Without any doubt, SC09 conference in Portland was marked with further development in HPC space with largest accent on reducing the cost of TFLOPS. As such, GPU technology is taking pole position in HPC space when it comes to processing power.
In our discussions with several HPC OEMs, we learned that approximately 90% of new HPC installations on order will feature at least one GPU board or whole GPGPU server "for evaluation purposes" i.e. Tesla S1070, S2070 or an 8GPU machine such as Tyan’s FT72B7015 barebone carried by HPC vendors such as Colfax International or system integrators such as AVADirect.
Intel’s GPGPU SNAFU – Reccomending Server that doesn’t support GPGPU at all
Thus, a picture where Intel Corporation, maker of 80% CPUs inside all Top500 systems is promoting GPGPU was somewhat a surprise to us. At first, small-scale events such as SC09 aren’t the ones clad by deceiving clouds of marketers and marketing messages but events where multi-million dollar contracts are signed and where there is business – there is no talk for "messaging".
As an example of the ideal link between Intel’s Xeon 5500 processors and GPGPU products, Intel was promoting HP ProLiant DL160se G6 server, and the demos in front of the panel were all performed by nVidia Quadro FX 5800 and Tesla C1060 boards. According to Intel, ProLiant DL160se G6 should pack up to six GPUs in a single server.
And everything would be fine, if there wasn’t for one small thing: according to HP’s website, ProLiant DL160se G6 is a 1U server incapable of hosting a GPGPU card. Unlike Supermicro’s new line of dual-GPU capable servers, DL160se G6 comes with very limited expansion options and nVidia Tesla C1060 or AMD FireStream 9170/9270 boards are not featured on the approved upgrade list.
Last minute banner SNAFU, we’d say.