Looks like the first stand-alone Larrabee chip has been canceled.
The August 2009 issue of ;login: from USENIX has a nice article on programming video cards by Tim Kaldewey entitled “Programming Video Cards For Database Applications”. Sadly, the article is only available to USENIX members until August of 2010.
Kaldewey surveys the past and present of programming video cards for non-graphics purposes – from the early days of using the graphics APIs to fool the GPU into thinking it is rendering graphics when it is really performing a general-purpose calculation, to the present era of general-purpose APIs such as CUDA.
He also shows a back-of-the-envelope calculation for building out a 100 teraflop data center using 100 GPUs versus 1400 CPUs, including power consumption differences.
If you are a USENIX member, the article is a good read. Sadly, it won’t be current when it finally becomes freely available.
[The same issue of ;login: also has a nice article by Leo Meyerovich: “Rethinking Browser Performance“.]
I’ve previously posted on the topics of CUDA and Larrabee. I continue to be intrigued by the possibilities that open up as multi-core GPU programming becomes available. For applications that need many threads this should present interesting opportunities. Why bother struggling to run your parallel application in the meager 4 or 8 cores of your CPU when you can offload the work to 32 cores?
While I’m on the subject of Dr. Charles Forgy’s talk at ORF 2008…
We know from ORF that Dr. Forgy is working on a 4-core machine with his parallel version of OPS/J. I’m curious to see the same ideas applied to 32+ core video cards such as the CUDA architecture and the upcoming Larrabee.