NVIDIA's 7.1 Billion Transistor Coprocessor
Since the release of NVIDIA’s GK104-based Tesla K10 coprocessor, we’ve heard surprisingly little from the firm’s high performance computing division. Despite this, NVIDIA has been working on a new GK110-based unit it promised earlier in the year, the Tesla K20. Cray has in fact already upgraded its Jaguar XK7 supercomputer (now known as “Titan”) with a jaw dropping 18,688 16-core AMD Opteron 6274 processors and Tesla K20 units, boosting the performance by a factor of ten, while reducing power consumption by a factor of five. The Tesla K20 features 2,496 32-bit and 832 64-bit floating-point units (FPUs), six times the number of 64-bit units than the Tesla K10 had, and 5GB of GDDR5 memory over a 320-bit memory bus, however this is reduced from the Tesla K10’s 8GB memory across two 256-bit memory buses.
Although the Tesla K20 provides just 4,577GFLOPS of single precision power, which is less than its Tesla K10 counterpart, it also offers 1,173GFLOPS of double precision power, compared to the Tesla K10’s 190GFLOPS. Despite this, NVIDIA has managed to reduce the power consumption by 10% from 250W to 225W. As of now, the Tesla K20 is not listed on NVIDIA’s website (only marketing the Tesla K10 with superior single precision power), with the Tesla K20’s 7.1 billion transistors expected to be the culprit.