It’s becoming tough to sort out competing AI performance claims and to keep track of new AI initiatives. Perhaps the new MLPerf benchmark will eventually help. Yesterday, Nvidia claimed its Volta tensor core architecture achieved the fastest performance on ResNet-50 at the chip, node (DGX-1) and cloud level. On the same day, Microsoft previewed Project Brainwave which will leverage Intel FPGA technology to accelerate AI calculations.
Here’s a brief excerpt from yesterday’s Nvidia Developer Blog. “Our results demonstrate that:
- A single V100 Tensor Core GPU achieves 1,075 images/second when training ResNet-50, a 4x performance increase compared to the previous generation Pascal GPU.
- A single DGX-1 server powered by eight Tensor Core V100s achieves 7,850 images/second, almost 2x the 4,200 images/second from a year ago on the same system.
- A single AWS P3 cloud instance powered by eight Tensor Core V100s can train ResNet-50 in less than three hours, 3x faster than a TPU instance.
Microsoft previewed Brainwave – a new hardware architecture – at its developer conference (Microsoft Build) being held in Seattle this week. “We’re making real-time AI available to customers both on the cloud and on the edge,” said Doug Burger, in a Microsoft blog. Burger is a distinguished engineer at Microsoft who leads the group “that pioneered the idea of using FPGAs for AI work.”
In the same blog, Mark Russinovich, chief technical officer for Microsoft’s Azure cloud computing platform, said the preview of Project Brainwave marks the start of Microsoft’s efforts to bring the power of FPGAs to customers for a variety of purposes. “I think this is a first step in making the FPGAs more of a general-purpose platform for customers,” Russinovich said. Intel also issued a short article on Brainwave.
Link to Nvidia blog: https://devblogs.nvidia.com/tensor-core-ai-performance-milestones/
Link to Microsoft blog: https://blogs.microsoft.com/ai/build-2018-project-brainwave/?utm_source=press&utm_campaign=75592