Getting started

Overview

HBM is building the memory control plane for AI. Proof of Memory, live today, measures how fast a machine moves data through GPU memory, verifies it and ranks it against identical hardware.

Reads stream from memory to compute (cyan); writes return (violet). HBM measures how fast this bus moves data in your browser.

The platform

HBM is the memory control plane for AI: the layer that measures, verifies and, over time, optimizes the memory AI runs on. It starts in the browser and grows toward fleets.

HBM: The company and the platform.
Proof of Memory: The open measurement protocol: real WebGPU kernels, validated runs and a daily public proof.
MEM Score: The standardized score. Log scale; the reference machine scores 50.
HBM Network: Every verified machine. The Stack is its public shape; GPU results are its data.
HBM Enterprise: Fleet-level memory observability and optimization for AI infrastructure. In development.

What it measures

A run executes real WebGPU compute kernels against 512 MiB of GPU memory for about 50 seconds and reports effective throughput: what this browser can actually move, not what the spec sheet promises.

Phase	Measures
Read	Streaming loads. The path token generation takes through the weights.
Write	Streaming stores. KV-cache appends and activations.
Copy	Buffer to buffer, read and write bytes both counted.
Random	Scattered 16-byte gathers, then a dependent pointer chase for latency.
Sustain	18 seconds of continuous copy. How much throughput the device holds as it heats.

The product

Test: Run the benchmark. Get a MEM Score, a class percentile and a MEMPRINT.
The Stack: Every verified run becomes one die in a public, growing structure.
Leaderboard: Rankings inside each device class, plus daily, weekly and efficiency boards.
Supply chain: Who makes, packages and consumes AI memory. Live Stock Token availability.
Rewards: Capped points for verified runs, personal bests and weekly seasons.

Principles

Measured, not claimed. Every figure is labelled as measured, published or estimated.
Judged inside the class. A laptop never ranks against a workstation card.
Visuals never touch a measurement. The 3D stack draws only between timed GPU passes.
Nothing personal. No accounts, no IP addresses, no hardware fingerprints.

Where to start

Run a test: Requirements, the five phases, reading the result.
Workload: Exactly what the kernels do and how bytes are counted.
Architecture: How the app, the engine and the database fit together.