Radeon Vega Architecture Preview Thread

[IllogicalGlory]

Vega seems like a more Pascal/Maxwell-esque design with high clockspeeds along with a fancier architecture, taking up space where CUs might have gone in older AMD designs.

 

[sm625]

This is probably a good thing because it means the changes must be more significant than I first suspected. I figured Vega was simply two polari strapped together but that is clearly not the case. A 4096 core polaris would only be 390 sq mm. They spent an extra 150 sq mm on something, and we just have to assume they did it for a good reason. Who knows, I like to entertain the thought that maybe they stuffed 4 zen cores in there and fused them off... If they're really having budget problems that might be the only way to get their APU out the door before back to school season. Or heck, dont fuse them off, use them to make games run faster.

 

[dogen1]

how many compute units would be in this though. 14 nm 500+mm^2. not likely 4096 right?



in useful cases, not just "I want 64 x tessellation and i will have it!"

Yeah it'll definitely be useful.

 

[Vesku]

I'm with IllogicalGlory, less dense CUs would point to a high frequency target.

 

[turtile]

So it looks like Navi must be another low end series like Polaris (takes up less die space) and then the true replacement of Vega will be after that.

 

[tential]

So it looks like Navi must be another low end series like Polaris (takes up less die space) and then the true replacement of Vega will be after that.

I've already said AMD is on a segmented release cycle. Instead of releasing a full line up over 2 years like the competition. They are releasing a high end architecture then a low architecture and alternating. It's a new strategy. It will be interesting to see how it effects AMD and whether they do any better compared to the competition.

 

[Bacon1]

People seen these yet?

Aren't those the slides from months ago?

Yeah:

The slides we are sharing today are the same slides we described to you back in September (so they are not new). We are posting this because so far everything turned out to be (almost) correct.

 

[turtile]

I've already said AMD is on a segmented release cycle. Instead of releasing a full line up over 2 years like the competition. They are releasing a high end architecture then a low architecture and alternating. It's a new strategy. It will be interesting to see how it effects AMD and whether they do any better compared to the competition.

I wonder if it's only to do with their financial situation. The long time between nodes makes enhancements a lot less significant so they have the ability to save on R&D.

 

[IllogicalGlory]

Aren't those the slides from months ago?

Yeah:

We didn't see them months ago, they were described months ago.

 

[Bacon1]

We didn't see them months ago, they were described months ago.

Ya

 

[William Gaatjes]

I've already said AMD is on a segmented release cycle. Instead of releasing a full line up over 2 years like the competition. They are releasing a high end architecture then a low architecture and alternating. It's a new strategy. It will be interesting to see how it effects AMD and whether they do any better compared to the competition.

ATI did something similar in the old days. When a new process was announced, first the "low" architecture would be made on the process. After gaining experience with the process, the"high" architecture is made on the new process. It makes sense to do so, smaller dies and in the low architecture, more resources can be spend to solve issues of the new process. Maybe they have turned back to a similar strategy.

 

[Glo.]

ATI did something similar in the old days. When a new process was announced, first the "low" architecture would be made on the process. After gaining experience with the process, the"high" architecture is made on the new process. It makes sense to do so, smaller dies and in the low architecture, more resources can be spend to solve issues of the new process. Maybe they have turned back to a similar strategy.

Navi is scalable architecture. It means that it can scale from small to bigger GPUs without any problem, without much work to do with the silicon design. Raja touted however, that on smaller nodes we will see a shift to multiple smaller number of GPUs.

Think about it this way. Vega lift the limit of 4 Shader Engines in the GPU, so we will see with Vega even 8 Shader Engines in 64CU GPU. If you will split up the designs, into parts, you can scale the performance extremely easily.

For example it can be easier to get 250W GPU from dual 8 Shader Engine setup, than from single 16 Shader Engine.

 

[crisium]

Nvidia was genius to build a GP102 that is much smaller but has the same gaming performance as their monster GP100. But they also have the cash.

AMD is focused on gaming and HPC both, just like Nvidia. But the gaming focus in consoles now, imo.

I'm starting to believe the rumour the Vega 11 is basically Polaris 10 on NGCU.

http://i.imgur.com/ug5E6SE.jpg

Project Scorpio is 6 TFLOPs and 320GB/s. It's not Polaris, as that would require the full 2304 shaders at ~1300MHz - higher than 480 stock speeds. Consoles are always downclocked (see PS4 Pro's downclocked Polaris 10 equivalent). I think Vega 11 could very well be 2304 shaders again. But with Vega 10 at a supposed 1500MHz, moving that same NGCU to Vega 11 means 1300MHz is the console downclock. Throw in G5X on 256-bit and there is your 320 GB/s.* I know there's 12GB of RAM in Project Scorpio, so this would only work if the console is using segmented system and video memory. Could be a big hole in the theory, or it's simply 384-bit. =P

So AMD's gaming focus is Vega 11, but it's a slimmer chip because of consoles. Desktop version will be 1500MHz+ probably then, giving a RX 580 that's a decent bump over the 480 assuming the architecture has not regressed (and a much bigger bump if IPC has improved). Vega 10 is big, fat, and only 4096 shaders because of the HPC functions. If AMD had the money they might do a smaller chip with equal (or more) shaders without HPC stuff, but no console wants something that big.

Vega 11 = Microsoft Xbox. Vega 10 = HPC. Polaris 10 (modified) = Sony PlayStation. Polaris 11 = Apple. AMD is no longer focusing on dGPUs, but instead releasing dGPU versions of what they are making for their big 3 contracts + HPC.

 

[PhonakV30]

Are you sure It's 4096 ALUs? AMD didn't say anything about number of ALUs.

 

[crisium]

If you believe the 64 CU from this:
http://cdn.videocardz.com/1/2017/01/AMD-VEGA-10-specifications.jpg

 

[CatMerc]

Are you sure It's 4096 ALUs? AMD didn't say anything about number of ALUs.

Well, the Videocardz info that was right so far mentioned 64 CU's. And there was the Linkedin profile leak with graphics IP v9 and 4096 shaders.

 

[Glo.]

I have watched the PCWorld interview with Raja. One thing apparent about Vega is this: it is designed for the future. Tile-Based Rasterization will be immediately apparent as improvement. However, Primitive Shaders, and Geometry Pipeline are designed for the future. Doom for example has not been updated for this feature, and Vega does not benefit from it whatsoever.

 

[William Gaatjes]

I have watched the PCWorld interview with Raja. One thing apparent about Vega is this: it is designed for the future. Tile-Based Rasterization will be immediately apparent as improvement. However, Primitive Shaders, and Geometry Pipeline are designed for the future. Doom for example has not been updated for this feature, and Vega does not benefit from it whatsoever.

It might very well be that vega also ends up in consoles.
In consoles, it is close to the hardware programming, so in the console environment it is easy to implement it since vega also has backwards compatibility .
I am not sure about this , but does dx12 and vulcan support additional features that where not the immediate standard ?
Like for example the primitive shaders ?

 

[dogen1]

I am not sure about this , but does dx12 and vulcan support additional features that where not the immediate standard ?
Like for example the primitive shaders ?

Vulkan will most likely through extensions.

 

[ZGR]

I think it is safe to say Vega will be a great long lasting 4k card, specifically because of HBM2. We see how well Fury does with 'only' 4GB of HBM at 4k, but I still am skeptical about non-optimized games. 8GB of HBM will be enough for 4k for a few years I can imagine. At that point, 8k will be the next big jump in resolution.

An AMD 4k rig sounds like a great 2017 build! I haven't been this excited for AMD in a long, long time!

 

[sirmo]

It's clear AMD is targeting higher clocks with Vega. Which will likely increase the area needed per CU compared to Polaris, which is very densely packed.

 

[Karnak]

https://www.youtube.com/watch?v=Y8tDaPLHxiE

 

[Krteq]

I don't know if this was posted here before.



Only one PCI-E 6+2-pin

 

[crisium]

Then it may be like the RX 480, where reference uses exactly the TDP and virtually every aftermarket uses extra pins.

 

[Glo.]

Then it may be like the RX 480, where reference uses exactly the TDP and virtually every aftermarket uses extra pins.

That is correct.