2 comments

  • aabdi59 minutes ago
    5x throughput more than b4, same 10x tokens&#x2F;s as compared to standard.<p>looks like disagg&#x2F;wide ep is applicable to cerebras chips as well. kind of amazing.<p><a href="https:&#x2F;&#x2F;www.cerebras.ai&#x2F;blog&#x2F;how-we-deploy-trillion-parameter-frontier-models-on-cerebras" rel="nofollow">https:&#x2F;&#x2F;www.cerebras.ai&#x2F;blog&#x2F;how-we-deploy-trillion-paramete...</a>
  • augment_me1 hour ago
    Have they solved the issue yet where they need to keep running ghost tokens to keep the board hot to prevent it from cracking, which causes massive energy draw in idle, and reduces lifetime by 3 times the estimate?<p><a href="https:&#x2F;&#x2F;culperresearch.com&#x2F;wp-content&#x2F;uploads&#x2F;2026&#x2F;10&#x2F;Culper_CBRS_10-28-2026.pdf" rel="nofollow">https:&#x2F;&#x2F;culperresearch.com&#x2F;wp-content&#x2F;uploads&#x2F;2026&#x2F;10&#x2F;Culper...</a>
    • eh853 minutes ago
      The short seller report contains a bunch of errors:<p>1. The CS-3 operates on average at &lt;65c, well below its thermal limits. Considering the actual thermal parameters of the wafer, using the short seller&#x27;s own temperature equation would imply the wafer can operate at 85c for 33 years<p>2. &quot;Ghost tokens&quot; are not unique to Cerebras and the 3x lifespan reduction claim is based on extrapolating GPU figures which is a big logical leap<p>3. &quot;since the introduction of the CS-3 in 2024, less than half of one percent of all CS-3 wafers deployed have been decommissioned.&quot;<p><a href="https:&#x2F;&#x2F;cerebras.ai&#x2F;blog&#x2F;what-it-took-to-build-wafer-scale-and-why-it-lasts" rel="nofollow">https:&#x2F;&#x2F;cerebras.ai&#x2F;blog&#x2F;what-it-took-to-build-wafer-scale-a...</a><p>Disclaimer: I work at Cerebras
      • augment_me37 minutes ago
        Very interesting! Thanks for answering with your take, and the link.<p>Questions:<p>1) What is the idle power draw of the chip? Is the 80% number wrong?<p>Nvidia is idling at around 180W&#x2F;15% of their max for B300. So saying that &quot;it&#x27;s not unique to Cerebras&quot; is meaningless if we don&#x27;t consider the power draw itself<p>2) Decommissioned is not the same as RMA:d&#x2F;refurbished&#x2F;redeployed, so giving this number is kind of dodging the question IMO
        • wmf17 minutes ago
          None of these expensive chips should be idling.
          • augment_me8 minutes ago
            What a non-answer. And there should be perfect latency hiding, and there should always be predictable power-of-2 workloads, and all weights should always be homogenous, etc.
      • lukeify33 minutes ago
        So which Cerebras statement is in error then?<p>&gt; First, in Cerebras’s preliminary IPO prospectuses filed through December 2025, the Company assigned its computer equipment useful lives of just 1.5 to 3 years – in-line with both our Arrhenius modeling and what many of the Company’s hardware customers have described to us.4 Just four months later, in the Company’s April 2026 S-1, Cerebras nearly doubled that range to 3 to 5 years – a change that significantly inflated reported margins. The Company appears to have never explained the reason for this change.<p>Was the useful life of a wafer 1.5–3 years on your IPO filing or 3–5 years in your S-1?<p>Hard to believe a company with such skilled engineers couldn&#x27;t figure out the correct useful wafer life given existing industry data. Or (more likely) are we just looking at standard corporate BS?
        • wmf15 minutes ago
          That&#x27;s answered in the blog.