Deck 03

🔩 Hardware

GPUs, AI accelerators and the silicon race — from rackscale to desktop.

Modern gaming graphics card with triple fans

Supercomputing 2026 delivered a three-way shootout: AMD leads inference benchmarks, Nvidia answers with a record CUDA-core science GPU — and compact local-AI desktops finally go mainstream.

SC26: The GPU War Goes Rackscale

AMD
Instinct MI455X: 21,333 tok/s on DeepSeek-R1
Rackscale server system with dense cabling

Measured MLPerf-style offline inference on the Helios rackscale system — up to 30% over Nvidia's latest published Vera Rubin result, with up to 20% higher HBM bandwidth.

Nov 12, 2026
AMD
MI430X — the science-first GPU

Native FP64 (288 TFLOPS) plus 432GB HBM4 at 23.3 TB/s — 50% more memory than Vera Rubin — in a glue-less four-GPU Infinity Fabric mesh for HPC + AI convergence.

Nov 12, 2026
AMD
New MI450X joins the MI400 family

288GB HBM4 per GPU with the same compute and scale-up connectivity as MI455X — a lower-TCO option for workloads that fit the smaller memory envelope.

Nov 12, 2026
Nvidia
Rubin HPC: 50,000+ CUDA cores per GPU
Founders-edition style graphics card standing upright

Announced at SC26 — a Vera Rubin platform variant built for agent-driven scientific simulation: record CUDA core count, sparse data acceleration, 7th-gen NVLink. Available 2028.

Nov 16, 2026
Nvidia
RTX Spark compact desktops ship

ASUS, Dell and MSI systems out now: 20-core Grace CPU, Blackwell RTX GPU, up to 128GB unified memory, 1 petaflop AI — runs 120B-parameter LLMs locally. Acer, GIGABYTE, HP and a Microsoft Dev Box coming.

Nov 18, 2026
AMD
6th Gen EPYC "Zen 6" goes GA
AMD EPYC server processor in its socket

SP7 server CPUs shipping from all major OEMs/ODMs, with Azure and OCI instances announced — pitched at agent orchestration and concurrent agent sandboxes in the enterprise.

Nov 10, 2026

Context & Analysis

Roundup
SC26: Rubin HPC vs MI450X vs Intel D270

All three GPU makers used Supercomputing 2026 to unveil datacenter accelerators in the same week — a rare head-to-head moment for the AI silicon race. Tom's Hardware coverage →

Nov 21, 2026
Trend
Local AI hardware is finally mainstream

From RTX Spark desktops running 120B models at home, to open-weight laptop models and video generation on 8GB VRAM — the open-weights wave and the compact-hardware wave are converging. The sweet spot for self-hosters keeps moving down-market.

Nov 2026