Convequity on X
Strange short report on Cerebras $CBRS. Wafer-scale power and thermal constraints are longstanding engineering challenges. They can force compromises on clock speed versus a smaller die. That doesn't mean unacceptable overall performance. CS-4 is getting more out of WSE-3T, with the same core count and SRAM capacity as WSE-3, through higher clocks and improved I/O. Cerebras attributes the higher clocks to better power delivery. My read is that earlier systems left performance on the table to stay within power and thermal limits. Directionally, Cerebras looks to be working through those constraints better than I expected. Higher clocks alone don't answer the lifespan allegation. I’d still want to know how representative those failures are of current systems. I also give weight to Oracle, OpenAI and AWS choosing to work with Cerebras.
