Tag HBM

AI & Compute

Qwen 3.8 27B Needs Two RTX 5090s for Full Context

Three horizontal bars drawn to scale from one shared axis: two short gray bars for the 32,768 and 37,740 tokens of context a single RTX 5090 reached under the published vLLM and SGLang recipes, against a blue bar roughly eight times longer for Qwen 3.8 27B's full 262,144-token window.

Weights that fit in 32GB still left the RTX 5090 waiting half an hour for a first token, and the fix was a different inference engine rather than a different card.

Saya
September 9, 2026

Hardware

AMD Clears 30 Percent of Client x86 While DDR4 Sells Intel’s Older Chips

Headline above a donut ring whose green arc is exactly 30.3 percent of the full circle, AMD share of the whole client x86 market, to scale.

Mercury Research put AMD past 30 percent of client x86 for the first time, and two Intel chips returned to Amazon's top ten. DDR5 pricing links the two.

Saya
August 26, 2026

AI & Compute

AMD Helios Is Its First Real Rack-Scale Answer to NVIDIA

A grid of exactly 72 cells in twelve columns and six rows, one cell per GPU in the Helios rack.

AMD's MI455X and 72-GPU Helios rack finally challenge NVIDIA at the system level, but peak specifications are only the start of the comparison.

Saya
July 26, 2026

Market & Pricing

DDR5 Prices Have Tripled: What Happened, and How to Buy Memory in 2026

Headline above a staircase of pixel blocks whose last step is exactly three times the first, the tripling of DDR5 prices drawn to scale.

A 32 GB DDR5-6000 kit that cost around $100 in 2025 now lists near $500 at US retailers. The cause sits in the server room, not the retail channel. Here is the supply-side mechanism, the numbers behind it, and a practical way to buy memory while the squeeze lasts.

Saya
July 22, 2026