Shengyi Wei

AI DATA MOVEMENT · MEMORY SYSTEMS · INTERCONNECT · RTL

From the hardware's side, an inference run is a traffic pattern. I work on reading that pattern off real workloads, and on what follows from it: memory systems, on-chip and chip-to-chip networks, and the RTL underneath.

One decode step of a pre-norm decoder layer, drawn as traffic. Three bands: the operator order — RMSNorm, QKV projection, RoPE, KV append, attention, output projection, RMSNorm, MLP — each tagged with what it reads or writes off-chip; four HBM channels, all busy, with the served stripe sweeping them in turn; and weights multicast from the memory edge into a mesh of tiles whose partial sums reduce into an accumulator bank. A segmented gauge grows with the KV cache at every step, and the tiles stay far idler than the links.One decode step of a pre-norm decoder layer, drawn as traffic. Three bands: the operator order — RMSNorm, QKV projection, RoPE, KV append, attention, output projection, RMSNorm, MLP — each tagged with what it reads or writes off-chip; four HBM channels, all busy, with the served stripe sweeping them in turn; and weights multicast from the memory edge into a mesh of tiles whose partial sums reduce into an accumulator bank. A segmented gauge grows with the KV cache at every step, and the tiles stay far idler than the links.
FIG. 01 — ONE DECODE STEP OF A PRE-NORM DECODER LAYER, DRAWN AS TRAFFIC.WEIGHTS MULTICAST FROM THE MEMORY EDGE, PARTIAL SUMS REDUCE DOWN EACH COLUMN. THE LINKS STAY BUSY WHILE THE TILES WAIT. SCHEMATIC, NOT TO SCALE.

Life is a long trade, not a greedy algorithm.

WRITING16 ENTRIES

  1. 手搓GPU(二)
  2. GPU3
  3. 手搓GPU(一)
  4. 如何提问技术问题
  5. LifeManual
  6. 工业级RTL风格
  7. 高速接口PCIe无痛入门(五)物理层(2)
  8. 高速串行通信编码 Encoding
  9. 高速接口PCIe无痛入门(四)物理层(1)
  10. 高速接口PCIe无痛入门(三)数据链路层
  11. 高速接口PCIe无痛入门(二)事务层
  12. Resources
  13. Advanced Computer Architecture CA1 分支预测竞赛
  14. 高速接口PCIe无痛入门(一)总体架构
  15. 速读ASIC综合(一)
  16. Hello World

ARCHIVENOTES