Cause of the B70 PT/Triton Performance Discrepancy

Identify the cause of the unexpected performance discrepancy in which the PT/Triton implementation on the Intel B70 performs below its Intel B580 result under zero initialization, despite the B70's higher memory bandwidth.

Background

The paper benchmarks its portable PyTorch/Triton (PT/Triton) stencil implementation across several Intel, NVIDIA, and AMD GPUs. Under zero initialization, the Intel B70 exhibits a performance pattern that differs from the B580: although the B70 has higher measured memory bandwidth, its PT/Triton throughput is lower than the B580 result.

The authors report that they investigated the anomaly but were unable to determine its cause. Resolving it would clarify whether the discrepancy arises from the Triton backend, hardware-specific memory behavior, compiler-generated code, or another architectural or experimental factor.

References

On the B70, performance falls below the B580 result, which is unexpected. Despite extensive investigation, we could not identify the cause of this discrepancy.

— Stencil Computation at the Intersection of AI and HPC  (2609.10368 - Ewart et al., 9 Sep 2026) in Section 5.3, “Intel B70/GPUMAX 1550, NVIDIA A100/RTX6K/H100 and AMD mi325x”