As footprint of applications utilizing AI and complexity of neural networks used increases , FB is implementing customized HW using ASICs for Machine learning applications.
In this talk, we will discuss the compute platform being developed currently for Inference applications. We will go over the accelerator module form factors (M.2 and dual M.2) and specifications developed at Facebook to support ASICs with different size and power envelopes. The talk will also cover the Hardware platform design (Yosemite V2.50 and Glacier point V2) to support these applications. We will discuss the Glow SW stack that will run on these ASICs and the debug capabilities built into the platform.
The system is developed on the Yosemite 2.50 (Yv2.50) platform which is a modification of the existing Yosemite V2 single socket platform. Yv2.50 is a 2 host system, which supports higher system power and host network bandwidth (50G/host) to improve overall effective system performance. The network cards used have architecture improvements to optimize performance in multihost environments.
The modules are housed in a Glacier point V2 carrier card which are in turn connected to the Twinlake (Intel Skylake) hosts on the Yosemite V2.50 platform through a PCIE switch. Each Glacier point V2 carrier card can house up to 6 dual M.2 accelerator modules or upto 12 M.2 modules.
Glow accepts a computation graph from deep learning frameworks, such as PyTorch, and generates highly optimized code for machine learning accelerators. We will discuss the integration and validation of Glow frameworks on these systems.
