Some infrastructure technologies are difficult because the terminology hides what’s actually happening underneath. Concepts breaks them down one at a time — how they work, why they exist, where they fit, and what the trade-offs look like in an enterprise environment.
One concept at a time. No assumption that knowing the acronym means understanding the system.
New here? Read them in this order: GPU Architecture → GPU Data Path → GPU Time Slicing → MIG → GPU Performance → Reading DCGM Counters. They build on each other.
- MIG: An Old Idea in New Silicon
October 24, 2026
How NVIDIA MIG compares with IBM Power LPARs and HPE Superdome nPartitions, plus reliability, GPU sharing methods and profile sizes.
- GPU Data Path: Why Your GPU Is Starving
October 18, 2026
How a training batch moves through storage, CPU memory, PCIe, HBM, cache, and the GPU SMs — and where performance gets lost.
- Reading DCGM Counters
October 10, 2026
Two DCGM counters read as a pair tell me whether a GPU workload is compute-bound, bandwidth-bound, or broken — and my first measured run gave me a textbook example.
- GPU Performance: Why Utilization Lies
October 9, 2026
GPU utilization is not GPU performance. Understand the metrics that matter for AI infrastructure.
- GPU Time Slicing
October 7, 2026
Sharing one GPU across multiple workloads by slicing time — how it works, where it breaks, and when it's the right call.
- GPU Architecture
September 27, 2026
What's inside an H100: GPCs, streaming multiprocessors, HBM3, and the memory hierarchy that decides performance.