Today at Computex NVIDIA and Supermicro announced the product of their latest collaboration. Supermicro’s new SuperServer 6016T-GF Series is an evolution of the idea launched in the Tesla Preconfigured Clusters announcement earlier this month. Rather than interleaving S1070 GPU server boxes with CPU server boxes, the SuperServer puts 2 S1070s and 2 quadcore servers in a single 1U server package. The point: increasing the convenience of building out really big Tesla-based systems. insideHPC talked with Andy Walsh, the Director of Product Marketing for NVIDIA’s Tesla Business Unit, about the announcement and where NVIDIA is headed.
Today’s announcement for the Supermicro SuperServer 6016T-GF is an evolution that NVIDIA hopes will spark a revolution in high-end supercomputing deployments. The new 1U server packs two S1070-class GPUs, for 2 TFLOPS of GPU goodness, into a single rackmount enclosure with 2 quadcore CPUs. The result is a self-contained building block for large systems. This is the same compute density as the previously announced Tesla Preconfigured Clusters, which interleaved a single 1U NVIDIA box with 4 S1070s (4 TFLOPS) with a 1U CPU server box that serves as the host for the S1070s. That’s 4 TFLOPS in 2U, with cables in between.
The CPU part of the SuperServer today is a dual socket quadcore Nehalem system capable of supporting 96 GB (less than the maximum, but still healthy). NVIDIA and Supermicro are careful to talk about the “enterprise” features of this unit, by which they mean that the unit is IPMI 2.0 compliant so its fits into your management stack, and that it has other features like high efficiency power supplies that make it fit well in large configurations.
With the new release, the two companies are most certainly targeting large configurations. Andy Walsh says that this announcement is really about “making it more convenient for large-scale deployment.” Of course neither Supermicro nor NVIDIA will actually sell the clusters themselves, that will be the job of partners, who will add all the cabinets, network switches, and packaging that make a node a cluster. So who will those partners be? NVIDIA didn’t want to say too much, but I did get out them that there are over a dozen OEMs and system builders lined up, including Lenovo and Appro. One assumes there will be many announcements over the coming weeks.
This is an intelligent evolution of NVIDIA’s strategy, and puts products at all layers of the customer pyramid. The scientific desktop is covered by NVIDIA’s 1 TFLOPS cards and the Personal Supercomputer partner program, the mid-range server by the 16-32 TFLOPS Tesla Preconfigured Clusters, and now the high end is explicitly taken care of with the SuperServer building block. And with system builders lining up to be part of NVIDIA’s partner programs every time a new one is announced, I have to believe they are moving product. It may well be that this is a workable model for quasi-custom gear at the high end. Take a volume product, tweak it for HPC, but rather than develop your own cluster delivery and support business rely on partners to do the lifting and share the revenue (and split the expenses). With custom gear manufacturers pushed almost entirely out of the market (SiCortex, network manufacturer Quadrics being just the latest) it’s nice to see something working.
And there is evidence that NVIDIA’s strategy is working for high end computing. After introducing the Tesla S-series 1U devices for clusters in 2007 they showed up at #29 on November’s Top500 list and in about a dozen other large systems. The new SuperServer will be available later this month.