Omdia benchmarks HPE, Kamiwaza, and NVIDIA's AI stack: up to 20x faster time to first token, 17x higher GPU throughput.