Monday, December 08, 2008

gpu clouds

Will someone please build me a GPU cloud!


For research and commercial application cloud computing is a hot subject. It takes 10 minutes to boot an EC2 instance, it takes 10 minutes to boot 100 EC2 instances. I can get billed by the hour. It is a very scalable and cost effective model - I don't need to buy one hundred computers, but I can rent them by the hour. But these instances only have CPU's (because they are all sitting without monitors in a data center somewhere), and I want to buy time on a GPU cloud for several reasons:
  • Massive parallel processing of data - there are guys at Glasgow working on speeding up image recognition using the GPU, and many, many other applications around the world.
  • Production of real time 3D images - Companies like xtraNormal render simplistic graphics using something that looks like OpenGL/DirectX rendered to video. Modern video games show how fast very even more detailed content can be created.
  • Production of non real time images using Gelato or similar.
There are all sorts of fun things GPU clouds might be good for - we are seeing applications where we might not want to store rendered images, but produce and stream them on the fly. How about inserting custom Google-style individually targetted adverts into the billboards in a football game? Into a movie?

But no one has built me a cloud yet :(

It looks like amazon isn't really interested in this. And the NCSA has a massive...wait for it...16 node cluster. I guess this is a real business opportunity if someone wants to run with the ball!

The phrase to use when googling this subject is "GPU cluster" this gives more results, but none (that I found) for on demand use.

There's no such thing as a new idea - AMD want to build a cloud later this year. More or less for exactly the reasons above; lovely can't wait. It's strange that this is being done first by a hardware company (even if it's looking wobbly financially), does the rest of the cloud world know something AMD doesn't?

Hoopoe have a CUDA/Open CL cloud in alpha. They are giving away free usage, until they get the bugs out :) The interweb says it'll be announced at the start of november...


  1. What we need is rack fulls of nVidia Teslas:

  2. - consider it done, Twak, old friend

  3. post was updated as that slashdot "story" came along (non link jacked) ;). Still AMD are looking financially shaky and are only planning 1000 cores, not quite what you expect for a real scalable cloud (I want containers!)...

  4. finally these guys have one in alpha :)

  5. just got a reply from Hoopoe

    "Dear Twak,

    Thank you for your interest in Hoopoe.
    The registration is currently disabled for the Hoopoe service as we finished an alpha phase and preparing for the second one.
    Once the service will open again you will be notified."

  6. Just found out that you can rent GPU's in Europe :)