This week in Frankfurt, Germany, at the annual International Supercomputing Conference (ISC), NVIDIA is expected to announce the PCI Express Tesla P100 card.
According to HPC Wire, NVIDIA is mounting its efforts to replace the CPU as the top of the processor list for demanding datacenter workloads. One element of this strategy is introduction at ISC 2016 of a PCIe-based version of its new Tesla P100 card. ISC is one of the two major annual supercomputing conferences used for high performance processor announcements.
Since introduction of DGX-1 platform and Pascal chip last April, NVIDIA's strategy has been shifting. The company argues that the mainstreaming of AI/deep learning approaches in scientific computing, better performance scaling on GPUs and the large base of CUDA code are combining to lift GPUs to preeminent role from its previous supportive role.
According to AnandTech, NVIDIA first introduced the Tesla P100 card at their 2016 GPU Technology Conference back in April. The Tesla P100 graphics accelerator is based on NVIDIA's new 16nm GP100 GPU and Pascal architecture.
The new graphics card is a significant improvement from the Tesla K/M series and their respective 28nm Kepler/Maxwell GPUs. Tesla P100 introduces a number of new features including instruction level preemptive context switching, larger caches and double speed (packed) FP16 compute, besides being a bigger-still GPU.
The initial version of the P100 announced in April was NVIDIA's highest performing version, shipping with 56 of 60 SMs enabled, based on a 300W board using NVIDIA's new mezzanine connector.
The card's mezzanine connector was necessary to facilitate NVIDIA's high-speed point-to-point NVLink bus and marked a radical departure from traditional NVIDIA Tesla card designs. However not everyone wants to build systems specifically for the mezzanine connector or needs the features of NVLink. For this reason, NVIDIA announced now the PCIe version of the card in the Tesla P100 lineup.
NVIDIA will offer two versions of the PCIe Tesla P100 that will be available in Q4 of this year. Their pricing has yet to be announced.
