GCAT · GPU Cluster Acceptance Testing
Are your GPU nodes
healthy and performing as expected?
Run our acceptance testing on nodes you manage and create a report that can be shared. No Liquid Compute account needed.
What GCAT tests.
GPU node health
Hardware diagnostics and sustained-load checks for faults and stability.
GPU performance
Compute throughput, GPU memory bandwidth and PCIe transfer performance.
Cluster performance
Collective communication and network checks for supported multi-node configurations.
How it works
How we test your GPU nodes.
Select your hardware
Pick the GPU configuration that matches your nodes.
Add our key
Add our SSH key to authorized_keys for a user with root or passwordless sudo.
Preflight
We log in and check the nodes match the selected configuration: GPU model and count, driver and CUDA versions and OS.
Test
Click Start test. We copy our test files to the node, run the test’s checks without installing anything and stream progress to this page.
Report
We remove our files and verify cleanup, then report each measurement and whether it met its acceptance threshold.
Frequently asked questions
GPU node testing FAQ.
What is GCAT?
GCAT (GPU Cluster Acceptance Testing) is Liquid Compute's test suite for GPU nodes and clusters. It runs hardware diagnostics, sustained-load tests and performance measurements, and checks each result against reviewed thresholds. Its test files are copied to the node for the test and removed afterwards; nothing is installed.
What does GCAT test?
Every test runs NVIDIA DCGM diagnostics (dcgmi diag) and DCGM memory bandwidth, targeted stress and targeted power tests, a CUTLASS matrix-multiply (GEMM) benchmark, and nvbandwidth GPU-memory and host-to-GPU (PCIe) bandwidth tests. Tests on 8-GPU nodes add an NCCL all-reduce across the GPUs in the node. Cluster tests add an NCCL all-reduce across nodes over RDMA. Each configuration has its own reviewed thresholds.
What are the requirements for each GPU node?
Ubuntu LTS 22.04 or later with NVIDIA drivers, and SSH access to a user with root or passwordless sudo. GPUs must be idle for the whole test.
Which GPUs can I test?
GCAT has reviewed tests for NVIDIA A10, H100, H200, B200 and B300 hardware. Available tests depend on the current catalogue. Check the configuration list before starting.
What is the difference between a GPU benchmark and a GPU stress test?
A GPU benchmark measures performance, such as compute throughput and memory bandwidth. A GPU stress test applies sustained load to check behavior under demanding workloads. Hardware diagnostics check for faults. GCAT combines these checks in a reviewed plan for your hardware configuration.
Can I test a GeForce RTX GPU?
Consumer GeForce RTX GPUs are not currently supported. GCAT runs on NVIDIA server hardware.
Do I need an account?
No. You can start a test, follow it and read your report without an account. Your report stays available in this browser for 30 days. Save it as a PDF to keep a copy.
What access do I give Liquid Compute?
You add our key to authorized_keys for an account with root or passwordless sudo. This gives Liquid Compute SSH access to run our GPU node tests.
Can I test GPUs while they are running other workloads?
No. The tests put full load on the GPUs and need them to themselves. The test will not start if another process is using a GPU. We don't change your driver or GPU settings.
How long does a GPU node test take?
Duration depends on the configuration. Full diagnostics may take 90 minutes or longer. The selected configuration shows its maximum duration, and the test shows actual progress.
What happens if I close this page?
The test keeps running. Come back in the same browser to follow it and read your report.
What is left on my node afterwards?
Our key in authorized_keys, until you remove it. We remove our test files and check that they are gone. We install no packages, services or accounts, and we change no settings.