Nvidia launches dual-GPU Tesla K80 accelerator

November 18, 2014 // 12:13 p.m.

Tags: #amd #firepro #gk210 #gtx-titan #hpc #kepler #nvidia #supercomputer #tesla #tesla-k40 #tesla-k80

Nvidia has officially unveiled its latest Tesla GPGPU-driven accelerator board, a dual-GPU beast with 24GB of GDDR5 memory and claimed single-precision compute performance of 8.74 teraflops: the Tesla K80.

Aimed at the high-performance computing market, Nvidia's top-end Tesla cards can frequently be found powering some of the fastest supercomputers in the world. Based on the same GPU designs as the company's consumer graphics cards, the Tesla family dispenses with graphics outputs in favour of using the massive number of parallel processors to perform calculations required for everything from finding new sources of oil to simulating the after effects of a nuclear blast.

Nvidia's previous top-end model was the Tesla K40, launched in November of last year, which boasted 12GB of GDDR5 memory connected to 2,880 CUDA processing cores. With a claimed single-precision compute performance of 4.29 teraflops under base clocks and 5 under GPU Boost conditions, the Tesla K40 was the king of HPC until rival AMD announced its FirePro S9150 in August this year with a whopping 5.07 teraflops of compute performance and 16GB of GDDR5 memory.

Nvidia's answer should surprise nobody: glueing two Tesla K40s together to form the Tesla K80. Using the same design techniques as its dual-GPU graphics cards, the Tesla K80 combines two GK210 - newer than the GK110Bs of the original K40 - GPUs to form a single dual-slot card capable of pumping out up to 8.74 teraflops of single-precision and 2.91 teraflops of double-precision compute performance under GPU Boost. Each GK210 chip has access to its own bank of 12GB GDDR5 memory, for a total of 24GB, but not everything has doubled: rather than the 2,880 CUDA cores of the Tesla K40 each chip in the K80 has just 2,496 cores enabled for a combined total of 4,992. Despite this, architecture improvements - including a doubling of L1 cache on each core cluster and a larger 512KB register - although, as the performance indicates, the dual-GPU design does not provide a straight doubling of raw performance.

While previous Tesla-centric chips have found their way to the consumer market in the form of Nvidia's GeForce GTX Titan family, the new-design GK210s powering the Tesla K80 are likely to remain an HPC-market exclusive. More details are available on the official website.
Discuss this in the forums