Nvidia today said its graphics processors will be part of a new IBM server used for high-performance computing and webscale deployments. IBM’s iDataPlex servers will combine x86-based CPUs with GPUs in order to offer parallel processing on the GPU, which is more energy efficient. However, the announcement that IBM has turned to Nvidia GPUs after it stopped producing its own specialty parallel processor also offers a guide on how to succeed in the data center.
IBM’s own Cell processor, originally developed for the Sony Playstation with Sony and Toshiba, was killed late last year. It was used as an accelerator chip in the high-performance computing world to build cheaper, greener supercomputers. We wrote about the efforts IBM made to push it into the HPC market as well as into the data center, but it turns out the market didn’t want the Cell processor.
There are likely a few reasons for its death and the simultaneous survival of GPUs from which those pushing ARM-based servers should learn. In the Darwinian environment of the data center, where the search for energy efficiency, performance and low-cost computing collide, Nvidia had two advantages over the Cell. One, it had the CUDA programming tool introduced in 2007 that made it possible for developers to program in C and then watch their efforts run easily on the GPU’s parallel architecture. IBM had developer efforts, but programming for the Cell was still more difficult. And specialty programming raises costs.
Nvidia also has a huge consumer market for its graphics chips, which lowered the overall cost of the chips, making them cheaper to buy for large-scale computing efforts. IBM’s Cell processors cost a lot to develop and a market for the chips outside of the Playstation and IBM’s supercomputers never materialized despite efforts by Toshiba, Sony and IBM. Both the Cell processor and GPUs deliver incredible performance gains when dealing with parallel processing compared to a CPU, but huge performance gains alone are not enough.
So as chip and server vendors attempt to develop hardware for the new webscale companies, the tale of IBM’s cell and NVdia’s continued survival offers some solid lessons, namely: It’s not enough to be green; you have to be cheap, easy to program and orders of magnitude better at crunching data. As for bringing technologies once reserved for supercomputers, check out our Exascale Grail panel at Structure 2010 in June on where HPC is headed next and why it matters for software as a service, platforms as a service and other webscale vendors.