Graphics, Hardware

NVIDIA Tesla K20 i.e. GK110 is 7.1 Billion Transistors, 300W TDP, 384-bit interface


During the keynote session, NVIDIA?s CEO Jen-Hsun Huang introduced three products i.e. technologies from the company – GPU virtualization for consumers and professionals dubbed VGX i.e. VDI (Virtual Desktop Infrastructure), GeForce GRID (GPU Streaming Service) and the new cards for computation.

NVIDIA Tesla K10 (GK104-based) and K20 (GK110-based)

The launched products are Tesla K10 and K20 GPU Computing boards i.e. GPGPU cards. As we previously announced, Tesla K10 is actually a severely modified GeForce GTX 690 without display outputs for 100% maximized cooling output, lowered clock for longer longevity and equally high quality components.

When compared to Fermi-based Tesla C2075, the K10 brings 3x single-precision floating performance, i.e. you can expect around 3.1TFLOPS, instead of 1030 GFLOPS offered by the C2075. Since K10 is based on GK104 core, don?t expect stellar double precision i.e. FP64 performance.

Memory bandwidth grew by only 1.8x, which is expected due to the fact that ECC is still supported by software. What the scientific community should appreciate is the use of 8GB of GDDR5 memory, i.e. 4GB per graphics processor.

The big daddy though is the Tesla K20 board. This product is based upon new GK110 silicon, a 7.1 billion transistor manufactured using 28nm process node. According to Jen-Hsun Huang, this is the largest number of transistors on a single piece of silicon ? Xilinx has the second place with a FPGA chip with a bit over six billion transistors. Third one is a GPU from AMD, the 4.31 billion transistor monster, and then we get back to GK104 with 3.1 billion transistors. Top 5 are rounded with a 10-cor Intel Westmere processor, a 2.6 billion transistor chip manufactured in 32nm technology.

GK110 i.e. Tesla K20 brings over 1.5TFLOPS of double-precision FP performance, i.e. 3x more than Tesla C2075 with its 515 GFLOPS. The GK110 GPU brings an unannounced number of cores, 384-bit memory interface and an undefined amount of onboard memory. The memory combinations possible are 6GB, 12GB or 24GB. Jen-Hsun mentioned that the amount of memory depends on the DRAM industry. We expect to see the 12GB version, though.

All that performance requires a certain amount of power, and Tesla K20 requires a single 8-pin and 6-pin power for a grand total of 300 Watts.

Tesla K10 is available today, while the K20 will become available in the fourth quarter. As you might expect, there is no word on when will GK110 come life as a Quadro or a GeForce card.