Research
Experimental work across computing, AI and engineering. Each study comes with a public implementation and its measurements, on any chip.
Acceptance standard
Four checks on the real chip. Pick one to see the numbers.
Every kernel is checked against an FP64 reference across at least 64 shapes. Verify holds back 20 percent of those shapes, so a pass reflects real behavior on new inputs.
shapes per check
held back
A kernel passes at 90 percent of the vendor library's median time on the real chip. Kernel gain and application gain are shown separately in your own metric.
of vendor median time
gains shown apart
Each result is repeated 1000 times with identical inputs. Whole model ports keep 99.9 percent of the task score, and 99 percent for FP8 and INT8.
identical runs
task score kept
Green Computing reports the energy for each accepted result, so Green Computing becomes a number your finance team can track next to cost per result.
number per result
signed proof pack
Research areas
Compare kernel implementations on the same hardware and see how workload structure shapes performance.
Compare accelerator implementations with reference calculations and check the evidence behind each tolerance.
Study action model efficiency, deployment limits and the detection of unusual robot actions.
Train control policies and compare their behavior on evaluation episodes kept apart from training.
Examine model failure modes, training data quality and the reliability of automated evaluation.
Evaluate classification, forecasting and retrieval methods with examples that make model behavior easy to inspect.
Compare numerical models with analytical solutions and reference cases across physical engineering problems.
Model batteries, motor drives, power converters and electrical grids.
Examine the software around training and deployment, including experiment records, request batching and performance measurement.
Portfolio
Selected studies from our collection of 47 public projects.
Read the study
Read the study
Read the study
Read the study
Read the study
Read the study
Invention areas
Compare GPU outputs, measure numerical error and keep the evidence behind each result.
Connect allocated resources with workload execution and records of the work delivered.
Help interrupted GPU workloads resume from usable saved state.
Use technical evidence to assess GPU infrastructure and its usable computing capacity.
Put your kernels through the same standard.