以下部分引用 AI 总结，现阶段 AI 仍然有幻觉。请以内容的原文为准。

数据新鲜度群聊 07-14 ✓卖方 07-14 ✓Wrap 断档15天 ✕News 07-15 ✓X 断档12天 ✕页面生成 07-15

Cerebras CS-3

7 atoms · 跨 4 天 · 首见 2026-05-02 · 最近 2026-06-11

三色: 🟦 fact 6 · 🟥 take 1 · stance ▲3/▼2/◆2

来源: X 7

时态: fresh:7

标签: 好信源:4 · 好数字:4

2026-05-17 · system cost
assuming $2–3M per CS-3 system
tags: 好数字 · → daily
value: qty=$2-3M

2026-05-02 · SpMM 性能相比 AMD EPYC 9354P CPU 提升倍数
The results show that the CS-3 can outperform the baseline CPU (AMD EPYC 9354P) by 100x for SpMM with 90% sparse matrices with performance improving as sparse matrix dimensionality increases.
tags: 好数字·好信源 · → daily
value: qty=100x
2026-06-11 · inference tokens per second
CS-3 delivered 2,700+ tokens/sec
tags: 好数字 · → daily
value: qty=2700+ tokens/sec
2026-05-15 · 隐含租金每台$41.96/小时
每台 CS-3 系统的隐含租金还是要 $41.96/小时，是B200的差不多十倍租金
tags: 好数字 · → daily
value: qty=$41.96/小时
2026-05-02 · SpMM kernel 优化推出
In this paper, researchers have proposed novel SpMM and SDDMM kernels for the Cerebras CS-3, optimized to improve I/O performance, memory footprint, and scalability to large matrices.
tags: 好信源 · → daily
2026-05-02 · SDDMM 性能相比 CPU 提升幅度
"The SDDMM performance is typically an order of magnitude faster than CPU, scaling in performance at the same rate as CPU for increasing density."
tags: 好信源 · → daily
2026-05-02 · 超稀疏矩阵下性能劣势
"As matrices become hyper-sparse, the CS-3 suffers from high host-device communication overhead relative to computation, making the CS-3 slower than GPU and approaching baseline CPU performance."
tags: 好信源 · → daily