▦ GPU FIELDNOTES
Local AI hardware / PCIe research edition 02
Reading notes ↗
GPU Hardware Ladder · PCIe research edition

Memory is the limit.
Bandwidth is the pace.

Consumer, workstation, modified and datacenter PCIe cards for local AI, including Tensix accelerators and unified-memory comparisons.

NVIDIAAMDIntelAppleTenstorrent
NVIDIA GeForce RTX 5090 Founders Edition graphics card
GeForce RTX 5090 ↗NVIDIA · Founders Edition
AMD Radeon AI PRO R9700 reference product render
Radeon AI PRO R9700 ↗AMD · reference render, not GIGABYTE variant

Capacity gets the model in. Bandwidth helps get tokens out. Compute, batching and software also matter.

Bandwidth bars · fixed 4,800 GB/s scale
GPU memory and bandwidth specifications from the supplied research sheet
PlatformBandwidth · GB/sMemory · GBResearch notes
Before you compare

Read the
fine print.

About this edition. Curated from the supplied 6 September 2026 sheet for a PCIe-focused comparison. New and revised entries link their sources; retained source-sheet figures are marked where not reverified. Multi-die totals and unified memory are not equivalent to a single GPU’s usable VRAM.

Back to top ↑

Product imagery: NVIDIA and AMD. Brand marks via Simple Icons. Images and marks belong to their respective owners; shown for identification.