Original URL: http://www.theregister.co.uk/2013/04/15/clown_card_of_cpus/

63 TRILLION maths ops a second - in 5 inches? Mm, show me

3U rack box is a $70,000 clown car of GPUs

By Dan Olds, OrionX

Posted in HPC, 15th April 2013 09:38 GMT

If you need lots of compute power, you’re either already using GPU accelerators or taking a close look at them. And if you’re serious about tapping into the full processing power of the GPU, perhaps this gear will fit the bill.

Dubbed the ioMillennia, it’s a 3U, PCIe Gen3 switch from One Stop Systems that can handle sixteen full-size GPU cards. We glimpsed this behemoth when walking the floor at GTC 2013 and were attracted like moths to the 42.3 billion candle-power Luxor Sky Beam. (Or, more accurately, like me to a doughnut display case.)

One Stop System sales veep Jim Ison walked us through the box (as seen in the video here), discussing both technical specs and customer usage. It packs 16 hot and power-hungry GPUs into a cramped 3U space. And it does take a little bit of power to run the high-performance GPUs – about 6,000 watts provided by three 3,000-watt hot-swappable power supplies. (One is a redundant backup.)

At a retail price of about $15,000 for the base unit, you get a lot of bang for your buck. Let’s do some maths. A single Nvidia K20x has a single precision peak of 3.95 teraflop/s, and 1.31 teraflop/s at double precision.

Multiply this by 16 and you get a potential max of at least 63 TFLOP/s single precision and at least 20 TFLOP/s double precision per 3U rack mount unit. The ioMillennia costs around $18,000 with all four host adapters, and 16 Kepler K20s will cost you an additional $51,200 (16 times $3,200). Let’s call it an even $70,000 purchase price. (I’m assuming you already have some fast host systems hanging around.)

So, if you have the code to push all 16 Keplers to their peak performance, your price performance would be a $1.11 per gigaflop/s single precision or $3.34 per gigaflop/s double precision.

Let's put these numbers into context: I wrote an article just over a year ago comparing LINPACK price performance on several supercomputers versus common workstations and home PCs. Getting accurate cost figures on supercomputers isn’t easy; these things don’t come in shrink-wrapped boxes with price tags. However, I managed to find some ballpark numbers. Here’s what that chart looks like with One Stop’s ioMillennia added in:

Table comparing hardware: HTML version to come

The stats are in (click to enlarge)

At waaayyyy (technical term, meaning ‘a lot’) under a penny per megaflop/s, the ioMillennia handily beats every system on our chart – even thrashing my wife’s generic desktop that used to lead the pack at just over 4 cents per MFLOP/s.

Keep in mind that the numbers for the ioMillennia were arrived at by simply adding up the theoretical peaks for all 16 GPUs. It’s not the result of a benchmark or any real-world workload. There also isn’t any host system or other infrastructure cost in that number. However, it’s still orders of magnitude less expensive than other systems in terms of price/performance.

Speaking of performance, the ioMillennia host adapter allows GPUs to directly exchange data without running the request back through the CPU. Nvidia introduced RDMA (Remote Direct Memory Access) with the Kepler launch last year, but we haven’t seen much real-world data on how well it performs until some sessions at GTC 2013. The advantages can be profound, reducing latency by 5x on large transfers (2MB) and almost 16x on small (16KB) chunks of data. Throughput also rises dramatically. (See El Reg's Timothy Prickett Morgan's discussion of RDMA here.)

This product is already available directly from One Stop Systems and should be appearing soon in at least a few major vendor price books. ®