Rosetta: The Engine Behind Cray’s Slingshot Exascale-Era Interconnect

Re:Rosetta: The Engine Behind Cray’s Slingshot Exascale-Era ... by wmf on Wednesday May 06, 2020 @12:37AM (#23087940)

Supercomputers are incrementally less custom over time. Slingshot is a good example of this, reusing many parts of the Ethernet protocol and even parts of a commodity Broadcom chip design. The rest of the Cray Shasta system is built with off-the-shelf CPUs and GPUs (although custom HPC GPUs are coming from AMD and Intel).

Clearly the customers (mostly governments) think it's worth it. They also buy commodity clusters in addition to supercomputers so they're very aware of the differences. The biggest difference is the network and that's why Cray is developing that part themselves; some simulations are network-bound and using a commodity network it would either cost more than a supercomputer (due to lower efficiency) or it would never reach the desired performance at all (due to poor scaling).

[ Parent | on HN ]

Re:Rosetta: The Engine Behind Cray’s Slingshot Exascale-Era ... by gnufx on Wednesday May 06, 2020 @05:40AM (#23089276)

It's odd to class slingshot/dragonfly as less custom for HPC than even infiniband (and Mellanox is all-out on HPC-specific "offload"). Other interconnects designed for HPC include Bull's BXI, Fujitsu's Tofu, and European stuff, though they're even as common as Cray's, and I don't know how they compare.

[ Parent | on HN ]

Re:Rosetta: The Engine Behind Cray’s Slingshot Exascale-Era ... by wmf on Wednesday May 06, 2020 @01:10PM (#23093575)

I agree that Infiniband is off-the-shelf technology at this point, but previous Cray interconnects like SeaStar/Gemini/Aries were 100% custom and had no interop with standards.

[ Parent | on HN ]

Re:Rosetta: The Engine Behind Cray’s Slingshot Exascale-Era ... by gnufx on Wednesday May 06, 2020 @02:46PM (#23094790)

Right, if that's all custom means.

[ Parent | on HN ]