POV-Ray Benchmarks
POV-Ray
can not just used for generating great images,
it includes a benchmark, which can be used to compare the CPU speed of real
world applications (at least if they use floating point operations a lot).
Since Debian Jessie includes POV-Ray 3.7.0 for all architectures from i586 up to a
modern 64Bit processor, it can be used to see real speed improvements,
and not just a few percent, which many websites like to report when comparing
CPUs. The time given here is the Trace Time from the POV-Ray output. The power
was measured on the primary side, so the whole power of the desktop PC
(including usual USB devices, excluding screen) or
laptop (including screen on darkest value, battery removed) was measured.
So the values can only be considered a rough estimate on efficiency.
Fine tuning power saving settings or improving the PSU would result in a lower power consumption.
All values were measured with a Voltcraft SBC-500.
Debian Stretch dropped support for the very old processors (AMD K6, Pentium I). So the POV-Ray version of Debian Jessie will be the last to use as comparison for old systems. Jessie and Stretch both deliver POV-Ray 3.7.0, however the binary is build with a newer version of gcc in Stretch. Tests on the same system showed, the speed difference is less than 5% between the two versions. So the values are still comparable.
Comparing the K6-2 CPU with the Ryzen, gives a huge speed improvement of a factor 542 within 18years. Trying to estimate which type of technology improvement contribute how much to this factor, let's first look at the 32bit multicore performance of the Ryzen. The 32bit code is the same binary as run on the K6-2, so no new CPU command set like sse2 were used. Therefore it can be assumed that a factor of 2.46 is gained by introducing new CPU commands and 64bit registers. Next, compare the multicore to single core performance, it is a factor 7.2. This fits very well to the six cores with SMT. SMT is said to give a speed improvement of about 20% (6*1.2=7.2). The CPU frequency increased from 400MHz to 3.2GHz, giving a factor of 8. Now there is a factor of 3.83 left (542/(8*7.2*2.46) = 3.83), the assumption here is, this is the optimization done over the yeas while executing code: Branch prediction, reordering, multiple execution units. Of course this is all just a guess, as there are too little information available to analyze the source of speed improvements in detail. For example, memory speed of caches and main memory and the size of the cache memory are expected to have a big influence too.
| CPU | 32bit single thread | 32bit multi thread | 64bit multi thread | Relative speed compared to 400MHz | Power consumption (during the fastest test) | Energy consumed for benchmark | Idle power consumption |
| AMD K5 PR133 100MHz | Estimated 133757s (31h + 35min) based on 3854 rendered pixels | - | - | estimated 38.5% | 52.4W (desktop) | estimated 1946Wh | 39W |
| AMD K6-2 400MHz | 51519s (14h + ~26min) | - | - | 100% | |||
| Raspberry Pi1 700MHz | 38376s (10h + ~39min) | - | - | 134% | 3.23W (desktop like) | 34.4Wh | 3.03W |
| Samsung Galaxy Trend 7580 - Dual Core Cortex A9 1,2GHz | - | 9006s (2h + ~30min) | - | 572% | jumping between 0.15W and 3.11W (smartphone battery is fully charged) | less than 7.8Wh | 0.15W |
| AMD Athlon XP 1800+ 1,5GHz | 4403s (1h + ~13min), Debian Stretch | - | - | 1170% | 120W (desktop) | 146.8Wh | 92.5W |
| Intel Atom N550 1,5GHz | - | - | 2770s (~46min) | 1860% | 12.5W (laptop) | 9.6Wh | 9.3W |
| Intel Centrino Duo T2050 1,6GHz | - | 1834s (~30min) | - | 2809% | 38.5W (laptop) | 19.6Wh | 15-17W |
| AMD Athlon X2 3800+ EE SFF 2,0GHz | - | - | 1248s (~20min) | 4128% | |||
| Intel Core2 Duo E6550 2,33GHz | - | 1112s (~18min) | 830s (~14min) | 6207% | 85W (desktop) | 19.6Wh | 49.8W |
| AMD Athlon 5150 1,6GHz | - | - | 801s (~13min) | 6432% | 40W (desktop) | 8.9Wh | 28.5W |
| AMD Athlon II X3 415e 2,5GHz (4th core unlocked) | - | - | 410s (~7min) | 12566% | 102W (desktop) | 11.6Wh | 61-65W |
| AMD Ryzen R5 1600 3,2GHz | 1687s (~28min), Debian Stretch | 234s (~4min) | 95s | 54231% | 147W (desktop) | 3.9Wh | 60.5W - 62.7W |
Debian Stretch dropped support for the very old processors (AMD K6, Pentium I). So the POV-Ray version of Debian Jessie will be the last to use as comparison for old systems. Jessie and Stretch both deliver POV-Ray 3.7.0, however the binary is build with a newer version of gcc in Stretch. Tests on the same system showed, the speed difference is less than 5% between the two versions. So the values are still comparable.
Comparing the K6-2 CPU with the Ryzen, gives a huge speed improvement of a factor 542 within 18years. Trying to estimate which type of technology improvement contribute how much to this factor, let's first look at the 32bit multicore performance of the Ryzen. The 32bit code is the same binary as run on the K6-2, so no new CPU command set like sse2 were used. Therefore it can be assumed that a factor of 2.46 is gained by introducing new CPU commands and 64bit registers. Next, compare the multicore to single core performance, it is a factor 7.2. This fits very well to the six cores with SMT. SMT is said to give a speed improvement of about 20% (6*1.2=7.2). The CPU frequency increased from 400MHz to 3.2GHz, giving a factor of 8. Now there is a factor of 3.83 left (542/(8*7.2*2.46) = 3.83), the assumption here is, this is the optimization done over the yeas while executing code: Branch prediction, reordering, multiple execution units. Of course this is all just a guess, as there are too little information available to analyze the source of speed improvements in detail. For example, memory speed of caches and main memory and the size of the cache memory are expected to have a big influence too.