Measured GPU systems work

Evidence-backed GPU systems work across serving infrastructure and kernel optimization. Each project page carries the story, the architecture calls, and the measured runs behind them.

Serving infrastructure · 7 experiments · 9 decisions

GPU Inference Decision Lab

An EKS/vLLM lab that turns serving measurements (Qwen2.5-0.5B on single-T4 g4dn.xlarge replicas) into architecture decisions for concurrency limits, autoscaling, context limits, scheduling, and quantization.

Closed-loop burst clients dropped nothing by construction but completed 1,656 requests vs 2,632 open-loop
1.20 req/s long-context knee repeats with 36.8s p95 queue delay
FP8 KV rejected on the current g4dn/vLLM path

View project GitHub

Kernel optimization · 9 experiments · 9 decisions

CUDA Kernel Lab

A Triton kernel lab for LLM-shaped primitives on A10G and H200: correctness-checked benchmarks against PyTorch/cuBLAS, Nsight Compute counters for selected A10G kernels, timing only on H200.

115 operator rows plus 27 decode replay rows on A10G
H200 matmul rows keep best standard Triton around 88-90% of PyTorch/cuBLAS
RMSNorm fp16 reached 5.901x over an unfused eager PyTorch reference (A10G, 4096x8192)

View project GitHub