EcoCompute — Notes https://quantenergy.tech/blog/ Evidence-based notes on the energy cost of LLM inference, from direct NVML GPU power measurements. en Sat, 25 Jul 2026 00:00:00 GMT We ran our own container on a rented RTX 4090: NF4 crosses over at 7B, INT8 never wins https://quantenergy.tech/blog/we-ran-our-own-container-rtx4090.html https://quantenergy.tech/blog/we-ran-our-own-container-rtx4090.html Sat, 25 Jul 2026 00:00:00 GMT 15 of 15 configurations measured with direct NVML on a rented RTX 4090 using our own open MLCube container. NF4 flips from a +36% penalty at 0.5B to a -29% saving at 7B; INT8 costs more energy at every size because decode throughput collapses to 9-19 tok/s. Data archived at DOI 10.5281/zenodo.21528103. When not to quantize: the small-model energy penalty https://quantenergy.tech/blog/when-not-to-quantize.html https://quantenergy.tech/blog/when-not-to-quantize.html Thu, 25 Jun 2026 00:00:00 GMT Weight-only NF4/INT8 quantization shrinks memory but can raise LLM inference energy by 25–55% for small models. What 360+ real GPU measurements show about the crossover point, and when quantization is still the right call.