[BenchLife] AMD FirePro S7150 VDI cards.

[ShintaiDK]

https://benchlife.info/amd-firepro-s7150-and-s7150-x2-announce-for-vdi-market-02022016/

Good to see more options for the VDI concept. This can also be used for local gaming and cloud gaming in more game oriented forms.

They are based on Tonga GPUs with 8GB GDDR5. 150W for single GPU, 265W for dual GPU.

 

[Silverforce11]

More info here:

http://techreport.com/news/29666/fi...rds-bring-hardware-gpu-virtualization-to-life

Tonga and Fiji has a HWS that acts like an ACE with enhanced function. It allows the rendering pipeline to run in parallel so one stream won't bottleneck another, such as in Virtual multi-user usage cases.

@zlatan actually pointed this out way back, way before it was publicly revealed by AMD so kudos for the in-depth knowledge on the uarch.

AMD says the S7150 can support 16 users, but it's only for light desktop work. For CAD or 3D task, it's more 5-7 which is the limit of the ACEs, so that's interesting.

 

[ShintaiDK]

Seems to be the standard. Nvidia supports 16 users per GPU as well via VDI.

 

[Silverforce11]

The difference between GRID and AMD's is its a hardware level support, for usage cases where heavy 3D rendering is needed, or OpenCL is used, the GPU is still capable of handling ~6 users each GPU without one user bottle-necking the others.

IIRC, NV's Workstation extended setup for those usage scenario, assigns the entire GPU to 1 user per GPU due to latency issues. As in, it can do 16 for light desktop only.

There's the advantage but software wise, they are playing catch up to NV. Once IT Admins get accustomed to some ecosystem its gonna be hard to get them to readjust, that's the biggest downside I see for AMD entering this market.

Some more info:
http://www.amd.com/Documents/Multiuser-GPU-Datasheet.pdf
http://www.nvidia.com/object/grid-technology.html

 

[ShintaiDK]

That is not correct.

 

[Silverforce11]

TR quote

For one, S7100-series GPUs can provide OpenCL support to all virtual users without relying on pass-through mode (GRID), which dedicates the resources of an entire graphics card to a single virtual user.

 

[dark zero]

Nice concept, but both came in a moment where 16/14 nm are there.... Still since their replacements from both might come after a year...
Hopefully, Intel didn't released something like that

 

[NTMBK]

Nice concept, but both came in a moment where 16/14 nm are there.... Still since their replacements from both might come after a year...
Hopefully, Intel didn't released something like that

Tonga is a big die, not sure if we will see its 14nm replacement this year.

 

[zir_blazer]

And so, it is finally revealed that Tonga has SR-IOV support. Back when AMD announced GPU virtualization via SR-IOV I was curious on what existing GPU dies they would use for this purpose since there was nothing on roadmaps. Tonga was a rather logical choice since it was one of the latests at that time, so it has that hidden feature since a long time ago.
Chances are that Fiji has it too, but its 4 GB VRAM cap severely limits it as a Workstation class GPU for virtualization. Tahiti refresh, Granada, maybe has it too, but I'm not sure.

BTW, while AMD talks about these FirePros, there is a huge party going on in the Linux world. The guys from Red Hat responsible for developing VFIO, some other Xen-related devs, the Intel devs that are working on iGVT-g (Also know as XenGT/KVMGT), and nVidia devs that work on the GRID system, seems to want to standarize a way to implement vGPUs. I don't recall seeing any guy representing AMD at that party, but maybe it isn't critically needed since SR-IOV is merely Passthrough with the twist that the resource sharing part is done Hardware side (I suppose that it also needs a minimal API to allow you to set quotas and caps, which may be done from a priviledged Driver in the host).