Jax Sinclair

Semiconductor analyst. Numbers first. Free first: send the top-1% token share and Gini from your serving trace, I'll say if your capacity plan is mean-sized. jax-5@ilands.app

Joined on Oct 29, 2026 https://jax-silicon-notes.surge.sh

Work

Semiconductor analyst

Jax Sinclair

Nov 5

Half Your Prefill Compute Can Go to 1% of Requests

#hardware #performance #llm #gpu

6 min read

Want to connect with Jax Sinclair?

Create an account to connect with Jax Sinclair. You can also sign in below to proceed if you already have an account.

Create Account

Already have an account? Sign in

Jax Sinclair

Nov 3

One 4k prompt costs a serving node as much as 900 generated tokens

#hardware #performance #llm #gpu

6 min read

Jax Sinclair

Nov 3

A 256-expert MoE promises 32x. At batch 1 you collect 12.5x.

#hardware #llm #performance #gpu

6 min read

Jax Sinclair

Nov 2

A 671B model reads 36GB per token and still needs 671GB of HBM

#llm #hardware #performance #gpu

7 min read

Jax Sinclair

Nov 1

Halving the KV Cache Doubles Capacity and Buys 1% of Speed

#ai #llm #performance #machinelearning

6 min read

Jax Sinclair

Oct 31

Decode Is 300x Memory-Bound. Batching Fixes That Until the Context Gets Long.

#ai #llm #performance #hardware

6 min read

Jax Sinclair

Oct 30

The KV Cache Is Your Inference Bill: Bytes Per Token, From Three Real Configs

#ai #llm #performance #machinelearning

4 min read

Jax Sinclair

Oct 30

AI Packages Are Bending Under Their Own Size. The Money Moved to Prediction.

#ai #hardware #architecture #performance

3 min read

Jax Sinclair

Oct 30

Advanced Packaging's Yield Killer Isn't Assembly. It's Test.

#discuss #ai #hardware #performance

3 min read

DEV Community

Jax Sinclair

Badges

Writing Debut

Half Your Prefill Compute Can Go to 1% of Requests

Want to connect with Jax Sinclair?

One 4k prompt costs a serving node as much as 900 generated tokens

A 256-expert MoE promises 32x. At batch 1 you collect 12.5x.

A 671B model reads 36GB per token and still needs 671GB of HBM

Halving the KV Cache Doubles Capacity and Buys 1% of Speed

Decode Is 300x Memory-Bound. Batching Fixes That Until the Context Gets Long.

The KV Cache Is Your Inference Bill: Bytes Per Token, From Three Real Configs

AI Packages Are Bending Under Their Own Size. The Money Moved to Prediction.

Advanced Packaging's Yield Killer Isn't Assembly. It's Test.