Systalyze Open-Sources Utilyze, Beats nvtop Accuracy
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Systalyze is open-sourcing Utilyze, a free GPU monitoring tool that uses hardware performance counters to report true compute and memory bandwidth utilization via a Compute SOL % and Memory SOL % framework derived from the roofline model.
- Standard GPU utilization metrics from nvidia-smi, nvtop, rocm-smi, Weights & Biases, Amazon CloudWatch, Google Cloud Monitoring, and Azure Monitor only indicate whether the GPU is active, not actual throughput — the article reports real compute can be as low as 1% while these tools read 100%.
- In Systalyze's tests, nvtop read 100% for three matrix multiplication workloads on an H200, while Utilyze showed 2.6% at N=256, 32% at N=1024, and 88% at N=4096, all within 2% of theoretical ground truth.
- On a memory-bound H200 workload with ground truth arithmetic throughput of 6%, nvtop reported 100%, NVIDIA's DCGM "SM Active" reported 99%, and only Utilyze reported 6%.
- Manya Ghobadi, MIT Professor and CEO of Systalyze, said the gap is "tooling" — engineers who write CUDA kernels know what accurate utilization looks like but have lacked a way to see it continuously in production.
- The misleading metrics drive wasted AI spend on unnecessary hardware and energy, per Systalyze, because a 100% reading signals to buy more GPUs when orders-of-magnitude headroom actually exists.
Why it matters: AI infrastructure teams have trusted a single 100% reading from every major vendor and cloud dashboard to decide when to buy more GPUs. If that number is structurally wrong — Systalyze shows it can be off by more than 90 percentage points on real workloads — procurement and energy decisions across the industry are being made on a metric that doesn't measure what it claims to, and a free tool now lets anyone verify it.
