MixPert: Optimizing Mixed-Precision Floating-Point Emulation on GPU Integer Tensor Cores
summary
MixPert implements a 3-integer shared-exponent format and an optimized data layout plus augmented pipeline to emulate FP32 on GPU Integer Tensor Cores, using 6–9 tunable emulation steps and a lightweight tuner. Implemented on NVIDIA A100 and RTX GPUs, it builds matrix-multiply kernels and an autotuning flow, delivering single-matrix speedups up to 2.1×, average app speedups ~1.6× and 1.72× versus cuBLAS on general-purpose cores, and beating APE and CUTLASS by ~1.22× and ~1.21× while controlling error.
AI-written summary · about
chapters (6)
- 0:00MixPert introduction
- 2:13FP32 emulation challenges
- 3:40Integer emulation techniques
- 6:09Integer-based floating-point emulation
- 9:42MixPert evaluation
- 11:49Conclusion
related talks
05mApex
28mArray Programming on GPUs: Challenges and Opportunities
25mAI Acceleration Capabilities Using MMA(Matrix Multiply...
01h 05mReal In-Memory Processing
25mAI acceleration in POWER10
18mArbitor: A Numerically Accurate Hardware Emulation Tool for DNN Accelerators
20mMetis: Fast Automatic Distributed Training on Heterogeneous GPUs
26mEfficient MoE Pre-training at Scale on AMD GPUs With TorchTitan