Tag

LLM serving.

Articles: 3

An HBM memory stack beside an idle tensor core array, illustrating the bandwidth bottleneck in AI inference accelerators

AI Inference Chips: Why Memory Bandwidth Beats Peak FLOPS

Emerging Technology

Decode reads the whole model to produce one token. Working from NVIDIA's own published rack figures, here is the arithmetic that explains why serving throughput ignores the number on the box.

10 Sept 2026 · 12 min read

Stacked GPU memory modules beside a rising bar chart of model parameter counts, illustrating the memory cost of serving large open-weight models

Noorain Fathima

Open-Weight Model Serving: The Memory Bill Behind the Licence

Generative AI

Parameters times bytes per parameter decides feasibility before any licence review does. What 375B, 552B and 753B models actually cost to hold in HBM, and why active-parameter counts mislead.

10 Sept 2026 · 12 min read

A sealed silicon package with a signed certificate chain extending from it, representing hardware attestation of a trusted execution environment

Noorain Fathima

Confidential Computing Attestation: What It Proves, and What It Doesn't

Emerging Technology

Attestation proves genuine hardware and a known measurement. It does not prove the code is safe, the data protected either side of processing, or that anyone checked the report — and 2026 research…

10 Sept 2026 · 12 min read