By the way, this is why extrapolating TFLOPs from a Steel Nomad Light benchmark is a bad idea.
Here are the TFLOPs and average Steel Nomad Light benchmark scores for various GPU's in that range. Notice that an R9 285 has a similar score to a GTX 1050ti, despite the R9 285 being a 3.25 TFLOPs card and the GTX 1050ti being 2.2 TFLOPs.
| GPU | TFLOPS | Benchmark |
| RX 560 | 2.611 | 1769 |
| GTX 950 | 1.825 | 1708 |
| Radeon 7870 | 2.56 | 1530 |
| R9 270x | 2.688 | 1739 |
| GTX 1050 | 1.862 | 1830 |
| Radeon HD 7950 | 2.867 | 1958 |
| GTX 1630 | 1.828 | 1966 |
| GTX 960 | 2.413 | 2180 |
| Radeon R9 285 | 3.29 | 2434 |
| Radeon HD 7970 | 3.789 | 2253 |
| Radeon R9 380 | 3.476 | 2492 |
| GTX 1050ti | 2.138 | 2312 |
In fact if you do a simple linear regression on these 12 data points (with Steel Nomad being the dependent variable), you get an R^2 of only about .287, and an insignificant p-value (assuming .05 alpha.) I hope this table also shows why TFLOPs aren't a good measure of performance, even for synthetic benchmarks like Steel Nomad Light. Even still, using the equation from that simple linear regression you get something like a predicted 2.825 TFLOPs +/- 0.4 TFLOP given a Steel Nomad score of 2205.
But really we don't need to do this to know what SW2's TFLOPs are. Again, TFLOPs are exactly a function of max clock rate and core count. We know both of these.
Docked FP32 TFLOPs = 1536 Cores * 1.007 Ghz x 2 ~ 3.09 TFLOPs.
Handheld FP32 TFLOPs = 1536 Cores * 561 Mhz x 2 ~ 1.72 TFLOPs.
Last edited by sc94597 - on 08 March 2026
