ESL Tutorial: Performance Validation of Graphics Processing Units (GPUs)
Consistent, high performance unleashed by every generation of NVIDIA’s Graphics Processing Units (GPUs) makes them invaluable to systems ranging from supercomputers to gaming rigs to self-driving cars. Validating GPU performance with high accuracy and confidence is key to delivering high performance GPUs for the above systems. Over the years, though GPUs’ complexity has grown manifold, NVIDIA engineers have finessed the art of validating GPU performance and ensuring that performance goals are met, long before production-ready chips tape out.
In this tutorial, attendees will learn about the following aspects of performance validation at NVIDIA:
1) Importance of validating performance at various levels of the full-system hierarchy prior to first tape out, using both directed tests as well as full applications, and using a “Speed of Light” validation philosophy to push performance as close as possible to peak.
2) Engaging with SW teams throughout the validation process to successfully navigate through HW/SW interface changes.
3) The types of tools we use to greatly ease the burden of performance validation. Together, the above points ensure that we tape out a high-quality netlist and help avoid any surprises when first silicon returns to our labs. Some or all aspects of this tutorial may be of potential interest to system designers and computer architects with even a passing interest in high performance systems. Prior knowledge of GPUs is not required.