3 ms·

AMD 7840HS benchmarked at various TDP

11 points by sva_ 3y ago

wtallis 3y ago
So that's halfway to generating a useful performance/power curve; just missing measurements on the power axis. Geekbench is a highly variable workload with very different power consumption across its different subtests, so as the power limits are lowered the portion of the test that is affected by the power limits will gradually increase. And it's never safe to assume that the power consumption estimations used for the throttling mechanism are accurately calibrated.
- justsomehnguy 3y ago
  In a bit more useful form https://ibb.co/86J5rBr https://ibb.co/86J5rBr https://imgur.com/a/nS1pwSu https://imgur.com/a/nS1pwSu It's easier to see here what you can cap to 20W without a performance sacrifice (13% takes 7.8s per minute or 1 additional minute per each 10 minutes of 100% load all this time) and if you don't need multi-core performance much then you can go even gor 15W.
  - wtallis 3y ago
    Graphing it like that makes it more likely that an unsuspecting reader will assume the values denominated in Watts are actual power consumption rather than merely being input parameters to a control system.
- sva_ 3y ago
  I've never really done a comprehensive benchmark, but yeah this is only the TDP, which appears the only parameter that I have control over in the 7840HS. I should try to do one with actual power utilization if I can figure out how to accurately measure it. HP is not great with delivering information via ACPI, sadly. It seems like HP is a bit like Samsung - great hardware but 'ok' software. I just wish they built their BIOS in a way that the OS gets more access to it. > Geekbench is a highly variable workload Indeed, I noticed it seems to give fairly different scores based on the SSD in the system, so it doesn't appear to be a great CPU score.
a-french-anon 3y ago
1) Does this limit also contains the iGPU? A bit of mixed GPU/CPU workload (some video game, I guess) would be more than useful, then. 2) Does Geekbench use all SIMD units? I remember them being power hogs. 3) Why not a more reliable (open-source and compiled from source) benchmark? There are a lot of choices, I personally like using image/video processing stuff like ImageMagick, vips, darktable and encoders like x264/x265/aom.
- wtallis 3y ago
  AMD's iGPU is on the same power rail as the CPU cores, and the same throttling mechanisms apply to iGPU power consumption. Intel puts their iGPU on a separate power rail, but the chip's turbo power limits still apply. In both cases, the iGPU cannot draw as much power as the CPU cores so the limits have to be lowered quite a bit before they start affecting the GPU (unless you're running a workload that heavily uses both GPU and CPU).

11 points by sva_ 3y ago

So that's halfway to generating a useful performance/power curve; just missing measurements on the power axis. Geekbench is a highly variable workload with very different power consumption across its different subtests, so as the power limits are lowered the portion of the test that is affected by the power limits will gradually increase. And it's never safe to assume that the power consumption estimations used for the throttling mechanism are accurately calibrated.

justsomehnguy 3y ago
In a bit more useful form https://ibb.co/86J5rBr https://ibb.co/86J5rBr https://imgur.com/a/nS1pwSu https://imgur.com/a/nS1pwSu It's easier to see here what you can cap to 20W without a performance sacrifice (13% takes 7.8s per minute or 1 additional minute per each 10 minutes of 100% load all this time) and if you don't need multi-core performance much then you can go even gor 15W.
- wtallis 3y ago
  Graphing it like that makes it more likely that an unsuspecting reader will assume the values denominated in Watts are actual power consumption rather than merely being input parameters to a control system.
sva_ 3y ago
I've never really done a comprehensive benchmark, but yeah this is only the TDP, which appears the only parameter that I have control over in the 7840HS. I should try to do one with actual power utilization if I can figure out how to accurately measure it. HP is not great with delivering information via ACPI, sadly. It seems like HP is a bit like Samsung - great hardware but 'ok' software. I just wish they built their BIOS in a way that the OS gets more access to it. > Geekbench is a highly variable workload Indeed, I noticed it seems to give fairly different scores based on the SSD in the system, so it doesn't appear to be a great CPU score.

a-french-anon 3y ago

1) Does this limit also contains the iGPU? A bit of mixed GPU/CPU workload (some video game, I guess) would be more than useful, then. 2) Does Geekbench use all SIMD units? I remember them being power hogs. 3) Why not a more reliable (open-source and compiled from source) benchmark? There are a lot of choices, I personally like using image/video processing stuff like ImageMagick, vips, darktable and encoders like x264/x265/aom.

wtallis 3y ago
AMD's iGPU is on the same power rail as the CPU cores, and the same throttling mechanisms apply to iGPU power consumption. Intel puts their iGPU on a separate power rail, but the chip's turbo power limits still apply. In both cases, the iGPU cannot draw as much power as the CPU cores so the limits have to be lowered quite a bit before they start affecting the GPU (unless you're running a workload that heavily uses both GPU and CPU).