Profine automatically profiles and optimizes PyTorch training jobs on real GPUs, delivering measurable speedups and lower GPU costs before teams waste days tuning configs by hand.

NAV extracts and analyzes GPU performance traces from NVIDIA Nsight™ Systems (NSYS), enabling comparative analysis and visualization for efficient performance profiling and regression testing.

Communication cost modeling for tensor parallel LLM inference with TP vs PP vs hybrid comparison, VRAM analysis, pipeline bubble modeling, regime detection, and cost-efficiency. Shows TP dominates on NVLink, PP has 47% bubble at 8 GPUs, and LLaMA-70B needs 8× A100 or 2× H100 for VRAM.

Per-precision XMX (matrix engine) profiling for Intel Arc GPUs via Level Zero metric streamers — observes any workload without wrapping it

Collection of examples and links that uses different profiling tools to show memory usage and timings.

NAV extracts and analyzes GPU performance traces from NVIDIA Nsight™ Systems (NSYS), enabling comparative analysis and visualization for efficient performance profiling and regression testing.

Automated GPU profiling analysis for Adreno — turns Snapdragon Profiler captures into actionable insights with LLM

Unified benchmarking and profiling framework for the JAX scientific ML ecosystem. Timing, GPU/energy monitoring, FLOPS counting, roofline analysis, statistical testing, regression detection, and CI integration.

Cross-platform .NET performance engineering skill for coding agents, covering CPU, memory, GC, benchmarking, concurrency, startup, native profiling, GPU rendering, and production diagnostics on macOS, Windows, and Linux.

"The GPU Watchers swore upon their shared memory hierarchy, from L1 to global memory, which also served as their mandate as lords of parallel computation."

Long-context benchmark pushing Qwen2-0.5B from 4K to 32K tokens on RTX 2070 using SDPA + chunked prefill. Shows 40x speedup at 8K, FP16 beating INT4 at long context, and that quantization is NOT a long-context solution — KV-cache is the real bottleneck.

Profiling and Triton-based KV-cache optimization for protein language model inference on consumer GPUs.

Hand-built MoE expert parallelism: routing-skew capture, fused grouped-GEMM Triton kernels, 2-all2all dispatch/combine, overlap ablations — proving by counter-example why DeepEP exists.

Real implementation of speculative decoding showing that speedup is strongly prompt- and target-dependent: GPT-2 → GPT-2-medium achieves mean best speedup 1.013×, while GPT-2 → GPT-2-large reaches 1.253× and up to 1.846× on high-agreement prompts.

Attention backend benchmark on Turing GPUs comparing Vanilla, SDPA Math, SDPA Efficient, and a custom Triton FlashAttention implementation. SDPA efficient achieves 130× memory reduction and 10× speedup; Triton FA achieves O(n) memory but is 64× slower than SDPA efficient on RTX 2070.

Empirical profiling of FP32, FP16, INT8, and INT4 quantization on GPT-2 and GPT-2-medium: throughput, decode latency, model memory, and perplexity across batch sizes 1–16. FP16 wins on throughput; INT4 wins on memory with negligible quality loss.

AI for Science 全栈评测平台:用 JAX/PyTorch 实现 PINN、DGM、ML-CFD 等科学计算求解器,配套毫秒级 GPU 性能剖析工具链与 Django 在线评测平台(任务调度→指标采集→可视化→排行榜)。

Empirical profiling of chunked prefill showing two regimes: for isolated requests, full prefill always wins TTFT; under mixed workloads, chunked prefill improves fairness by reducing short-request TTFT, with chunk256 offering the best tradeoff.

Profile-first CUDA lab for killing GPU pipeline bubbles: stream overlap, CUDA Graphs, torch.compile, spec-decode overlap — Nsight-verified

Benchmark of KV-cache eviction policies on GPT-2-medium showing two clear regimes: sink-preserving heuristics dominate at tiny budgets, while attention-based eviction becomes near-lossless at moderate budgets. attention_384 matches full-cache quality with ~33% less cache.