GPU systems and the measurement around them.

Continued pretraining of a 1.7B-parameter Indonesian language model, trained and evaluated end to end on hardware I built myself. Eleven controlled runs isolating five variables at matched token budget.

• 14.4 / 16 GiBPeak VRAM, fitting the run onto one consumer card through bf16 mixed precision, gradient checkpointing, and sequence length tuned against batch size
• 250 W / 285 WSustained draw across 50-hour runs at full utilization, held at 60 to 75 °C with no throttling
• 0.7 ppBenchmark drift caused by inference batch size alone, comparable to the training effects under test. The evaluation environment was corrupting the measurement
• 11 / 11 nullEvery arm returned null, which established the corpus rather than the training recipe as the binding constraint

The machines GARUDA runs on, specified and assembled from bare components. A Proxmox host with a GPU passed through to a container, a second node for evaluation, and the network between them.

• 100 Mb → 1 GbLink between nodes was negotiating a tenth of its rated speed. Compared against a known-good reference link, isolated it to two cables joined by a coupler rather than to configuration
• 19 → 43 tok/sFirst-request throughput, after tracing a cold-start penalty to CUDA kernel compilation and moving it into service startup
• 48 h recoveredBuilt out-of-band management on a Raspberry Pi after a display-driver hang cost a run that only physical access could clear