IBM will soon be offering NVIDIA’s latest P100 GPU on its Bluemix cloud platform. The new option is aimed at users running AI and other types of performance-demanding analytics workloads.
In the announcement, IBM declares it is “the first major global cloud provider to make the NVIDIA Tesla P100 GPU accelerator available on the cloud.” Tencent and Nimbix also offer the P100 in their respective clouds, but they don’t quite make the cut as “major global” providers. Amazon EC2 and Microsoft Azure are currently offering the older NVIDIA K80 GPUs in their respective clouds, as is Google, which just recently rolled out its GPU cloud service. No doubt, all of them will upgrade to the P100 devices at some point.
It looks like the IBM will be using the PCIe-based version of the P100, rather than the slightly more powerful NVLink-optimized version. Nevertheless, the PCIe P100 is a good deal more performant than the previous-generation K80. In double precision performance, the P100 can deliver 4.7 double teraflops versus 2.9 teraflops for the K80. For AI workloads, which rely on lower precision math, the P100 also outruns the K80: 9.3 to 8.7 teraflops at single precision. In addition, the P100 delivers 18.7 teraflops worth of half precision processing, a capability missing in previous GPUs.
Maybe even more important than the additional flops, the new GPU is also equipped with second-generation High Bandwidth Memory (HBM2), which provides much high bandwidth than conventional GDDR5 memory – in the case of the P100, up to 732 GB/second. For memory-bound workloads, this represents a significant advantage, especially where the application’s data can fit into the 16GB footprint of the HBM2 module.
Although this is a cloud offering, the Bluemix bare metal set-up is not a virtualized platform, so there’s no performance lost through any hypervisor layer. According to IBM, you can specify configurations with up to two P100 devices per server. All are equipped with two Intel Xeon processors, which act as the hosts for the GPUs.
Pricing has not been established, although the current K80-accelerate bare metal servers start at $5.30 per hour or $1,359 per month, so one would expect the P100 servers to cost at least as much. For the time being, IBM will continue to offer the K80, K2 and M60 GPUs, in its cloud options. The P100 GPUs will be available on Bluemix servers later this month.