Comparing GPU architectures of GRID vs Workstation Models

Hi All,

I am searching for diagrams that show the architecture of GRID GPUs vs any of the other NVIDIA GPUs for workstations.

I found the diagram in the below link. The author of this article implies that the NVENC (encoder) is only present on a GRID GPU as that’s what’s needed to encode in H.264 and send it down the wire. Is this the case? If someone could provide another diagram showing the corresponding components for a non-GRID GPU, that would be great. Different diagrams highlight different components so I am looking for one that is similar to this diagram.

Also, I thought that H.264 encoding on the GRID GPUs (available in XenDesktop 7.6) was only an experimental feature and needs a special registry key to enable.

Thanks!

Richard

The K2 has the same GPU architecture as the K5000 inc. NVENC / NVDEC
The K1 has the same GPU architecture as the K600 inc. NVENC / NVDEC

The addtional technology added is to allow for multiplexing and time slicing of multiple GPU’s on a single PCIe card.

NVENC is a feature of all current GPU’s. Including the Quadro and GeForce line, and is utilised by applications to encode video streams (e.g. GameStream on GeForce)

It is exposed for remoting protocols by the GRID SDK and some vendors do make use of it for encoding h264 in their protocols. Whether a vendor chooses to do this or not is very dependent on their own roadmaps and requirements.

Citrix demonstrated ICA/HDX using hardware encoding recently at GTC, and a version of VMware’s BLAST protocol was showcased at the VMworld events in 2014 in conjunction with Google Chromebooks which also used NVDEC on the Tegra K1 for the decode.

I can’t comment on the timelines for those particular technologies becoming available, you’d have to speak to Citrix or VMware for that detail. However, when used there is a substantial reduction in the load on the CPU.

That’s a great answer. Thanks Jason and thanks for being so active on the GRID forums!

It was good meeting you at GTC this year at your session.

Cheers,

Richard

Great to meet you too. Always great to be able to put a face to a name and I hope you got a lot out of being at GTC.

One other question, Jason. In your session, you mentioned that the K2 had faster CUDA core clock speeds and the RAM was faster too. I have searched for "clock" "frequency" and "Ghz" in all the official GRID release notes and guides but have not yet found the clocks speed of the K1 and K2. I also looked through the K600 and K5000 documents. I can find there that the K600 has DDR3 RAM and the K5000 has GDDR5. Those Quadro documents don’t reference the clock speeds though.

Could you please point me in the direction of the docs that have the clock speeds. . . or just let me know what they are? :)

Thanks again!

Richard

Richard

I found the answer from a combination of techpowerup.com and anandtech.com. Here are the specs for anyone interested. I was surprised that the core clock speed of the K1’s was higher than the K2’s.

RAM:

-K1 = DDR3 @ 1.8 GHz

-K2 = GDDR5 @ 5.4 GHz

Core Clock Speed:

-K1 = 876 MHz

-K2 = 706 MHz

Cheers,

Richard

PS - To clarify, these are the specs from the K600 and K5000 that I just posted. They should be the same as the K1 and K2 respectively.

The card specs are all published here:

There’s links to the pdf’s that go into more depth for both cards on that page.

I don’t think I called out specifics of the cards in my session, just that there is a difference in the clock speeds you should be aware of when converting K2 & K1 profiles… I’ll listen to the recording now.

Thanks for pointing out where those docs lived. I am updating what I posted above, that were the K600 and K5000 specs, with the official K1 and K2 specs from the below two documents.

RAM:

-K1 = DDR3 @ 891 MHz

-K2 = GDDR5 @ 2.5 GHz

Core Clock Speed:

-K1 = 850 MHz

-K2 = 745 MHz

Thanks again, Jason!

Richard