01August 28, 2026/Jamie Dborin

What is a Megawatt Worth?

How much are the tokens worth that you can generate with 1 MW of power?

02August 28, 2026/Fergus Finn

What happens when a GPU writes memory

Following an STG.E SASS instruction through the hardware units in an RTX 4090.

03August 13, 2026/Fergus Finn

What happens when a GPU reads memory

Following an LDG.E SASS instruction through the hardware units in an RTX 4090.

04August 11, 2026/Radostin Stoyanov

On-the-fly snapshot compression for elastic inference at scale

Accelerating snapshot-based model serving with on-the-fly LZ4 memory compression in CRIU.

05August 11, 2026/Fergus Finn

The case for disaggregated LLM serving

Three checkable conditions under which disaggregated prefill is never worse than aggregated serving, and the reasons it's usually better. Plus an inventory-theory model of traffic drift: cold starts, safety stock, and a map of the shocks a deployment can ride out.

06July 30, 2026/Peter Bhabra

Gigatoken: making tokenisation care about the hardware

A first-principles tour of the SIMD scanners, cache-aware layouts, huge pages and worker scheduling behind Gigatoken.

07July 27, 2026/Jamie Dborin

You Could Have Come Up With Kimi Delta Attention

Guide to the DeltaNet Family of linear attention mechanisms.

08July 23, 2026/Fergus Finn

Throughputmaxxing: DeepSeek-V4-Flash on Isambard-AI

Single-node serving throughput for DeepSeek-V4-Flash: baseline, MLA/DP-attention, and MoE kernel choice.

09July 21, 2026/Fergus Finn

NVLink, NVSwitch, and all that

Why scale-up links are fast and short, how NVSwitch grew the domain from a board to a rack, and what TPU, UALink, and scale-up Ethernet do differently.

10July 9, 2026/Fergus Finn

Reverse-engineering NVIDIA's cuda-checkpoint for faster cold starts

Freezing a live CUDA process to host memory and thawing it again, what the driver does — and doesn't — do to make that work, and how understanding that lets us restore CUDA processes up to 4x faster.

11July 2, 2026/Fergus Finn

Width vs. depth: speculating on the margin

Some thinking about how to trade off batching and speculative decoding in a running inference engine.

12June 30, 2026/Peter Bhabra

The swarm that designs itself

We rebuilt Moonshot's Kimi agent swarm and pointed it at a real codebase: ~53× fewer tokens and ~45× cheaper than one long-context agent.