Authors
Philipp Gerasimov, GPU Software Development Engineer, Intel® GPA Product Owner
Pamela Harrison, Software Technical Consulting Engineer
Introduction
The Metrics Tab is an important element of Intel® Graphics Performance Analyzer’s (Intel® GPA’s) signature tool, Graphics Frame Analyzer. This feature displays GPU metrics and allows the understanding of performance for rendering and compute calls, particularly on Intel GPUs. There are hundreds of hardware GPU metrics representing various GPU and 3D/Compute pipeline stages.
In the newest release of Intel® GPA, version 2023.4, the Metrics Tab was significantly re-designed to improve the user experience as well as performance analysis quality. This post describes key new features and options.
Metrics Selection
The Plethora of Metrics
As mentioned, Intel GPUs have hundreds of metrics that provide a very deep understanding of rendering and compute performance. The metric collection time and ability to easily navigate across the metrics is important for doing performance analysis quickly and conveniently.
Initial Metrics Panel Implementation
Originally, Graphics Frame Analyzer was designed for Intel Integrated Graphics (onboard GPUs) which had less than a hundred metrics. It was possible to calculate all of these metrics quickly, so early versions of the tool were designed to display all metrics without the ability to enable or disable any of them.
Implementation of Pinning in the Metrics Panel
With the introduction of Intel discrete GPUs the number of metrics significantly increases. For example, higher SKUs of Intel® Arc™ GPUs can have more than six thousand metrics. Collecting and calculating thousands of metrics for complex rendering and compute workloads can take minutes or even hours.
We introduced the Metrics Pinning technique not long ago. It allowed the pinning of metrics to enable their collection and calculation, omitting metrics that are unpinned. That significantly reduced the time required for metrics calculation, improving analysis speed. Users were able to decide precisely which metrics they needed to view. User presets were also introduced to allow saving sets of selected metrics for future use, switching between saved presets based on the analysis needs at the moment.
However . . .
The initial Metrics Pinning implementation had a number of shortcomings.
- Searching through the lengthy list of metrics was required;
- Selected metrics were shown at the top of the list without any indication of which categories they belonged to;
- User presets were saved in the registry, so it was not possible to extract the lists to move them across different systems or add them back after a reinstallation of the operating system.
Leveling Up From Pinning to Selection
To improve the user experience and fix those shortcomings, a new metrics selection technique was introduced. Now, instead of pinning, it is possible to select metrics individually or select entire groups of metrics by checking the boxes that appear to the right of the metrics. All selected metrics still appear at their position, allowing the understanding of the metrics categories.
What you need at a glance: Hiding and Aggregation
To further improve the experience you can now hide the unselected metrics by clicking on the eyeball to the right of the Delta column. This changes the view so that you can see just the metrics that you have selected.
In addition, we have implemented aggregation. The biggest metrics number increase in Intel discrete GPUs comes from the fact that many metrics are calculated per Xe-core. The more Xe-cores present in the GPU, the bigger the metrics list. In many cases, however, it is enough to understand the average metrics value for all Xe-cores, and view only the individual Xe-core metrics to see if there are big discrepancies across Xe-core values. Hence, in the new Metrics Tab, metrics aggregation was introduced. When Intel® Arc™ GPUs are used, per Xe-core metrics are displayed as one aggregated value in the list, with the second level allowing us to see metrics for individual Xe-cores. This feature has allowed us to reduce the top level list of metrics to a few hundred which significantly simplifies your metrics search.
Metrics Search
The metrics search was also re-designed and improved. Fuzzy search was introduced. Now it is possible to type just few characters of the metrics name to see all names with that character combination anywhere in the name. They will be found and displayed at the top of the list, from the closest similarity to the least.
Presets
Finally, the metrics preset functionality was improved. Now all user presets are saved in individual files which are located in the …/Documents/GPA folder and can be copied across PCs or used for new installations in the future.
In addition, a special automatic Bottleneck preset was introduced. This preset includes all the metrics found in the Bottlenecks Tab for easily finding all metrics related to currently detected bottlenecks. This set will be different for different bottlenecks, based on which metrics apply in each situation.
Summary
With the re-designed Metrics Tab performance analysis becomes simpler and more convenient for both Intel® Arc™ GPUs and Intel Integrated Graphics GPUs.
Software engineer for 20+ years. Excels in all things software, plus connecting people and teams for optimal synergy.