GPU comparison
Compare GPUs by what your job needs
Rank cards by token speed, training throughput, memory bandwidth or tokens per dollar at today’s price. Specs come from the vendors; speed and training scores are indicative, so run your own job before you commit to a long lease.
Token speed: Llama 3 8B, llama.cpp, Q4_K_M, batch 1, from public community runs. Training score: relative throughput on a mixed LoRA fine-tuning workload with H100 = 100. Bandwidth, VRAM and TDP: vendor specifications. Tokens per dollar uses the cheapest live hourly price including the Kilnara fee.