Sam Altman-backed startup Volantis is building an SI accelerator where light instead of wires links memory to the compute chip
Prepared by Sintelligent with the help of SI. Publication time restored from the build log. . Updated: .
If memory stops being the bottleneck, models dozens of times larger than today's will become possible.
The San Francisco company (founded in 2022) is betting on an optical substrate and licenses the other blocks. It expects to fit more than 10 TB of memory in a single package with a transfer rate of up to 240 TB/s — by its estimate, enough to serve a 20 trillion-parameter model at 10 thousand tokens per second per user.
Sources 1