AWS Pushes the Limits of GPU Computing with P3 Instances

Amazon Web Services (AWS), just over a month after launching its largest-capacity EC2 instance in April, reached another milestone last week with the launch of its latest GPU-optimized instance group, the P3.
The new P3 instances are available in three configurations from AWS’ Northern Virginia and Oregon, Ireland, and Tokyo regions. They are optimized for highly compute-intensive and advanced workloads. Machine learning, genomics and high-performance computing (HPC), cryptography and financial analysis are just a few examples of the use cases.
The P3 instances are run on Nvidia’s Tesla V100 GPUs. These GPUs are based on Nvidia’s Volta technology. Nvidia claims that each Volta GPU can run 100 CPUs. This makes it ideal for complex AI and deep-learning workloads.
AWS promises up to 14x the processing power of its new P3 instances with the new P3 instances. These instances run on older Tesla K80 GPUs.
Jeff Barr, AWS evangelist, explained in a blog post that each NVIDIA GPU is equipped with 5,120 CUDA cores, another 640 Tensor Cores, and can deliver up 125 TFLOPS mixed-precision floating points, 15.7 TFLOPS single-precision floating points, and 7.8 TFLOPS double-precision floating. The GPUs can be connected to each other via NVIDIA NVLink 2.0, which has a maximum data rate of 300 GBps. This allows the GPUs exchange intermediate results and other data at high speeds without the need to move it through the CPU/PCI-Express fabric.
The table below lists the three configurations of P3 instances.
[Click on the image to see a larger view.] Source: AWS/Jeff Barr According to Barr, the largest configuration, the P3.x16xlarge is 781,000x faster at performing vector operations that the Cray-1 supercomputer. However, that system is more than four decades old. He said that it is harder to compare modern supercomputers to the P3. “You can think of the p3.x16xlarge as a step and repeat component of a supercomputer you can launch on an as-needed basis.”
Cray recently announced a partnership to bring Cray’s supercomputing capabilities into Microsoft’s Azure cloud platform. AWS is Microsoft’s cloud rival.
Nvidia announced last week that P3 instance users can take advantage of Nvidia GPU Cloud (NGC), a container registry for AI developers. Nvidia claims that the NGC will enable deep learning development by providing no-cost access and fully optimized deep learning software.
You can find more information about the NGC here.

Author: Victoria