Questions

HBM is building the memory control plane for AI: the layer that measures, verifies and, over time, optimizes the memory AI runs on. Compute is useless if you can't feed it. It starts with Proof of Memory, which anyone can run in a browser.

HBM's open measurement: real WebGPU kernels measure how fast a machine moves data through GPU memory. Every run is validated, scored as a MEM Score and ranked only against identical hardware, and each day's verified runs go into a public proof.

Fleet-level memory intelligence for AI infrastructure: which GPUs are memory-starved, where VRAM sits idle and which machines underperform their class. It's in development, and teams can request early access. HBM Enterprise.

How fast your browser can move data through your GPU's memory. Five phases run real WebGPU compute kernels over 512 MiB: read, write, copy, random access and a sustained copy. How the workload works.

About 50 seconds, in five phases. Keep the tab visible while it runs. What happens during a run.

Chrome and Edge 113 or later on Windows, macOS and ChromeOS; on Linux, turn on WebGPU in chrome://flags. Safari 26 and Firefox 141 on Windows run the test too, but hide the GPU model, so their runs are measured and unranked. Requirements.

It runs ordinary GPU compute inside the browser's sandbox, like a game or a WebGL demo, using about 512 MiB of GPU memory for under a minute. Your fans may speed up, and closing the tab stops it at once.

Published figures are theoretical peaks. HBM measures effective throughput from a browser, which lands below peak on every device; the efficiency board shows how far.

One number for the whole run, on a log scale. The reference machine scores 50, ten times faster scores 100, and each doubling adds about 15 points. How the score is built.

It was measured but cannot join a class ranking: the browser hid your GPU model, the model has no published memory spec yet, or the run used quick mode. Unranked runs still join The Stack. Validation.

Its numbers were physically implausible: well above the published peak, inconsistent between phases, or finished faster than the workload allows. Rejected runs are kept for review and do not join The Stack. Validation.

Only against identical hardware. Each device class has its own leaderboard, and percentiles appear once a class has five verified runs. Leaderboards.

A public record built from results: every accepted run adds one die, and sixty-four dies make a layer. Open The Stack.

No. The first run creates a random key in your browser; the server keeps only a hash of it, shown as OP-XXXXXX. Identity and privacy.

Your measurements, the GPU details your browser reports, a coarse browser and system name, and the hash of your key. No names, emails or IP addresses. Privacy Policy.

Copy the recovery key on the Rewards page and restore it in the other browser. Without the key, a cleared browser starts a new ledger.

No advertising or tracking cookies. A few essential entries live in your browser's storage, and cookieless analytics load only if you allow them. Cookie Policy.

Verified runs, first runs on a new device class, early data for small classes, personal bests and weekly seasons. Rewards.

Points have no cash value themselves. Each week, fixed pools are split by points among linked wallets and paid on Robinhood Chain while funded, and a verified run can also pay a few cents of NVIDIA stock instantly, outside the US. Rewards.

Instant NVDA rewards and weekly pools pay only wallets linked to an operator. Linking is a free signature, not a transaction, and you can unlink at any time on the Rewards page.

HBM's token on Robinhood Chain, launched on Pons v2. Token docs.

Only the one shown in the token docs. Anyone can launch a token with the same name, so check the address before you trade.

Connect a wallet with ETH on Robinhood Chain and trade on Pons v2 while $HBM is on its bonding curve. After graduation, trading moves to its Uniswap v4 pool. Trading.

The launch graduates: its liquidity moves into a Uniswap v4 pool and is locked permanently. How the launch works.

No. It can lose all of its value, and HBM makes no promise about its price, returns or listings. Read the disclosures before trading.