Fabric EDM Optimization (to 10+ GB/s per direction):

[SeanNijjar]

Stability

• Resolve ND (but mostly reproducible) hang in some multi-chip tests with EDM fabric #17432
  • likely culprit identified but fix makes more sense to apply with EDM Fabric (sender side): decouple all pointers (wrptr, ack_ptr_from_remote_receiver, completion_ptr_from_remote_receiver, wrptr_from_sender) #17427
  • Tue. Feb. 4: latest update is that this issue was resolved after a fix to one of the worker->EDM fabric APIs that was accidentally missing a writes_flushed for a flush barrier mode. Item left open until fix confirmed in main with all latest EDM and all-gather changes
• Add 1D fabric user configurable context switch frequency with pybindings #18222

Micro Architecture

• EDM Fabric: Decouple worker acks from EDM acks. #17425
• EDM Fabric (sender side): decouple all pointers (wrptr, ack_ptr_from_remote_receiver, completion_ptr_from_remote_receiver, wrptr_from_sender) #17427
• EDM Fabric (receiver side): add write transaction IDs and decouple all pointers (wrptr, ack_ptr_to_sender, completion_ptr_to_sender) #17428
  • PR: Enable transaction IDs on the write side (receiver channel) of EDM fabric kernel #17631
• EDM Fabric: Update outbound <-> inbound flow control to pointer exchange via eth reg writes (to stream registers) #17430
  • PR: Update EDM Fabric flow control between sender and receiver channels over Ethernet to use stream registers and eth reg write #17577

Low-Level and Code Implementation Optimizations

• EDM Fabric: Merge command type and noc command type in packet header. #17429
  • PR: optimize edm fabric packet header structure #17579
• EDM Fabric: Enable write-flush bits to atomic write operations, default all other writes to non-blocking on receiver #17431
• Migrate to new eth-word-size based eth_send_packet API and track packet sizes internally in terms of eth words (rather than 4B words or bytes) #17686
  • EDM Packet Header Optimization: Simplify packet size storage and access (+other misc low-level optimizations) #17724
• Update EDM fabric to use power-of-two buffer slot count on sender and receiver channels to optimize pointer management
• Evaluate and migrate to word level flow control in EDM instead of buffer-slot level flow control
• Re-evaluate main EDM fabric control loop for low-level code optimizations (including value caching, auditing use of volatiles, etc.)
• Run EDM fabric kernel in Erisc IRAM
• Evaluate EDM fabric kernel compile with -O2 and -O3; currently erisc kernels are compiled with -Os (which typically leads to less performant binaries than the other optimization levels)
• 1D Fabric: use stream registers for flow control between EDMs on same chip (for packet forwarding path) #18098