Tiger Lake + Renoir On Ubuntu Linux For Battery vs. AC Performance

[StefanBruens]

So "race to sleep" is a common trend now, where you fire the CPU core up to a high-speed, get the work done quickly, and then park the cores again. But conversely, if "getting the work done quickly", means you're slamming the CPU cores well out of their favorable V/F range and running up against the maximum power limit, you're probably going to absolutely demolish battery life.

If you look at the CPU alone, the efficiency becomes lower at higher clock speeds. When you look at the system power higher clocks speeds may be still favorable, especially for a laptop. When you are waiting for your computer to finish, your display is likely just wasting energy.

I'd like it if the Phoronix Test Suite could calculate the total energy consumed (milliwatt-hours) from each machine to complete the test. Maybe turn the screens off or set them to the same brightness to try and null out extraneous power draws. That would tell the real story.

If Intel can complete the test with their super boosting technology while consuming less overall power than the AMD system, then they are in the right. If their system is doing this while consuming significantly more total energy, well, then they have nothing.

Yes exactly (well, even comsuming the same amount of energy while being faster is still favorable).

What is totally missing here is a look at the actual efficiency. What did you buy the laptop for - being able to do as much work as possible in the shortest possible time. This of course depends on the workload, e.g.:

1. Office work (presentation, trivial tables, i.e. mostly idle CPU with active display. CPU should ramp up fast and the idle system power is important. Likely more or less single-threaded
2. Interactive development (compiling, debugging, etc.). Idle power is still important. CPU likely does not run into power/thermal limit. Scales over many cores. Energy consumed is about 50% Idle vs Busy (10% of the time at 10 times the power).
3. Heavy batch jobs, like rendering. Runs probably long enough for the display to go off or at least dim. Thermal throttling becomes relevant. How many scenes can be rendered with a single battery charge? How long does it take? There obviously is a trade off between speed and energy efficiency (i.e. 80% of the speed at 70% of the power due to lower voltage).

Looking at the (very superficial) power figures, Intel still works very well for the first two cases. Idle power is significantly lower. The average power is 15% higher, as is the performance, so roughly the same efficiency (effective work per amount of enery).

[Slartifartblast]

Tell you what son. Why don't you make your next shit post a public condemnation of the owner of this site for originally posting a detailed benchmark article on Apple's M1. Oh, yeah, condemn his SECOND article detailing how he added the M1 to his benchmarking suite.

Piss off junior. My statement stands as factual. x86 CPUs will NEVER achieve power per watt performance of ANY high end ARM SoC much less the M1.

Hmm, that's the face of a bulldog licking a piss soaked nettle.

I assume you were addressing me, are you capable of using the quote button or does it not work on your M1 test unit ?

[skeevy420]

Hello,

maybe someone commented it, but: https://arstechnica.com/gadgets/2020...lay-heres-why/

Br,

That's what I came to post as well. Don't know about you but I'm pretty sure the Ryzens would perform pretty damn well if they weren't capped at almost half the power or if AMD had configured them to keep on truckin with temps above 90C.

That said, if a person was to treat this laptop like a portable workstation that really only lives on battery power during the up to an hour between the office and home then those results aren't that bad. I still think we're in the shenanigans level of bullshit with the configurations that Intel was using, but having the option of not being battery throttled because you won't be off the AC for very long is actually a feature that I like.

I'd appreciate it if Intel would configure the Ryzens as close as possible and then re-run their tests. I'd then be curious which one has the best battery life with the conservative and performance modes in effect.

[skeevy420]

starshipeleven will be back from vacation soon. You are giving him a lot of raw material to work with.

I can't find any links to the GFLOPS/watt rating for the M1. Do you know if anyone independent has run a LINPACK or other FP tests on the M1 to see what its actual PPW rating is?

I'm glad you wrote that. I noticed he hasn't been posting lately and was hoping nothing bad happened.

[edwaleni]

I'm glad you wrote that. I noticed he hasn't been posting lately and was hoping nothing bad happened.

I posted something with his hashtag a few days ago which alerted him. He came on to say he was on vacation and don't expect anything for the duration. As for Jumbotron, he said it was up to all of us to fact check him until he returns. I still can't find any LINPACK results for the Apple M1 with the PPW metrics. According to the Green500, a system with an Epyc 7741 and NVidia GPU's have the overall highest TFLOPS per dollar as of November 2020.

[Guest]

So Ryzen performance is the same/more predictable regardless of % of the battery, while in Intel the the performance degrades when below 50% of battery?

If so, the argument of Intel saying that AMD reduces performance to reach the same battery life applies more to them than AMD, as Ryzen keeps the same perf throughout the remaining battery

Well no, as battery charge level drops, the voltage it can supply drops as well which does affect what frequencies the CPU can run at. Same reason why mobile phones shut down cores and/or throttle CPU frequencies when their battery charge decreases.

Only reason Ryzen isn't as affected by battery charge, is because their frequencies are already low to begin with. Given that they have higher core/thread count though, it's worth checking how the active core/thread count changes as battery charge decreases (and tests are run), as well as the frequency graph per core as battery charge decreases.

[Guest]

Whoah!
I just ran Geekbench5 on my Lenovo X13 with a Ryzen 4750U Pro and an Ubuntu 20.10.

Gives me
on battery: 825/4968
on AC: 1201/5665

Never thought that the impact would be that big - particularly with the single core performance where I assume the max frequency is seriously capped.

When comparing the figures to my Ryzen 3600 with a single core score of around 1300 or my 3700X (1350) this is really impressive.

Yeah, I've been reading about Zen 2 clock behaviour, and their turbo boost clock behaviour is a hot mess. While base clocks are guaranteed as long as cooling and power requirements are met, boost clocks are an absolute random die roll.

While with Intel, the Turbo Boost 2.0 frequencies are guaranteed as long as the CPU's cooling and power requirements are met.

[Guest]

So "race to sleep" is a common trend now, where you fire the CPU core up to a high-speed, get the work done quickly, and then park the cores again. But conversely, if "getting the work done quickly", means you're slamming the CPU cores well out of their favorable V/F range and running up against the maximum power limit, you're probably going to absolutely demolish battery life.

Not necessarily, "do it quickly and go back to sleep" can consume the same or less amount of power as "do it slow and steady". Both approaches will demolish battery power, one gets you a faster result.

[Paradigm Shifter]

Yeah, I've been reading about Zen 2 clock behaviour, and their turbo boost clock behaviour is a hot mess. While base clocks are guaranteed as long as cooling and power requirements are met, boost clocks are an absolute random die roll.

While with Intel, the Turbo Boost 2.0 frequencies are guaranteed as long as the CPU's cooling and power requirements are met.

The thing that most stands out for me with how my 3900X behaves is how even minor changes in load type result in wild swings in temperatures. It spends most of its life at full load crunching FFTs, and will stabilise on what I consider to be acceptable temperatures, but rsync kicks in in the background and for a few seconds temperatures (and core clocks) go nuts before it all works itself out again.

[Guest]

The thing that most stands out for me with how my 3900X behaves is how even minor changes in load type result in wild swings in temperatures. It spends most of its life at full load crunching FFTs, and will stabilise on what I consider to be acceptable temperatures, but rsync kicks in in the background and for a few seconds temperatures (and core clocks) go nuts before it all works itself out again.

What kind of all-core frequencies do you get, and does unlocking power limits in the UEFI improve that situation? (assuming sufficient cooling).