TORswitch

High-performance, open, and enterprise-ready — a 32-port 100G switch optimized for AI, storage, and cloud networking.

AI fabrics are not general cloud networks. A training job produces a small number of enormous, synchronised, all-to-all bursts, and because every rank waits for the slowest, a local hotspot becomes cluster-wide idle time rather than graceful degradation.

TORswitch places custom FPGA logic in the data path. That is a deliberate trade rather than a component choice: a fixed-function ASIC commits to its congestion control and path selection behaviour at tape-out, while the algorithms in question are still actively changing. Programmable logic can follow them on installed hardware.

Specific accelerated functions, measured figures, and the platforms they were validated against are shared under evaluation — ask and we will tell you what it does and what it does not.

• High-speed disaggregated memory bank
• Custom FPGA acceleration
• Low-latency interconnect

Also available on Pantheon for GPUs, turbines, networking and memory.