Published in Graphics

Nvidia officially launches Tesla K20 and K20X GPUs

by on13 November 2012

nvidia tesla logo

GK110 finally arrives

We heard about this one back at Nvidia's GPU Technology Conference and since it was recently used in the Oak Ridge National Laboratory's Titan supercomputer, Nvidia has now finally and officially launched its new Tesla K20 GPU family based on the elusive GK110 GPU.

Back at GTC 2012, Nvidia released a short announcement of the K20 and shed some details regarding its power, but has kept all those juicy GK110 specs to itself. Now, as it is officially announced, it is clear that Tesla K20 lineup will consist of two products, the Tesla K20 and the Tesla K20X.

Both of the new Tesla graphics cards are based on the same GK110 Kepler GPU and it is now clear that GK110 features 15 SMX clusters with six memory controllers and 1.5MB of L2 cache. The Tesla K20X, a more powerful one, uses 14 active SMXs with all six memory controllers and full 1.5MB of L2 cache. This adds up to a total of 2688 CUDA cores. The Tesla K20X works at 732MHz for the GPU and features 6GB of GDDR5 memory clocked at 5.2GHz and paired up with a 384-bit memory interface. All these specs will offer 3.95 TFLOPS of single- and 1.31 TFLOPS of double-precision floating-point performance and it is the same GPU that is a part of the Titan supercomputer.

The less powerful, Tesla K20 has one less SMX, five memory controllers and 1.25MB of L2 cache. That adds up to 2496 CUDA cores and it works at 706MHz for the GPU and 5.2GHz for 5GB of GDDR5 memory paired up with a 320-bit memory interface. It will provide up to 3.52 TFLOPS of single- and 1.17 TFLOPS of double-precision floating-point performance.

The TDP is set at 235W for the Tesla K20X and 225W for the K20. Although the performance of both Nvidia's new Tesla's are away from AMD's FirePro S10000 dual-GPU beast, the TDP is also nowhere near the incredible 375W.

When compared to the previous Fermi based Tesla M2090, the new Tesla K20X raises the actual architecture efficiency from 65 percent to up to 93 percent. This was achieved via new Hyper-Q, Dynamic Parallelism, new ECC algorithm and OpenACC, MPI and other HPC libraries and technologies.

Of course, Tesla family was never cheap so, although there is no official price, the estimated price is set at around US $3199 for the K20 and somewhere over that number for the more powerful K20X.

You can check out an in-depth analysis over at

nvidia teslaK20 1

Last modified on 13 November 2012
Rate this item
(0 votes)

Read more about: