Nvidia: What would you do with a petaflops super?
One freebie Kepler GPU down, 999 to go
Nvidia is being a tease, but your good idea could land you with a Telsa GPU coprocessor based on the future "Kepler" GPU chip.
Ahead of its GPU Technical Conference, which runs from May 13 through 17 in San Jose, the graphics chip and math motor maker has announced a competition called "What Would You Do With A Petaflops Supercomputer?" But don't get too excited. Nvidia is not going to give you a $10m petaflops system if you answer the question cleverly.
In announcing the contest, Sumit Gupta, senior product manager of the Tesla line at Nvidia, gave the required "how many laptops is a petaflops" metric (by Nvidia's math, it is around 20,000), and said that Nvidia would give the people who come up with the three best ideas for using a petaflops super early access to the Kepler GPU coprocessors that are slated for delivery later this year for server nodes. In its rules and regulations for the contest, Nvidia says these are worth about $3,000.
Nvidia does not provide list pricing on the Tesla M series GPU coprocessors because they are only available from server OEM partners, but this is consistent with what we have heard server makers are charging for the current Tesla 20 GPU coprocessors, which are based on the 448-core and 512-core "Fermi" GPUs. These GPUs deliver 515 gigaflops and 665 gigaflops of number-crunching power.
The "Kepler" GPU coprocessors are expected to weigh in at around one teraflops of double-precision floating point power, and are providing most of the flops in the "Titan" hybrid CPU-GPU machine going into Oak Ridge National Laboratory, which should hit around 10 petaflops, and the "Blue Waters" machine at the University of Illinois.
Nvidia has been pretty tight-lipped about the Kepler design for servers, but released some specs on the Kepler GPU variants used in desktop and laptop machines for graphics processing rather than number crunching. These Kepler1 GPUs have 1,536 cores and run at slightly more than 1GHz, and are designed to have three times the performance per watt of the Fermi GPUs they replace.
Nvidia has not released single-precision or double-precision floating point numbers for the Kepler1s, and that is probably because it doesn't want HPC buyers to jump to any conclusions about the performance of the Kepler2 GPUs that will be used as server math coprocessors, not for graphics.
El Reg figures that Nvidia wants to break one teraflops of double precision performance while getting the wattage down below the 225 watts that a Tesla 20-class card eats. If you can pack four of these into a server chassis without melting it, you could build a fairly compact parallel super with lots of math muscle.
Provided you have the CUDA chops to port your code.
Anyway, Nvidia is giving early access to the Kepler2 GPU coprocessor to those who give it the best ideas. If the three pool them together, then they only need another 997 or so to get to that 1 petaflops performance level. ®