Memory bandwidth is the silent killer of LLM performance, but it's almost impossible to "see" why your tokens-per-second drop when you switch hardware. WatchMachineGo solves this by providing a visual simulator that maps out exactly how a local model loads, prefills, and executes inference across different hardware configurations.
The tool is currently under construction and free, making it a low-friction way to perform a deep dive into how weights and KV caches actually move through your system. It's a solid companion for anyone building a local AI workflow who wants to move beyond just "it feels slow" to "I see why it's slow."
Instead of staring at a CLI output, you can actually watch the simulation of data movement. It lets you toggle between setups—like running on a CPU alone versus dual GPUs—to see the tangible impact on speed. It's essentially an explorable explanation for the physical side of LLM deployment.
If you're trying to figure out if a hardware upgrade is actually worth the cost for local hosting, this is a great way to get a mental model of the bottlenecks. Core capabilities:
Hardware Simulation: Compare no-GPU, single-GPU, and multi-GPU setups.Bottleneck Visualization: See how memory bandwidth affects the actual generation process.Lifecycle Tracking: Visualizes the transition from model to prefill and finally to inference.
The tool is currently under construction and free, making it a low-friction way to perform a deep dive into how weights and KV caches actually move through your system. It's a solid companion for anyone building a local AI workflow who wants to move beyond just "it feels slow" to "I see why it's slow."
Next El Niño: Tracking Pacific Warming Signals →
All Replies （4） #
N
G
Tried a similar tool last month; it just lagged my system and gave useless graphs. Overhyped.
0
S
Spent three hours debugging a slow prompt only to realize my RAM was basically a straw.
0