Research paper · November 2026
Biology Does Not Run on One Chip
Abstract
Data-centre GPUs are now designed for training and serving large language models, and each generation's headline gain is in low-precision tensor arithmetic. Life-science computing is not one workload. Molecular dynamics runs in single precision with fixed-point accumulation. Electronic-structure codes need double precision. Structure prediction is limited by memory capacity. Biological foundation models look like any other large model. Genomics pipelines, cryo-electron microscopy and optical and acoustic simulation each have their own binding constraint. This paper sets public specifications for six NVIDIA data-centre GPUs from four architectures, from the 2017 V100 to the 2025 Blackwell Ultra, against the published hardware needs of eight classes of biological computation. Per watt, dense 16-bit tensor throughput (FP16 on the V100, BF16 since) rose 5.4 times from the V100 to the B200. Non-tensor single-precision throughput per watt peaked with the H100 and has since fallen 22 percent. Native double-precision throughput per watt also peaked with the H100, fell 61 percent on the B200, and fell to about 1 percent of the H100's level on Blackwell Ultra. The energy cost of a computation therefore depends on its precision as much as on the chip's age. At peak rates, 10¹⁸ double-precision operations cost about 2.9 kWh on an H100, 7.5 kWh on a B200 and roughly 280 kWh on native Blackwell Ultra, while 10¹⁸ BF16 operations cost 0.20 kWh on an H100 and 0.12 kWh on a B200. An AlphaFold 3 prediction of 5,120 tokens runs 1.8 times faster on an H100 than on an A100 but uses about the same GPU energy. The paper ends with a matching table that a research organisation can use to buy or rent hardware by workload rather than by generation.
Cite
Abhay Chakra Sadineni (2026). Biology Does Not Run on One Chip. Research paper. https://doi.org/10.5281/zenodo.24210929 https://www.abhaychakra.com/papers/biology-does-not-run-on-one-chip/