Late Winter Benchmarks

2025/04/18. Y'know, amateur astronomers used to want bigger telescopes. Now, we (also) want faster computers. I've spent this pollen season indoors working on just that. Previously on "the slow blog" you can read all about fun in overclocking my Intel CPU. Comes now upgrading the graphics card to accelerate particular routines that know how to run on Nvidia GPUs. Here are some benchmarks (four images of different sizes and bit depths running under two and soon-to-be-three configurations) to see where advantages do and don't, may and might lie:

SPECS & Caveats: The computer is an HP Envy 850-063 prebuilt (c. 2016) with an i7-5820k CPU capable of overclocking (see previous page). It's running at its factory-default speed (in this case, the "36x Default" profile in the HP Extreme Tuning Utility). It has 32GB of RAM, 6GB of which are assigned as a RAMdrive for PixInsight's swap files. There are 12 paths to the swap directory (one for each thread in the hexacore i7 CPU). I'm running PI 1.8.9-3. BlurX and NoiseX are Russell-Croman's BlurXterminator and NoiseXterminator respectively. Starnet++ is a PixInsight process. All defaults have been accepted for BlurX and NoiseX. Starnet++ is set to non-linear data; all other parms are default values. In one case, I noticed that a second invocation of BlurX ran much faster than the first. So I adopted the convention of running that routine twice and tallying only the second timing (go figure). The image source files are sitting in the C:/!pwork directory on the SATA3 system disk which is a 2TB Samsung 870 EVO SSD. This GTX 960 is an HP branded variant of an Nvidia-based graphics card supplied with the computer all those years ago. This GTX 1660 Super is also an HP variant first offered about three years later. Both graphics cards' specs appear identical to non-HP versions in terms of clock speed, RAM, chip-types, etc. If I could think of anything else that might be relavent, I'd include it.

Note that using the GTX 960 roughly triples processing speed. It's a hugely important speedup that makes all the difference in interactive workflows. Clearly it's not just the GPU processsor that determines the degree of improvement: bandwidth to move data on and off the card matters; the smallest images show the smallest improvements, etc. Anticipating the GTX 1660S: the clock speed is increased by 50% over the GTX 960; there are roughly 30% more CUDA cores; the GPU is upgraded; and the data path is broader. I'm expecting another factor of 3 and hoping for more (of course). Getting these processes down to just several seconds rather than ~1 minute would make another big difference when tweaking parms to produce improved images. I opted for a refurbished 1660S (thanks, eBay) rather than a more modern, more powerful graphics card to avoid having to upgrade the computer's power supply. The 1660S requires essentially the same power as the 960 (approximately 125W vs 120W) while other affordable options need 2-3x as much. This computer uses a standard ATX power supply, so if and when the day comes that an even more potent graphics card needs to live here, it should be relatively easy to boost the wattage. But first, the 1660S... Stay tuned.

So: the 1660S did not provide as spectacular an improvement over the 960 as I had hoped, but it's not bad, the price was right, and the exchange only took a few minutes (plus hours trying to readjust the color, but that's another story; and getting the sound back on, but that's still another story -- all is well enough). BTW, to tweak the color, the easiest way is via the Nvidia Control Panel (right click on the desktop, select the NvCP and "adjust desktop color settings" and there you go).

:: top ::