● RESEARCH & SYSTEMS PUBLICATIONS

Systems, Shaders & Craft.

Deep technical investigations into GPU memory architectures, OpenAI Triton kernels, high-throughput LLM serving, and tactile mathematical UI.

/

GPU Systems // 02

Featured Physical MonographHardcover Edition

Dissecting Memory Bandwidth & SRAM Tiling in Triton: From GEMM to FlashAttention-3

Why LLM inference is strictly memory-bound in decoding, how HBM3e vs SRAM bandwidth dictates latency, and writing a fused Triton kernel with online softmax.

Target Silicon

H100 SXM5

Bandwidth

3.35 TB/s (98%)

Compiler

Triton 3.0

SRAM Tiling

-93.6% DRAM

Nov 11, 2026·7 min read

Read Full Monograph

CODENUCLEI PRESS · VOL. 012026

TRITON & SRAM TILING

FROM HARDWARE TO SCALE

HARDCOVER EDITION

Design Systems // 02

DISTRIBUTED SYSTEMS · VOL. 022026

SPECULATIVE DECODING

CONTINUOUS BATCHING vLLM

HARDCOVER EDITION

Design Systems2m read

Matte white, one theme

One light theme, no dark mode toggle, built from three Tailwind neutral ramps: stone for the canvas, zinc for ink, neutral where something must be colourless.

Nov 11, 2026Read

Engineering // 02

CODENUCLEI PRESS · VOL. 012026

TRITON & SRAM TILING

FROM HARDWARE TO SCALE

HARDCOVER EDITION

Engineering1m read

Demos next to the blog

Every side project ends up with a demo that needs to live somewhere. This site hosts them under /demos/<name> on the same domain as the writing.

Nov 11, 2026Read

AI Inference // 02

DISTRIBUTED SYSTEMS · VOL. 022026

SPECULATIVE DECODING

CONTINUOUS BATCHING vLLM

HARDCOVER EDITION

AI Inference6m read

Continuous Batching, PagedAttention, and Speculative Decoding in High-Throughput LLM Serving

How virtual memory paging eliminates KV-cache fragmentation, while draft-model speculative decoding breaks the sequential token latency wall on Modern GPUs.

Nov 10, 2026Read