Revision as of 13:31, 16 July 2018

Lake Crest is a neural processor microarchitecture designed by Nervana.

Process Technology

Lake Crest is fabricated on TSMC's 28 nm process.

Architecture

Lake Crest was designed from the ground up for deep learning. The architecture itself is a tensor-based architecture, meaning it's optimized for blocks of compute instead of operating on scalars (as would a traditional Intel CPU would).

• Tensor-based architecture
  • Nervana Engine
• Flexpoint number format
• No caches
  • Software explicitly manages all on-chip memory
• HBM2 memory
  • 32 GiB of in-package memory
  • 8 Tbit/s bandwidth
• 12 x Inter-Chip Links (ICL)
  • bi-directional high-bandwidth direct chip-to-chip interconnect
  • 100 GB/s (1,200 GB/s aggregate)

This list is incomplete; you can help by expanding it.

Block Diagram

Chip

Processing Cluster

This section is empty; you can help add the missing info by editing this page.

Memory Hierarchy

• 32 GiB on-package HBM2
  • 1 TiB/s

Die

Nervana stated that Lake Crest is "near-reticle size" implying the die size is likely around the 650-750 mm².

• TSMC 28 nm process
• 650-750 mm² die size

Additional Shots

• 

  Lake Crest Accelerator PCIe card (internal view)

Bibliography

• Rao, N. (2016, November). Pathfinding and Hardware Deep Dive. 2016 AI Day, San Francisco.
• Rao, N. (2018, May). Keynote presentation. 2018 AI DevCon, San Francisco.