PIM GPT

 PIM GPT paper

Session 23 AI accelerators ISSCC 2025

https://fpga.org/2014/12/31/the-past-and-future-of-fpga-soft-processors/

Improve GPU

https://notes.ekzhang.com/events/nysrg - computer systems

16 bit CPU from scratch in C

https://cacm.acm.org/research/deconstructing-the-bakery-to-build-a-distributed-state-machine/

https://www.developing.dev/p/turing-award-winner-on-working-with

Design of a GPU with Heterogeneous Cores for Graphics

How to overlap CPU and GPU

Fast GEMM from scratch

SRAM/DRAM bifurcation

Both test time training (a form of online learning) and linear attention operates at O(1) memory, O(n) compute complexity. Ever wonder why they have the exact same properties? Turns out they are just the same thing and we proved it!

https://dyahadila.github.io/blog/2026/industry-job-search-cs-phd/

https://pages.cs.wisc.edu/~fredsala/cs639/schedule.html

DeepSeek gpus' vram only needs to hold a single layer's worth of kv vectors to serve a request

https://www.lesswrong.com/posts/MNpBCtmZmqD7yk4q8/my-understanding-of-anthropic-strategy

stage the KV's in the DRAM on the decode GPU servers, and then transfer the KV's to the prefill GPUs via GDRDMA