This project catalogues estimated int8 compute accessible to different actors. You will notice that OpenAI is absent; they don’t own any hardware, they rent it from Microsoft. This is a list of people or groups who own and control decisions about how AI compute is used.
Each entity now includes a compute_metric tag in front matter so the table can distinguish
normalization quality:
dense_int8_tops: Value is intended to represent dense INT8 operations per second.vendor_ai_tops: Value comes from vendor AI TOPS style marketing/spec numbers and may not be
strictly dense INT8-equivalent.non_int8_proxy: Value is a proxy (for example GFLOPS) because no direct INT8 TOPS figure is
published.This figure outlines DBSCAN clusters for compute versus stakeholder counts. It looks for groups of entities with similar characteristics which are sufficiently separate from other groups to be considered distinct. Learn more about the DBSCAN analytical technique here.
This is the list of the data from all the entity files when the analysis was last run, along with their estimated int8 compute capacity and number of stakeholders. You can sort the table by clicking on the column headers. You can edit the data in the appropriate entity file, linked below.