3 ms·

Anandtech measured the max TDP of the CPUs in the M1 as being around 28W, with the GPU adding ~17W to that figure. https://www.anandtech.com/show/16252/mac-min

by pja 5y ago

Anandtech measured the max TDP of the CPUs in the M1 as being around 28W, with the GPU adding ~17W to that figure. https://www.anandtech.com/show/16252/mac-mini-apple-m1-tested https://www.anandtech.com/show/16252/mac-mini-apple-m1-teste...

hajile 5y ago
That figure is compute workloads. Also, those measurements are “at the wall” and include RAM, SSD, and everything else. At the very least, you must subtract most idle power, but you also need to subtract more when DRAM/SSD kicks up from low power state.
- pja 5y ago
  That figure is after subtracting idle power. It’s right there on the chart. The workload is described as "Compute MT" (i.e. multithreaded), compared with "Memory MT (i.e., memory intensive Multi-threaded workload), with a third bar on the chart being the average of the two. It seems to me that the "Compute MT" is the closest to the measurement that Intel or AMD would describe as TDP. The M1 is a ~25+W chip, not a 10W chip, if we’re going to make a fair comparison with Intel or AMD equivalents. It’s still a great CPU of course, but it’s not magic pixie dust.
  - hajile 5y ago
    Compute MT is using both CPU and GPU together. To compete, any other CPU needs to use a dedicated GPU. Also, at idle, the DRAM and SSD are powered off. Samsung 980 uses 4.5 watts of power when active. I'd guess a similar power consumption applies here. LPDDR4x ranges from 0.6v to 1.8v This should give 2-3w per 8gb chip at full-power. Anandtech's article agrees with ST Memory adding 4.2w of power over their other ST benchmarks. Now, 26.8 - 4.5 - 4.2 = 18.1w at PEAK. For just CPU loads, we get 22.3 - 4.5 - 4.2 = 13.6w.
    - pja 5y ago
      I don’t think your logic is based on sound assumptions: You can’t completely turn off DRAM at idle if it’s to store data - it has to self-refresh continuously. Some power savings are available by running it in a power saving mode, but you can’t get around the leakage of the DRAM cells, so the extra power used to drive the DRAM hard is not the full DRAM power usage, but a fraction of it. (LPDDR5 is nifty in that you can tell it to power down parts of the chip that aren’t currently being used.) The SSD isn’t being hit by either workload: why would a CPU benchmark hit disk? The SSD is going to remain idle & be part of the 4W that Anandtech assigns to the system idle power drain. Where does it say that Compute is CPU + GPU? (I’m happy to believe it, but I don’t see it in the article & they already break out the GPU power usage seperately using a different benchmark.) (It would really help if Anandtech told us what the benchmarks were - some CPU intensive benchmarks don’t hit memory at all, they just hammer a few cache lines. It’s entirely possible for a pure computational benchmark to leave most of the LPDDR5 in deep sleep!) If by MT Compute they mean specifically the Geekbench 5 compute benchmark then that is a Metal / GPU benchmark, so I will concede this point in that case :) Even conceding that the Compute benchmark is really a GPU benchmark, it’s still the case comparing a CPU benchmark load against the TDP of another processor isn’t entirely correct, as the TDP is meant to be a power ceiling (although Intel has rather breached this in recent years - AMD TDPs are more honest I’m told.) - if we accept the 22.3W of the average workload benchmark, then the TDP for the chip is still going to be more than that & that’s the appropriate point of comparison.
      - hajile 5y ago
        LPDDR4x has various power levels based on whether it's actually being used or just keeping data from corrupting. The latter is much less memory intensive and runs at much lower voltages. You may have a point about SSD depending on how much read/write is happening. Check out this comparison of the passively-cooled M1 Macbook Air vs both Intel and AMD Surface 4 machines. https://youtu.be/7yTWGjYFiC0?t=509 https://youtu.be/7yTWGjYFiC0?t=509 TL;DR -- AMD loses 25-50% of their performance as soon as the plug is removed. Their actual power usage spikes at 40w (for a 10-25w U-series CPU). No matter how you cut it, Peak M1 power usage is about half for similar performance. Insult to injury, the M1 not only has far better performance, but it's also cheaper.

by pja 5y ago

hajile 5y ago

That figure is compute workloads. Also, those measurements are “at the wall” and include RAM, SSD, and everything else. At the very least, you must subtract most idle power, but you also need to subtract more when DRAM/SSD kicks up from low power state.

pja 5y ago
That figure is after subtracting idle power. It’s right there on the chart. The workload is described as "Compute MT" (i.e. multithreaded), compared with "Memory MT (i.e., memory intensive Multi-threaded workload), with a third bar on the chart being the average of the two. It seems to me that the "Compute MT" is the closest to the measurement that Intel or AMD would describe as TDP. The M1 is a ~25+W chip, not a 10W chip, if we’re going to make a fair comparison with Intel or AMD equivalents. It’s still a great CPU of course, but it’s not magic pixie dust.
- hajile 5y ago
  Compute MT is using both CPU and GPU together. To compete, any other CPU needs to use a dedicated GPU. Also, at idle, the DRAM and SSD are powered off. Samsung 980 uses 4.5 watts of power when active. I'd guess a similar power consumption applies here. LPDDR4x ranges from 0.6v to 1.8v This should give 2-3w per 8gb chip at full-power. Anandtech's article agrees with ST Memory adding 4.2w of power over their other ST benchmarks. Now, 26.8 - 4.5 - 4.2 = 18.1w at PEAK. For just CPU loads, we get 22.3 - 4.5 - 4.2 = 13.6w.
  - pja 5y ago
    I don’t think your logic is based on sound assumptions: You can’t completely turn off DRAM at idle if it’s to store data - it has to self-refresh continuously. Some power savings are available by running it in a power saving mode, but you can’t get around the leakage of the DRAM cells, so the extra power used to drive the DRAM hard is not the full DRAM power usage, but a fraction of it. (LPDDR5 is nifty in that you can tell it to power down parts of the chip that aren’t currently being used.) The SSD isn’t being hit by either workload: why would a CPU benchmark hit disk? The SSD is going to remain idle & be part of the 4W that Anandtech assigns to the system idle power drain. Where does it say that Compute is CPU + GPU? (I’m happy to believe it, but I don’t see it in the article & they already break out the GPU power usage seperately using a different benchmark.) (It would really help if Anandtech told us what the benchmarks were - some CPU intensive benchmarks don’t hit memory at all, they just hammer a few cache lines. It’s entirely possible for a pure computational benchmark to leave most of the LPDDR5 in deep sleep!) If by MT Compute they mean specifically the Geekbench 5 compute benchmark then that is a Metal / GPU benchmark, so I will concede this point in that case :) Even conceding that the Compute benchmark is really a GPU benchmark, it’s still the case comparing a CPU benchmark load against the TDP of another processor isn’t entirely correct, as the TDP is meant to be a power ceiling (although Intel has rather breached this in recent years - AMD TDPs are more honest I’m told.) - if we accept the 22.3W of the average workload benchmark, then the TDP for the chip is still going to be more than that & that’s the appropriate point of comparison.
    - hajile 5y ago
      LPDDR4x has various power levels based on whether it's actually being used or just keeping data from corrupting. The latter is much less memory intensive and runs at much lower voltages. You may have a point about SSD depending on how much read/write is happening. Check out this comparison of the passively-cooled M1 Macbook Air vs both Intel and AMD Surface 4 machines. https://youtu.be/7yTWGjYFiC0?t=509 https://youtu.be/7yTWGjYFiC0?t=509 TL;DR -- AMD loses 25-50% of their performance as soon as the plug is removed. Their actual power usage spikes at 40w (for a 10-25w U-series CPU). No matter how you cut it, Peak M1 power usage is about half for similar performance. Insult to injury, the M1 not only has far better performance, but it's also cheaper.