As I recieved my Jetson TK1 few days ago, I downloaded some apps from Software center. I thought it would be cool to compare Tegra K1 performance to my laptop graphics performance. For testing I used “Sauerbraten” ,free game compatible with ARMv7 and x86.
Tegra K1 configuration:
GPU forced (?) 852 mhz peak (should be) 326 gflops
CPU forced 2.32ghz (I havent forced all cores online, usually 2 are used)
L4T stock release I think 19.2 armhf (How do I update to 19.3? Help)
Both maxed out graphics, same amount of bots, same map (most demanding map)
i7+GT 540M 100 fps min. 190 fps max avg 160 fps
A15+GK20A 15fps min. 59 fps max. avg 30 fps
Another benchmark: Globs
On higher reolutions in most benchmarks Jetson TK1 performed better.
average difference in percent is similar to this: 230 fps vs 190 fps
Why is performance difference in some applications and games so large? Is it because of Jetson’s BETA driver performance is so bad for some things?
Is it because i7 is much more powerful? Can anyone please explain me why is higher spec’d GPU so much slower in 3d game(s) than its lower spec’d competitor?
Just speculating, consider that the i7 is 64 bit, the ARM A15 32 bit. The i7’s video card also has more ram, including dedicated video ram. Dedicated video ram is much much faster. The Jetson has only 2 GB of ram, and video shares with that ram…any video ram is basically cutting from system ram.
Everything to do with memory is at a severe disadvantage on the Jetson, with neither the A15 cores nor the Kepler at fault…you’d need to at least have the same video ram get remotely close to anything but an “apples and oranges” comparison.
If CUDA cores matter, your GT 540M has far more cores…compare it to a tablet or smart phone though, and Jetson will probably run circles around them.
What you said is very likely true. Do you maybe know how much video ram does Tegra K1 allocate?
IF CUDA cores matter my GT 540M has 96, Tegra K1 has 192 so there it has big advantage. I just dont realize how those UE4 demos on Tegra K1 work so much better than what I tested on my laptop. I would really want to try them on Jetson but cant find ARM compatible packages of UE4 and other demos that Nvidia ran on Jetson to show of TK1.
I don’t know how much ram is being taken and used…there is probably a listing of some sort in either /proc or /sys. Within /proc/meminfo is a listing of various memory uses, and I see a couple of NvMapMem*, but I don’t know how this relates to video usage, OpenGL, etc. A big hit though is from fetching and storing from system memory rather than dedicated “really fast” (that’s a technical term! :P) video memory.
CUDA cores are completely unused for applications not programmed for this. The CUDA samples on Jetson are compiled to use those cores…the cores on the laptop are likely completely unused, at least in the sense of using them directly the way the Jetson uses them. Something like OpenGL would use those cores (the OpenGL driver which is not a software emulation is programmed to use the cores), but in different ways. Apples and oranges comparison.
If you compiled the testing apps directly on your laptop using CUDA, then the comparison of core count matters…any game out there is very unlikely to be compiled this way…in which case apples and oranges.
Since Jetson is to allow experimentation with new CUDA programming/hardware mixes, it does what it is supposed to. There may come a time when the K1 is combined with dedicated very fast video ram, and does not have to leach off of system memory, but this will add a significant dollar cost. When that day comes, people will be ready to go with their software because Jetson led the way…Jetson itself will never be available as a commercial product (it isn’t even FCC type accepted, purchasing Jetson requires acknowledging this).