Haha. Yes Jimmy, the release is coming soon.
Nuvia: don’t hold your breath
- Thread starter Cmaier
- Start date
I've been poking around and as far as I can tell there's no way to measure the memory bandwidth of a pre-compiled CPU program on the Mac. However, I'm not an expert and may be missing something. Obviously Anandtech found a way to measure it (https://www.anandtech.com/show/17024/apple-m1-max-performance-review/2), but I suspect that's for a custom program not measuring a precompiled one. So I can't really tell how much GB/CB stress my Mac's memory bandwidth. Given that, the best approach would indeed require a Qualcomm Nuvia device in hand, which obviously I don't have since they aren't available yet ... and am not ever going to have as my curiosity isn't enough to buy one, not even temporarily. If I DID have one, the solution would be rely on CB's linearity per core and its ability to run a custom number of threads. Start with the multicore test on all 12 cores, then 11, then 10, then 9, and so forth until a jump in performance over linearity is seen. This wouldn't necessarily pinpoint the memory bandwidth as the cause but you could try to track clock speed and power and maybe PCs also have the ability to track memory bandwidth utilization.
Jimmyjames
Site Champ
- Posts
- 822
- Reaction score
- 929
I just saw an old Linux command (available on homebrew) called “sysbench” which has a memory test. No idea if it will achieve what you want though.
Not really unfortunately. That has its own memory test, it won’t test the memory bandwidth of a precompiled application like Cinebench or Geekbench. I would need something like this:
Measuring the GPU’s use of memory bandwidth | Apple Developer Documentation
Check whether your Metal app correctly reads and writes to memory by measuring the GPU’s memory bandwidth.
But unfortunately that’s tooling for the GPU. However, I should just try the instrumentation package it’s a part of to see. I don’t see it CPU memory bandwidth described, but maybe it’ll have it!
Jimmyjames
Site Champ
- Posts
- 822
- Reaction score
- 929
Ok understood.Not really unfortunately. That has its own memory test, it won’t test the memory bandwidth of a precompiled application like Cinebench or Geekbench. I would need something like this:
Measuring the GPU’s use of memory bandwidth | Apple Developer Documentation
Check whether your Metal app correctly reads and writes to memory by measuring the GPU’s memory bandwidth.developer.apple.com
But unfortunately that’s tooling for the GPU. However, I should just try the instrumentation package it’s a part of to see. I don’t see it CPU memory bandwidth described, but maybe it’ll have it!
x.com
x.com
x.com
x.com
Yes, it absolutely could from a total platform perspective. DRAM and power delivery are part of the power consumption for the chip. If you want to see what happens when AMD and Intel guys do this, I have terrible news (directionally it isn’t good for Apple CPUs if we start comparing like that.)
I’ll paste Andrei’s comments on this when asked if these M4 ST numbers were reasonable (and he has the curves for M2/3):
“if they're idle-norm then yes”
(Meaning yes 11W is possible and they’d be higher if they didn’t subtract idle due to display).
Mind you he’s measured from the VRMs in their other chips. Sorry fellas but I am going to take my own views, most of Chips n Cheese, or Andrei over this zealotry.
Jimmyjames
Site Champ
- Posts
- 822
- Reaction score
- 929
I find It very strange that you are somehow trying to bring my conversations with you from twitter on here. We don’t agree on this and it seems unlikely we will. I don’t appreciate being called a zealot because I disagree with you.x.com
Yes, it absolutely could from a total platform perspective. DRAM and power delivery are part of the power consumption for the chip. If you want to see what happens when AMD and Intel guys do this, I have terrible news (directionally it isn’t good for Apple CPUs if we start comparing like that.)
I’ll paste Andrei’s comments on this when asked if these M4 ST numbers were reasonable (and he has the curves for M2/3):
“if they're idle-norm then yes”
(Meaning yes 11W is possible and they’d be higher if they didn’t subtract idle due to display).
Mind you he’s measured from the VRMs in their other chips. Sorry fellas but I am going to take my own views, most of Chips n Cheese, or Andrei over this zealotry.
I won’t be engaging with you any further.
- Posts
- 5,561
- Reaction score
- 9,080
Let’s chill out, please
Jimmyjames
Site Champ
- Posts
- 822
- Reaction score
- 929
Happy to. I just don’t appreciate being engaged with in this way. As I said. I won’t be engaging with them and appreciate if they don’t engage with me.
I’m really not even following you, we’re in the same place. I had no idea you were in this forum and you followed me before I even posted here realizing you were here.
Please stop slandering me because you’re mad about power. This is actually a lie from you, I’m not doing anything, and you posted about the 11W GB6 ST after I innocuously posted about it in the Nuvia thread after a convo. Sheesh.
Anyways, I’ll chill out if Jimmy can.
Jimmyjames
Site Champ
- Posts
- 822
- Reaction score
- 929
I did not “follow you”. That is a complete lie.
- Posts
- 5,561
- Reaction score
- 9,080
Guys, let’s just avoid the personal stuff. We are pretty chill here with the moderating, but this sort of stuff can quickly go too far. If you don’t want to engage with someone just block them.
Thanks for your cooperation.
Thanks for your cooperation.
Followed *back*. I had no idea who JimmyJames was and that’s not why i started posting here. I noticed you were the same guy. You’ll note my account has been here for a long time anyway, only recently did I decide to post. You’re not that important dude.
Yeah I’m done, sorry Cliff. Anyway I really don’t appreciate being slandered when a lot of this stuff (tech forums) is correlated. I’ll note my position is the standard position anyway of tech enthusiasts but I’ll let Jimmy be since he disagrees.
I tried running Instruments from XCode for Cinebench GPU benchmark using the GPU profiler to confirm if the GPU profiler gave the memory bandwidth information, but it crashed twice.
Running the CPU profiler on the CPU benchmark has all sorts of load and store counters, but I'm not sure I can turn that into memory bandwidth information easily, especially across all cores. It didn't crash though!
Basically the information might be there but sadly I'm not good enough to access it. I might try again at some point ... but ...
exoticspice1
Site Champ
- Posts
- 320
- Reaction score
- 123
I agree.
I very much do not want to continue an argument between members of this forum, but if I could piss everyone off and offer my own controversial opinion:
All of these power level measurements are valid and have utility ... except CPU manufacturer "TDP" ratings, those are largely fictional.
It depends on what you actually want to measure and who you want the measurement to matter to: wall power measurements, powermetrics, HWInfo, VRMs and power lines, hell even runtime to shutdown for laptops/mobile, etc ... and indeed for laptop/mobile devices with batteries and screens: measurements made with or without batteries, with or without screen power ... EDIT: and for all computers yes with or, yes, without DRAM (especially for on-package vs modular systems) - these all can provide useful information. Except CPU "TDP" ratings given by the manufacturer (GPUs weirdly seem okay), I think everyone can agree by now that those are often mostly fictional and sometimes completely fictional. Regardless, each of these other power measurements is fine as long as the person who is reporting the power measurements is clear on the method chosen (and the circumstances of the method, e.g. test bench vs case, extra cooling, etc ...), is understanding on what the method is actually measuring (and what it is not) as well as the caveats of that method (and all of them have caveats), is clear of the goals of the measurement and the intended audience (and those goals/audience align with the method), and, if making comparisons, is as consistent as possible when comparing power measurements across different devices. This last bit is naturally where we run into the most problems - consistency across devices with vastly different hardware, software, and expected use cases can be quite a challenge.
All of these power level measurements are valid and have utility ... except CPU manufacturer "TDP" ratings, those are largely fictional.
It depends on what you actually want to measure and who you want the measurement to matter to: wall power measurements, powermetrics, HWInfo, VRMs and power lines, hell even runtime to shutdown for laptops/mobile, etc ... and indeed for laptop/mobile devices with batteries and screens: measurements made with or without batteries, with or without screen power ... EDIT: and for all computers yes with or, yes, without DRAM (especially for on-package vs modular systems) - these all can provide useful information. Except CPU "TDP" ratings given by the manufacturer (GPUs weirdly seem okay), I think everyone can agree by now that those are often mostly fictional and sometimes completely fictional. Regardless, each of these other power measurements is fine as long as the person who is reporting the power measurements is clear on the method chosen (and the circumstances of the method, e.g. test bench vs case, extra cooling, etc ...), is understanding on what the method is actually measuring (and what it is not) as well as the caveats of that method (and all of them have caveats), is clear of the goals of the measurement and the intended audience (and those goals/audience align with the method), and, if making comparisons, is as consistent as possible when comparing power measurements across different devices. This last bit is naturally where we run into the most problems - consistency across devices with vastly different hardware, software, and expected use cases can be quite a challenge.
Last edited:
thanks!
- Posts
- 5,561
- Reaction score
- 9,080
To me, as long as we are comparing chips with built-in DRAM to chips with off-package RAM, i’d like to know the number without DRAM in both cases (i.e. for both M* and its competitors), just because that tells me more about what’s going on with the chip, how it might scale in different scenarios, etc. Knowing the full package power is also useful. It just depends on what you are using the numbers for.
Yeah I meant to include that in my list of with and withouts ... especially since it was so central to the discussion. I've put it in now.
Last edited:
