AMD Posts Open-Source Vulkan Driver Code For Vega 12 GPU

[phoronix]

Phoronix: AMD Posts Open-Source Vulkan Driver Code For Vega 12 GPU

AMD developers have done their weekly code drop to their official open-source Linux Vulkan driver code. This week there are fixes while most interesting is initial support for the yet-to-launch Vega 12 graphics processor...

http://www.phoronix.com/scan.php?page=news_item&px=AMD-Vulkan-Vega-12-Posted

[hetzbh]

There were few interviews in the HotChips conference, and although VEGA12 wasn't mentioned in name, my guess is that this VEGA 12 will be a laptop version of Ryzen 2400G with 45W TDP and built using the 7nm.

[Michael]

There were few interviews in the HotChips conference, and although VEGA12 wasn't mentioned in name, my guess is that this VEGA 12 will be a laptop version of Ryzen 2400G with 45W TDP and built using the 7nm.

Actually the Vega12 LLVM patches indicated already it's a discrete GPU - https://www.phoronix.com/scan.php?pa...20-AMDGPU-LLVM

[juno]

There were few interviews in the HotChips conference, and although VEGA12 wasn't mentioned in name, my guess is that this VEGA 12 will be a laptop version of Ryzen 2400G with 45W TDP and built using the 7nm.

What they called "Raven Ridge 2018" at Hot Chips was Ryzen 7 2800H, a 35W TDP variant (e.g. for AiO) of Raven Ridge. So nothing special. Also it wouldn't make sense to go with 7nm and then push it to 45W.

They removed this detail from their page but the archive still has it:
https://web.archive.org/web/20180822...nologies/25x20 (open footnotes)

[Girolamo_Cavazzoni]

The driver code for Vega 12 does confirm though that it is indeed quite cut-down compared to Vega 10, which some rumors have suggested it being a lower-end part

It rather looks like a Vega 10 cut in half. Not bad if you ask me, should be near Polaris 10.

[juno]

It rather looks like a Vega 10 cut in half. Not bad if you ask me, should be near Polaris 10.

The code indicates 20 (4*5) compute units, Vega10 has 64 CUs.

I also think the mentioned hardware stereo rendering is interesting. Might allow for single-pass stereo rendering for VR.


I can think of two potential products AMD has in the pipeline:
Fenghuang Raven: https://www.anandtech.com/show/13163...-amd-ryzen-soc
Vega mobile: https://www.anandtech.com/show/12233...-vega-on-7nm/9

However, both should have more than 20 CUs...

[Girolamo_Cavazzoni]

Yes, you're right, I obviously looked at the wrong parameter.

[tildearrow]

Typo:

- A PAL optimizartion that should help MSAA anti-aliasing performance.

[sykobee]

The code indicates 20 (4*5) compute units, Vega10 has 64 CUs.

AMD have not put performance-leaking details in their code for a long time - they always read this from the AtomBios on the chip. This specifically includes the number of CUs in a particular hardware design.

This code shows that the arrangement within the GPU is different, but not how big the GPU actually is. This may be to optimise the GPU over the prior Vega design, i.e., to balance the design better.

If this is a 20CU chip (which feels a little low to me, I was expecting 28-32) then it could also serve as a Radeon R650 and R660 level chip, killing off Polaris 11/21.

[GreenReaper]

Looking at the massive commit, my best guess is that Vega12 is targeted towards VR headsets - either built-in, or powering them.

There are settings for stereo control (left eye, right eye, multi-plane, etc.) that apply only to Vega12, and a cut-down low-powered reduced-compute-unit version which nevertheless has plenty of shader engines (4 engines with 5 CUs, i.e . 20 CUs total, vs. 4 with 16 CUs on Vega 10 and 1 with 11 CUs on Raven). It's got half the number of texture cache blocks as Vega 10, but twice that of Raven Ridge.

So . . . figure twice the performance, which would bring it into line with current recommendations for VR, and be about a GTX 970 (compare to Raven Ridge Vega 11 - needs ~4x rendering, 2x everything else) or mobile 1060 (same comparison) level of performance. It'd need dedicated memory to back that up.