps memory bandwidth vs 11.3 tflops / 484GBps on 1080ti...

5.8 tflop rx 480 / 256GBps is generally slower / comparable to 3.8 tflop / 192GBps gtx 1060

Unlikely vega 10 beats gp102 unless there's some architectural miracle. Should beat 9 tflop / 320-352GBps gp104, but probably not by a lot.

Given the extravagant pricing of hbm2, how is vega not a disaster? I've been away what's the pro vega thesis here?

Vega is the first major revamp of the GCN architecture introduced in 2012. The architectural improvements in Vega are quite comprehensive and encompass every subsytem of a modern GPU. There is no way we can extrapolate Vega performance from Polaris or any other GCN architecture based GPU released till date.

 

[caswow]

13tflops vega 10 with 512 GBps memory bandwidth vs 11.3 tflops / 484GBps on 1080ti...

5.8 tflop rx 480 / 256GBps is generally slower / comparable to 3.8 tflop / 192GBps gtx 1060

Unlikely vega 10 beats gp102 unless there's some architectural miracle. Should beat 9 tflop / 320-352GBps gp104, but probably not by a lot.

Given the extravagant pricing of hbm2, how is vega not a disaster? I've been away what's the pro vega thesis here?

why did you left out the huge clock disparity? ~1250mhz vs ~2000mhz.

 

[Malogeek]

why did you left out the huge clock disparity? ~1250mhz vs ~2000mhz.

Clock isn't directly comparable between them as they both process data very differently per cycle. That's why metrics such as tflops are used instead.

 

[tential]

What was the preorder situation like for fury x? It's entirely plausible you won't get a Vega gpu til holiday season with shortages the way they are

 

[Bacon1]

There is no way we can extrapolate Vega performance from Polaris or any other GCN architecture based GPU released till date.

Yep. Worse case should be Fiji @ 1500 vs Fiji @ 1050. There are (supposedly) some massive changes to the architecture which will likely change a ton of the efficiency and perf/clock

 

[tamz_msc]

IMac Pro has Vega@11TF SP.

 

[Yakk]

IMac Pro has Vega@11TF SP.

If Apple can stick a Vega in the iMac screen like this new iMac Pro and still get 11TF, I'm impressed on the thermals.

 

[Bacon1]

IMac Pro has Vega@11TF SP.

https://twitter.com/GFXChipTweeter/status/871788766137995264/photo/1

Welp thats a huge boost

 

[Glo.]

1350 MHz core clock.

Expect that BIOS of the GPU High Sierra beta will leak soon, when devs or Beta testers will get the OS.

 

[Despoiler]

Vega FE launching first makes a lot of sense now. They had to for Apple to announce this.

 

[raghu78]

Vega FE launching first makes a lot of sense now. They had to for Apple to announce this.

Vega FE is a pro desktop card while the Vega in the iMac Pro is most probably a notebook type Mobile PCI-Express Module (MxM) . Apple could have even gone a step further and integrated the Vega GPU with HBM2 package on the iMac Pro motherboard . Moreover the iMac Pro is launching in December while Radeon Vega FE is launching in late June.

 

[Glo.]

There are two GPUs in iMac Pro.

Radeon Vega 56, and Radeon Vega 64. You know perfectly well what those numbers mean.