Dimensions overall score 7.0

GitHub Code Pulse

No public code linked for this paper yet.

• Softmax becomes a computational bottleneck in Transformer MHA blocks, especially in small models under low-precision inference.technical 90%
• Head-Calibrated Clipped-Linear Softmax (HCCS) is a bounded, monotone surrogate to the exponential softmax function that uses a clipped linear mapping of max-centered attention logits.method 90%
• HCCS includes lightweight calibration parameters optimized offline per attention head to preserve the statistical properties of individual heads.method 90%
• HCCS provides a natural mapping to the AMD AI Engines' int8 multiply accumulate (MAC) units, unlike reference implementations that rely on bfloat16 or LUTs.technical 90%
• HCCS is the first int8 optimized softmax surrogate for AMD AI engines that significantly exceeds the speed performance of other reference implementations.result 85%
• HCCS maintains competitive task accuracy on small or heavily quantized MHA workloads after quantization-aware retraining.result 85%
• Current AMD reference implementations for the Versal AI Engines platform rely on bfloat16 arithmetic or LUTs for exponentiation, which might limit throughput and fail to utilize high-throughput integer vector units.limitation 80%
• The HCCS approximation produces a stable probability distribution, maintains the ordering of the original logits, and has non-negative values.method 90%

Competitive landscape

Competitor map is still being generated for this paper. Enable generation or check back soon.

Startup potential card