Arbitor: A Numerically Accurate Hardware Emulation Tool for DNN Accelerators
summary
Arbitor is an open‑source compiler‑based emulator that rewrites ML compiler IR to insert custom low‑precision, approximate and sparsity policies, running on real GPUs to measure accuracy and performance. It reproduces FP16 matrix‑mul and conv results bitwise accurate, matches native FP16 training curves for ResNet50, and runs about 2.6× slower per kernel (end‑to‑end up to 6×). Sensitivity results show ResNet18 needs ~13 total bits (6 exponent+6 mantissa) to match FP32, Transformers/GNNs need ~10–13 bits, and posit yields no clear accuracy win.
AI-written summary · about
chapters (10)
- 0:00Arbiter introduction and motivation
- 1:05Emulation challenges
- 3:59Hardware optimizations accuracy
- 6:48Limitations of prior tools
- 7:31Arbiter’s compiler approach
- 9:56Emulation policy implementation
- 12:01Evaluation setup and accuracy
- 13:45Performance overhead studies
- 15:16Subnormal mode and posit impact
- 17:00Summary and future
related talks
25mAI Acceleration Capabilities Using MMA(Matrix Multiply...
12mMixPert: Optimizing Mixed-Precision Floating-Point Emulation on GPU Integer Tensor Cores
26mLow-Precision Inference without Quality Loss...
22mCentimani: Enabling Fast AI Accelerator Selection for DNN Training with a Novel...
01h 11mMachine Learning Acceleration in an Embedded World
11mSimAI: Unifying Architecture Design and Performance Tuning for Large-Scale Large...
17mMiniFloat-NN: A RISC-V ISA Extension for Low-Precision NN Training
34mOptimizing AI Inference with ML Compilers & Hardware