ykchen-intel

GEMM Shape Clustering — Skill Reference

• Machine: scaia254.sc.intel.com
• Absolute path: /mnt/20TB/home/ychen4/20260422-GEMMShapeCluster/
• Relative path: ~/20260422-GEMMShapeCluster/

Goal

Given a collection of GEMM workload traces from multiple LLM models, reduce 1300+ unique (M, N, K) shapes down to ~17 representative medoids that can be

ykchen-intel

Analytical GEMM Performance Study

1. Objective

Project GEMM performance for specific workload shapes on Intel Xe4 (48 XeCores) and Xe5 (176 XeCores) GPU configurations, sweeping HBM memory latencies from 650ns to 1000ns at 1.5 GHz core clock. Identify optimal tile configurations and characterize bottlenecks per shape.

---

2. Setup Steps

ykchen-intel

SKILL.md - GPU Analytical Projections

Overview

This repository provides analytical performance projection models and simulators for Intel Xe4 (JGS) and Xe5 (TGS) GPU architectures. It models GEMM, Flash Attention, Flash Decoding, and MoE (Mixture of Experts) workloads to predict throughput, latency, bandwidth utilization, and bottleneck classification at the XeCore and SoC level.

---

Prerequisites