Cerebras Watch posted an update
Cerebras chief executive Andrew Feldman says the company competes with Nvidia and Google by taking a different route: building processors across an entire wafer rather than linking multiple chips together.
Why it mattersFeldman says that design reduces data movement between chips and lets Cerebras use data centres of various sizes instead of relying on a giant campus. The Seoul Economic Daily reports that a Cerebras chip is about 50 times larger than a GPU, while still running in a relatively smaller data centre. It is a useful glimpse of the architectural trade-off behind an AI-chip challenger: not simply making the usual approach bigger, but changing how the processor is assembled. The report was translated from Korean, so its translated quotations may not match the original wording exactly.
Discuss: Is a wafer-scale design a convincing alternative to linking many smaller chips, or will scale and software ecosystems matter more?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.