The advantages of adding the parallel capabilities of graphics processors to dual- and quad-core CPUs in advanced server designs is on display this week, but you'll have to go to Taiwan to see it.
At Computex 2009, NVidia and Supermicro are showing the new 1U SuperServer 6016T that features dual- or quad-core Xeon 5500s plus Nvidia's new Tesla GPU, a combination that both companies claim can deliver up to 12 times the performance of traditional quad-core designs.
NVidia has been touting GPUs as general-purpose processors for some time now, arguing that the massively parallel approach is better suited to the heavy data loads found in high-traffic enterprise environments, as well as for the kind of number-crunching needed for scientific research, oil and gas exploration and medical imaging.
For a 1 U device, the SS6016T packs quite a punch. It delivers 2 Teraflops and is outfitted with three 3.5-inch drives, up to 48 GB of SDRAM and dual LAN controllers with Intel Gigabit Ethernet capability. The device comes in two flavors, the GF-TM2 that uses two of the Tesla M1060 cards, and the GF-TC2 sporting a pair of C1060 cards. The main difference is the cooling architecture. The C card has an active fan sink, while the M card uses a passive sink with airflow managed by the server, the chief advantage being you get to extend the host's monitoring capabilities directly to the GPU.
The Tesla itself provides up to 240 streaming processor cores running at 1.3 GHz with 4 GB of dedicated GDDR3 memory with a top memory bandwidth of 102 GBps. It features both single and double precision floating point performance, topping out at 933 singles and 78 doubles.
The device also runs NVidia's CUDA parallel architecture, a C-based system and toolkit designed to help developers better harness the GPU for general-purpose applications. The platform has already proven to be a hit in high-performance computing circles with more than a million users so far, but it remains to be seen whether it can draw an audience among the broader enterprise community. NVidia recently teamed up with Dell and Penguin Computing to deliver a Tesla-based desktop.
In coming out with the 6016T, Supermicro and NVidia are betting that demand for clustered server architectures will continue into the short-term at least. Right now, that looks like a pretty safe bet, as more and more enterprises power down their big iron systems in favor of the more flexible cluster.
Going forward, however, the risk is that enterprises will start to put more and more reliance on cloud-based architectures, pulling their resources from third-party providers. Whoever provides those resources will, of course, need an IT infrastructure, but if the trend away from home-grown IT gathers steam as many experts predict, than demand for servers of all kinds could drop even further.
Still, that's the kind of thing that could take a decade or more to play out. And in the meantime, the combination of small form factor and high output will be pretty hard to beat.