Empirical validation on NVIDIA GPUs and TPUs

Validate Primus Projection empirically on NVIDIA GPUs and TPUs, including the required adaptation of the TPU collective model to XLA collectives.

Background

Primus Projection is evaluated only on AMD Instinct MI300X, MI325X, and MI355X GPUs using the ROCm software stack. The framework represents hardware characteristics through YAML profiles and uses GPU microbenchmarks to anchor kernel and communication performance.

The paper states that NVIDIA support would require an appropriate hardware profile and an eight-GPU benchmark, whereas TPU support would additionally require adapting the collective communication model to XLA collectives. Consequently, the accuracy and portability of the projection method on these accelerator platforms remain empirically unresolved.

References

We have not yet empirically validated on NVIDIA or TPU; porting to NVIDIA requires only a hardware YAML and the 8-GPU benchmark (TPUs also need an XLA collective model).

— Beyond the Trial-and-Error Loop: Hybrid Projection and Automated Tuning for Distributed Training  (2610.19423 - Raina et al., 14 Oct 2026) in Section 4.12, “Scope and Limitations”