| Top500 | Computer | Manufacturer | Peak | LINPACK | HPCG | HPCG/Peak |
|---|---|---|---|---|---|---|
| Rank | Name/Location | [TFlop/s] | [TFlop/s] | [TFlop/s] | [%] | |
| 61 | Saturn | Nvidia | 4897 | 3707 | 67.4 | 1.38 |
| 72 | Jean Zay | HPE | 4867 | 3054 | 52.6 | 1.08 |
| 65 | HPC2 | IBM | 4605 | 3188 | 48.9 | 1.06 |
| 74 | Tetralith | ClusterVision | 4335 | 2969 | 65.2 | 1.50 |
| 49 | ECMWF | Cray | 4249 | 3945 | 65.5 | 1.54 |
| 50 | ECMWF | Cray | 4249 | 3945 | 65.5 | 1.54 |
| 71 | Circe | Nvidia | 4241 | 3057 | 199.2 | 4.70 |
| 73 | Mistral | Bull | 3963 | 3011 | 44.1 | 1.11 |
| 90 | Occigen2 | Bull | 3570 | 2495 | 45.5 | 1.27 |
| 66 | SORA-MA | Fujitsu | 3481 | 3157 | 110.2 | 3.17 |
The data of the previous page are given in a tabular form.
The exceptional efficiency of Circe (Nvidia) seems to be an error in the list Top500: Circe (71th in Top500, June 2019) is mentioned in Nvidia Circe Reference Architecture page as "The Circe supercomputer ... is now part of the NVIDIA DGX SATURNV supercomputer". The latter system is 22nd in Top500, June 2019, and the HPCG performance value (the column R of the Top500 list) is 199.196 TFlop/s both for the DGX SuperPOD (SATURNV) and Circe, which suggests that the HPCG data should not be attributed to Circe (J. Dongarra asked for comments). The HPCG efficiency of the DGX SuperPOD (11209.114 TFlop/s peak, 199.196 TFlop/s HPCG) is 1.78 %, i.e., well in the usual range 1.5-2 %.
Thus, the only exceptional value in the present list exhibits SORA-MA of Fujitsu with processors SPARC64 XIfx; the Fujitsu SPARC line of processors is known for a good (low) ratio of the peak performance and the memory bandwidth giving good HPCG efficiency (the earlier members of the SPARC64 line are used in K-computer that has the top HPCG efficiency 5.34 % among all Top500 supercomputers).