Table 3. Multi-GPU Throughput Timings (ns/day) for AMBER GB Simulations with a Time Step of 2 fs Using the Parallel CPU Version (12 Intel X5670 Cores or 32 AMD Opteron 3136 Cores on Each Node) and the Parallel GPU Version with the SPDP Precision Model (One Intel X5670 Core and One GPU Per Node)^a.

CPU/GPU	apo-myoglobin (2,492 atoms)	nucleosome (25,095 atoms)
GPU version
8 × M2090	135.1	3.95
4 × M2090	115.0	2.71
2 × M2090	93.1	1.80
1 × M2090	78.1	1.42
CPU version
2048 × Opteron 3136	–	0.53
1024 × Opteron 3136	–	0.78
512 × Opteron 3136	–	0.65
256 × Opteron 3136	–	0.55
128 × Opteron 3136	29.8	0.31
64 × Opteron 3136	18.3	0.17
32 × Opteron 3136	10.3	0.08
12 × X5670	6.6	0.07

^aFor details on the hardware and software stack, see the text. A dash indicates lower speed than with less nodes.