Chicago (IL) – Nvidia’s Tesla cards appear to be gaining more traction lately with Amax being the latest company to announce a Tesla supercomputer product. The company promises close to 87 TFlops from 21 Tesla nodes.
Amax’ cluster builds on Nvidia’s 1U Tesla servers containing four Tesla units with 240 processing cores each. The company offers entry level configurations with four nodes running four 1U systems with 16 Tesla cards for $79,000.
There was no pricing information on the high-end 20 node solution offer 86.94 TFlops, but we expect such a system to easily top $400,000. Also, the 87 TFlop number refers to single-precision computations only. In double precision mode, the performance drops to 7.2 TFlops.
Amax said that the Tesla systems can be equipped with up to 336 dedicated GPU memory and a bandwidth of 8.6 GB/s. Compared to a traditional CPU system offering roughly the same performance, the 20 node system costs 17 times less, consumes 21 times less power and 60% less space. Of course, a Tesla system alone does not make much sense for most supercomputing environments. Many applications require vastly more memory than just 336 GB.