Super Micro stuffs super node into pizza box
Thin crust CPU-GPU
Motherboard and whitebox system maker Super Micro is talking up its products at the International Supercomputing Conference in Hamburg, Germany this week, including a hybrid Xeon-GPU cluster node, blade servers sporting the new six-core "Istanbul" Opterons, and a baby super workstation.
Perhaps the most interesting machine that Super Micro is showing off is the SuperServer 6016GT-TF-TM2. This is a variant of the 1U SuperServer Twin line of machines, which puts two half-width, two-socket motherboards side-by-side in a single chassis, sharing power supplies and peripherals. The hybrid CPU-GPU box that Super Micro is showing off at ISC '09 puts one skinny mobo in the center of the box and then plunks two of nVidia's Tesla M1060 graphics co-processors in the empty space on either side of the board, with each one plugging into is own PCI-Express 2.0 x16 slot on the motherboard.
The motherboard used in this machine is a modified version of the normal Twin board using Intel's "Nehalem EP" processors and "Tylersburg" chipset since that regular Twin mobo only has one x16 slot for graphics cards or GPU accelerators like the Tesla M1060. When you are a motherboard maker with server aspirations, as Super Micro is, you can crank out a new motherboard pretty much at the snap of the fingers to suit your needs. This particular mobo supports 96 GB of DDR3 main memory and has room for three 3.5-inch disks, an integrated Matrox G200eW graphics card, and two Gigabit Ethernet ports.
The Tesla M1060 GPU cards were announced at the beginning of June, and they are dense enough to allow the GPUs to be put into the same rack enclosure as a server node. Prior rackable CPU-GPU combinations put four S1070 GPU co-processors in their own 1U chassis, then linked them via PCI-Express bus links to adjacent 1U x64 server nodes. With the design put forth by nVidia and Super Micro (and which will be mimicked by others for sure), the whole shebang can be crammed into a single 1U package.
Super Micro is billing the CPU-GPU box as being able to dish out 2 teraflops of number crunching performance, which - as you can see in this YouTube preso from nVidia - is about ten times the performance you can get using normal CPU-based nodes. The GPU has 240 cores spinning at 1.3 GHz, has 4 GB of its own GDDR3 memory, and is rated at 933 gigaflops on single-precision floating point calculations and a paltry 78 gigaflops on double-precision math.
It is basically half of an S1070 GPU appliance crammed into the same case as a half-width two-socket server. What Super Micro and nVidia are not saying is that this approach yields roughly the same teraflops as you would get using a twin server 1U box and an S1070 half-dedicated to each server node.
As El Reg previously reported back in May when supercomputer maker Appro announced its own CPU-GPU servers, the Tesla co-processors due for launch in the first half of 2010 are expected to be a lot more elegant than the PCI-Express versions sold today and are also expected to have much better double-precision math - something that really limits the appeal of GPUs right now.
At ISC this week, Super Micro is also showing off the SuperWorkstation 7046GT-TRF, a 4U tower machine that supports four double-width Tesla GPUs with three more PCI-Express slots for peripherals. This Nehalem EP box has up to 96 GB of main memory and four PCI-Express slots that are eaten up by the GPUs. The remaining slots include two PCI-Express 2.0 x4 slots and a single PCI-Express 1.0 x4 slot.
You can see the full lineup of GPU-enabled servers and workstations from Super Micro here.
Super Micro is also trotting out its support of the six-core Istanbul Opteron processors from Advanced Micro Devices. With its SuperBlade SBA-7141M-T blade server, Super Micro can cram 1,400 Opteron cores and 7.68 TB of memory into a single 42U server rack. The Istanbul chips are now supported in the A+ line of 1U Twin and 2U Twin2 rack servers, which pack two or four Opteron-based server nodes (each with two sockets) into a single chassis. And for companies that want fatter Opteron nodes, Super Micro has certified Istanbul chips to run on its four-socket Opteron boxes too.
It may seem a bit odd that Super Micro hasn't paired the Tesla GPUs with the Opteron boxes, but considering the love that has not been lost between AMD since it acquired graphics chip maker ATI, it is not really surprising. While AMD does its own riff on GPUs for supercomputing, with its Radeon chips and Stream software development kit, it just doesn’t have the traction that the nVidia Tesla GPUs and their CUDA programming environment have. ®