Summary
Alex Ziskind starts from AMD’s announcement of a workstation with up to 2 TB of RAM and asks whether that much system memory matters for AI. He tests it on a Threadripper Pro 9975WX build with two RTX Pro 6000 GPUs, and then on a much cheaper 32 GB Radeon R9700.
Instead of measuring tokens per second, he runs a growing number of real coding agents against one model at the same time, with each agent’s conversation parked in system RAM between turns. With 128 GB of RAM, eight agents ran fine and 12 was the ceiling; at 16, throughput fell sharply and 10 jobs failed. Swap stayed at zero and the CPU hovered around 6%, which indicates that system memory, not VRAM, ran out first. He also describes discarding his first benchmark because it leaned too heavily on the GPU.
With the smaller Radeon card, a 124-billion-parameter mixture-of-experts model ran at about 38 tokens per second, with only around 5.5 GB on the GPU and roughly 90 GB held in system RAM. The takeaway is that big models are not necessarily slow when only a few experts activate per token. The video also introduces Agent Headroom, a new open-source benchmark suite for agent workloads, available on GitHub in a 0.1 release. A sponsor segment for ChatLLM Teams appears midway through.
📺 Source: Alex Ziskind · Published November 05, 2026
🏷️ Format: Benchmark Test
