Published in Graphics

Nvidia Tesla K80 boasts two Kepler GK210 GPUs

by on18 November 2014

World's fastest data accelerator

Nvidia’s Tesla takes care of the big data, server/compute market and isn’t necessarily aligned with the processors Nvidia has offers professional graphics and the gaming market.

Nvidia just announced two new Tesla data accelerator cards, the K80 based on two GPUs and the K40 with a single one. It was interesting to point out that both are based on the previous Kepler architecture, while the most current GM204 Geforce GTX 980 GPU is based on the Maxwell architecture.

Meet the Tesla K80


The Kepler based K80 has two GK210 GPUs, 4992 cores (2496 per GPU) which allows the card a peak performance of 1.87 TFLOPS at the base clock, or 2.91 TFLOPS with GPU Boost. The GPU boost clock will only be possible if the card thermals allow it. The K80 peak single precision performance hits 5.6 TFLOPS at the base speed or 8.74 TFLOPS with GPU Boost. Maximum card bandwidth with ECC off is 480 GB/s (240 GB/s per GPU) and the K80 comes with 24GB memory or 12GB per chip.

K80nvidia

The slower, single GPU K40 has a single GK110B with 2880 steam processors and 12GB GDDR5 memory on a 384-bit bus. The card achieves 1.43 TFLOPS at base speed or 1.66 TFLOPS with GPU Boost (double precision), or 4.29 TFLOPS at base speed or round 5 TFLOPS with GPU Boost in single precision calculations. The maximum bandwidth of the card is 288 GB/s. it is clear that GK110B is a derivative of the GK110 we saw in Tesla K20 and K20X compute cards.

The K20X, based on GK110 28nm GPU ships with 2688 stream processors at 732 MHz clock and 6GB RAM, while the K20 comes with 2880 stream processors 732 MHz core clock and 12GB memory.

Let's not forget that the K80 has a 300W TDP and is passively cooled, while the K40 has is available in passive and active flavours, with a TDP of 235W.

AMD has great compute cards of its own


AMD has a competitor in the form of the AMD FirePro S9150 Server GPU and this one is based on the Hawaii GPU. The card will offer 2.53 TFLOPS peak double precision floating point at 1/2 rate.

Just for comparison, Nvidia offers 1/3 of the double precision rate. S9150 single precession peak rate is 5.07 TFLOPS and it comes with 16GB GDDR5 and 512-bit memory. The S9150 memory bus enables 320GB/s bandwidth and outperforms the K40's 288 GB/s. AMD’s single GPU based S9150 also has a 235W maximal TPD and it clearly outperforms K40, at least on paper. The downside of AMD card is that you can only do ECC externally.

AMD also claims that you can use up to 10 AMD FirePro S9150 server GPUs in a single system which will definitely give you a lot of power.

Rate this item
(0 votes)

Read more about: