A GPU job is not useful just because it starts. It is useful when it finishes, proves the GPU did the work, and gives you results you can actually use.
That is what I wanted to test with @ONcompute:
a matrix multiplication benchmark, using the same kind of repeated calculations that sit underneath many AI workloads.
I ran the job across different matrix sizes, compared GPU performance with CPU performance, and generated charts showing the difference in speed and processing time.
Here is what the job did:
1. detected the available GPU
2. ran the same matrix calculations on both the GPU and CPU
3. tested several matrix sizes
4. measured processing time and throughput
5. generated charts and a results summary automatically.
The result was easy to understand: as the calculations became larger, the GPU handled them much faster than the CPU.
What stood out to me was not only the benchmark itself, but the final detail required to complete the job successfully: every chart, CSV, and result file had to be saved to Ocean’s /data/outputs/ folder.
That may seem small, but it is an important lesson for anyone building with decentralized compute. Running the code is only part of the process, the output needs to be reproducible, available, and saved correctly after the job finishes.
I attached the benchmark charts below.
They show the GPU/CPU difference in throughput and processing time, along with a simple performance summary. For me, this was a useful first step before moving on to more advanced GPU workloads and AI experiment.
Start Building with Ocean Network https://t.co/3uTEmVlk9M
The team at @ONcompute asking the question many have been avoiding:
Why own an H200 when you only need it for 37 minutes?
Rent the compute. Ship the model.
#ETHPragma#PragmaLisbon@ETHGlobal