It is powered by a GK180 processor and has 12 GB GDDR5 VRAM

Oct 7, 2013 06:25 GMT  ·  By

You'd think we'd have heard about the GK180 graphics processing unit from NVIDIA by now, but we haven't, and neither has anyone else from what we can tell. Nevertheless, if NVIDIA's new Tesla supercomputer GPU accelerator uses it, we can't do anything but accept its existence.

NVIDIA has prepared a new GPU accelerator made for parallel processing, the Tesla K40 which will teach Tesla K20/K20X a thing or two.

The PCI Express card has twice the memory capacity of the Tesla K20X, 12 GB of GDDR5 versus 6 GB.

The board also happens to boast a maximum single-precision floating point performance of 4 TFLOP/s (TERA floating point operations per second). Meanwhile, the double precision performance is of 1.4 TFLOP/s.

Moving on, the memory bandwidth is of 288 GB/s and the TDP (thermal design power) is of 235W (as an add-on card) or 245W (as an SXM, more on the lines of a compact card like the MXM used in notebooks).

For the sake of comparison, the 235W Tesla K20X, based on the GK110, has 3.95 / 2.9 TFLOP/s single-/double-precision performance, 250 GB/s bandwidth and 6 GB GDDR5.

That's not all though. The Tesla K40 also has more CUDA cores in the GPU (2,880 vs. 2,688) and dynamic overclocking technology (works on ANSYS and AMBER workloads).

Finally, NVIDIA's Tesla K40 Atlas GPU compute accelerator works on the PCI Express 3.0 interface, while the Tesla K20X only has PCI Express 2.0 support.

All in all, we'd say that the GK180 and the card based on it are definitely superior to the GL110 and Tesla K20X, although some might say that the difference isn't as high as the difference in GPU numbers would imply.

After all, GK110 is well beyond GK104, so a GK180 would, presumably, be even more ahead of GK110. Alas, NVIDIA is still relegated to the 28nm manufacturing process technology, so it doesn't exactly have much room to move in.

Photo Gallery (2 Images)

NVIDIA Tesla K20
Tesla K40-K20X comparison
Open gallery