Thursday, November 8, 2007

Supercomputer On A Graphics Card (1 TFlops)

Not so long ago, who would have thought that a desktop computer would be able to process "trillion floating point operations per second (TeraFlops or TFlops). Most people not up-to-date with technology would say its impossible for a normal desktop computer to do it, but it isn't anymore!! Really..and I'm not marketing for any company...

Today AMD announced the FireStream 9170, a General Purpose Graphics Processing Unit (GPGPU) running at 775Mhz and performs 497 GFlops. That's about 0.5 TFlops!! The FireStream is based on the RV670 chip and has 320 double-precision floating-point (FP) units. The RV670 will also be available in desktop variants for gamers and will be known as the HD 3850 and 3870 as mentioned by me here.

But it is still just 0.5TFlops... right?? Yes, but in about a few months time, AMD (actually ATI) will be releasing R680 which will have 2 of these chips working together. Add a few more clocks to the chip and we'll have a chip that can perform 1TFlops!! Aint that great??

Consider this: The world's fastest supercomputer i.e. IBM's BluGene/L scores 280.6 TFlops and costs a few billion dollars to build whereas one GPGPU will cost some thousand dollars. Do the math and you'll know its some million times cheaper with the GPGPU!! But the entire comparison I did is comparing apple to oranges. And the reason is because BlueGene/L uses general purpose CPUs which are easily programmed, whereas GPUs require special programming. I don't say that programming GPU is tough, it's just that we are habituated to program CPUs and hence we currently don't have the same type of tools that are used to program CPUs. But as the tools come up and mature GPGPUs are going to be the next milestone in supercomputing.

Intel also about a year back showed a 80 core processor do some networking to process as much as 1.81TFlops at 5.7Ghz... But its still got a lot more tweaking to do and the yield I think won't be very good at the moment!!

All of this means that in the very near future, by the use of GPGPUs, each one of us may be having a supercomputer built into our computers. But where is the work that these fast mean machines would be doing??


AsHisH Kulkarni said...

The latest would be the nvidia GeForce Titan, with a capacity of about 1.48TFLOPS per second !!!

mahasiswa teladan said...

hi..Im student from Informatics engineering, this article is very informative, thanks for sharing :)