v0.9.0  ·  87.0% delivered  ·  1,891 hrs flagged last 30 days

Know if you're actually getting the GPU hours you paid for.

Goodput runs a lightweight agent on each GPU node, measures real hardware capability — clocks, ECC, InfiniBand, Xid faults — and turns it into availability weights and SLA claim packs.

Hardware counters only — no process names, code, models, or workload data ever leave the node.
FLEET VIEW · 60s cadence
weight 1.0 = full delivery
Capability weight vs. full delivery
Last 30 days · 17,280 hrs purchased
weight 1.0
Hours purchased17,280
At full capability87.0%
Infrastructure loss1,891 hrs
node-11 · live weight ● fault
0.58
Xid 79 · GPU off bus · 41 min @ 0.0
node-01
0.96
node-02
0.91
node-11
0.58
node-05
0.93
Scroll: install to SLA claim pack in 4 steps ↓
NVML clocks · ECC · PCIe· InfiniBand port state · counters· Xid numeric code + PCI bus· Never PIDs · commands · models · datasets
The problem

Uptime dashboards lie. Goodput measures capability.

Every monitoring tool says the node was up. Goodput asks whether the hardware could actually run at spec — and when it couldn't, whether you could have fixed it yourself. IB switch flaps, Xid 79 drop-offs, and thermal throttling count. Bad training code does not.

See it in action
Uptime dashboard saysNode up — 100%
Goodput saysCapability weight — 58%, Xid 79 on node-11
What you can doExport a timestamped SLA claim pack
How it works

Three steps. One honest number.

No changes to workloads. No SSH from outside. Data in ~60 seconds.

01

Install the agent

One binary per node. Reads NVML, InfiniBand sysfs, kernel Xid events.

$ curl -fsSL …/install-linux.sh | bash
02

Samples every 60s

Batches land in ingest, get classified into weights, roll up hourly. Crash-safe outbox.

weight 0.0–1.0 · hourly timelines
03

Prove it & claim

Fleet weights, node drill-down, timestamped SLA report mapped to provider tiers.

Export PDF → send to provider
See it in action

What Goodput does, end to end.

One cycle — install, fleet view, fault capture, proof.

01 / INSTALL
Trust

Hardware telemetry only.

Built for security reviews. The agent never reads process names, command lines, env vars, filenames, container images, model identifiers, or memory contents.

What leaves the node: GPU UUIDs, clocks, throttle reasons, ECC counts, IB port counters, attributed Xid codes — not raw dmesg lines.

Read the data contract
Get started

Goodput Cloud early access

We host ingest, database, and dashboard. Join the waitlist. When your spot opens, we email your install command and dashboard link.

Waitlist only: example commands below are placeholders until you are onboarded.

Step 2 · Install the GPU agent

After onboarding, run the command we send on each Linux GPU server. Example shape only:

Sample (not your command)linux
goodput-agent --endpoint=https://app.example.com/v1/batches --token=gp_live_example_token --cluster-id=your-cluster-id

Step 3 · Open your dashboard

Fleet weights, node detail, and SLA reports. Data usually appears within about 60 seconds of the first upload.

Sample URL
https://app.example.com/?cluster_id=your-cluster-id

Step 4 · Export proof

Generate timestamped SLA claim packs mapped to provider tiers and send them with your invoice dispute.

Monthly report PDF · infrastructure-attributable hours
17,280GPU-hours purchased / mo
87.0%delivered at full capability
1,891infrastructure-attributable hrs
60ssample cadence, crash-safe
Stop arguing about uptime

Prove what the hardware delivered.

Join the waitlist, install one binary per node, and walk into your next provider conversation with timestamps.