💡Networking Vendors Race to Innovate with New Interconnects and Switches
New tech could slash GPU communication overheads
TL;DR
Rival networking vendors are launching new interconnects and switches to compete with Nvidia's NVLink and AMD's UALink. Cornelis and Delos Data are leading the charge with technologies like SHARP and Ultra Ethernet, promising faster and more efficient data transfer.
Rival networking vendors are scrambling to bring alternative interconnects and switches to market, aiming to compete with Nvidia's NVLink and AMD's UALink. Cornelis' Active Compute Fabric (ACF) and Delos Data's near-packaged optics (NPO) tech are leading the charge, promising to slash GPU communication overheads and boost performance. The key technologies behind these innovations include SHARP, which allows for in-network collectives to be offloaded to switch ASICs, freeing up GPU compute. Delos Data's I/O chiplets offer over 30 Tbps of aggregate bandwidth and are protocol agnostic, integrating with optics engines from leading optics suppliers.
Key Points
Cornelis' Active Compute Fabric (ACF) seeks to establish an open architecture for scale up and scale out networking, expanding on technologies like UALink and Ethernet for Scale-Up Networking.
SHARP allows for in-network collectives to be offloaded to switch ASICs, freeing up GPU compute and cutting down on communication overheads, with Nvidia's NVSwitch ASICs supporting it.
Delos Data's I/O chiplets offer over 30 Tbps of aggregate bandwidth and are protocol agnostic, integrating with optics engines from leading optics suppliers.
Delos Data's near-packaged optics (NPO) tech integrates a 10-plus Tbps data interface with optics engines, offering a series of new network reference designs.
Delos Data's 400-plus Gbps NIC is protocol agnostic and builds on the company's existing compute reference design and Nonstop AI software platform.
Why It Matters
If you're running large-scale AI workloads on GPUs, these new interconnects and switches could significantly reduce communication overheads and boost performance. For example, SHARP offloads in-network collectives to switch ASICs, freeing up GPU compute. Delos Data's I/O chiplets offer over 30 Tbps of aggregate bandwidth, making them a strong contender for high-performance computing applications.
Comments
Be the first to comment
Enjoyed this article?
Get it daily. 7am. Free. Reads in 5 minutes.
Join 3,482 builders reading daily.