Wednesday, April 17, 2013

The Joy of RDMA?

In this follow-up podcast to the GPU Technology Conference, the Radio Free HPC team mulls over a talk by GE’s Dustin Franklin, GPU app specialist. Dustin’s topic was GPU-direct RDMA; was this a first look at real-world RDMA with GPU-to-GPU communications?

Follow along as the guys describe flow charts on technical slides that are not yet approved by viewing for the “great unwashed masses” – but make no mistake, they’re impressed by what they saw. Dan “knows a guy” who can divulge more, and offers to arrange an inquisition with Henry. Henry promised to “be nice,” whatever he means by that. Rich missed this GTC session and several others while “conducting interviews,” whatever he means by that. Dan offers another characterization.

And this just in: there’s a great deal of information available on the Internet.

Download the MP3 * Subscribe on iTunes * RSS Feed

FPGA’s and Stuff

In this episode of Radio Free HPC, Rich, Dan, and Henry discuss the recent buzz surrounding FGPAs. After being sidelined by accelerators, they’re increasingly being used in appliances.

Big vendors are talking about FPGAs not only for appliances but for general-purpose systems as performance assists. Are we headed back to the future?

The guys discuss the ins and outs of FPGAs and why, in some cases, they could be a huge win for the organizations that implement them. But is the architecture flexible enough? For enterprise and Big Data, perhaps it is. If you need to perform the same algorithms over and over again, FPGAs could be a perfect fit.

As with all things tech, there are a few cautionary notes to be sounded. Amassing more and more appliances can lead down a tricky road. Will their use in workload-optimized systems lead to vendor lock-in? Can you really teach an old FPGA new tricks? And can they be weaponized?

Most importantly: how are servers like cattle? Tune in to find out…

Download the MP3 * Subscribe on iTunes * RSS Feed