The Importance Of Thermald On Linux For Modern Intel Tiger Lake Laptops

[phoronix]

Phoronix: The Importance Of Thermald On Linux For Modern Intel Tiger Lake Laptops

Most Linux distributions including the likes of Ubuntu and Fedora have been shipping Intel's Thermald daemon the past few years as it's important to achieving good thermal/power behavior on modern Intel SoCs. For those curious about its impact, here are some benchmarks carried out with Intel Thermald or not when using an Intel Core i7 1185G7 Tiger Lake notebook.

https://www.phoronix.com/vr.php?view=30349

[chuckula]

Semi-related but has anybody noticed that the pstate driver doesn't work on Tiger lake? Even in the 5.13 kernel series there is no TGL support.

[perpetually high]

Interesting, especially in cases where disabling improved benchmarks.

Why'd that happen though? If it was able to sustain the GPU clocks and had the thermal headroom to perform and complete the benchmarks, how come thermald couldn't figure out the optimal solution?

Also, ever since a phoronix user (Linuxxx I believe) put me on to it, I've switched from intel_pstate performance to acpi_cpufreq performance and the system's noticeably snappier and actually reaches proper boost levels. Shocker

Phoronix Test Suite v10.4.0
System Information

PROCESSOR: Intel Core i5-4670K @ 4.30GHz
Core Count: 4
Extensions: SSE 4.2 + AVX2 + AVX + RDRAND + FSGSBASE
Cache Size: 6 MB
Microcode: 0x22
Core Family: Haswell
[B]Scaling Driver: acpi-cpufreq performance (Boost: Enabled)[/B]

Give it a shot if you haven't already. Add "intel_pstate=disable" to GRUB, update-grub2, and reboot. Profit

[MadCatX]

Also, ever since a phoronix user (Linuxxx I believe) put me on to it, I've switched from intel_pstate performance to acpi_cpufreq performance and the system's noticeably snappier and actually reaches proper boost levels. Shocker

Have you done any benchmarks on this or is it just a seat of the pants feeling? I find this very suspicious because both performance governors will run the CPU at its maximum clock regardless of the load. I don't see where the difference would come from.

[perpetually high]

Have you done any benchmarks on this or is it just a seat of the pants feeling? I find this very suspicious because both performance governors will run the CPU at its maximum clock regardless of the load. I don't see where the difference would come from.

I hear you, no I've actually done some unofficial benchmarks and they back up what my eyes are seeing also.

I even looked up many resources online, acpi_cpufreq performance is the business. Give it a shot! I've been very pleased. I know I throw around "this tweak made X faster" but all of these little damn tweaks have made my machine faster than your nephew's new Ryzen or Intel chip. And I'm not even trying to be funny.

ctx_clock:
pts/ctx-clock-1.0.0
Test 1 of 1
Estimated Trial Run Count: 3
Estimated Time To Completion: 1 Minute [08:12 PDT]
Started Run 1 @ 08:12:11
Started Run 2 @ 08:12:16
Started Run 3 @ 08:12:20

Context Switch Time:
142
142
142

Average: 142 Clocks
Deviation: 0.00%

Comparison to 3,645 OpenBenchmarking.org samples since 18 February 2019; median result: 255. Box plot of samples:
[ |*-----------------------------------------------####*#*#*##!*#* ]
This Result (83rd Percentile): 142 ^
Intel Celeron J4115: 3201 ^ AMD Ryzen 5 2600: 256 ^
AMD A10-7870K Radeon R7 12 Compute Cores 4C: 436 ^
AMD FX-8370E: 533 ^
AMD FX-6300: 658 ^

I challenge anyone to beat that!

edit: $ mkdir tmp && cd tmp && git clone https://github.com/phoronix-test-sui...test-suite.git && cd phoronix-test-suite && ./phoronix-test-suite benchmark pts/ctx-clock-1.0.0

[coder]

I wonder if they could somehow dynamically determine the system is more GPU-limited and favor it. Maybe it needs visibility into the work queue statistics for the GPU?

[MadCatX]

ctx_clock:
pts/ctx-clock-1.0.0
Test 1 of 1
Estimated Trial Run Count: 3
Estimated Time To Completion: 1 Minute [08:12 PDT]
Started Run 1 @ 08:12:11
Started Run 2 @ 08:12:16
Started Run 3 @ 08:12:20

Context Switch Time:
142
142
142

Average: 142 Clocks
Deviation: 0.00%

Comparison to 3,645 OpenBenchmarking.org samples since 18 February 2019; median result: 255. Box plot of samples:
[ |*-----------------------------------------------####*#*#*##!*#* ]
This Result (83rd Percentile): 142 ^
Intel Celeron J4115: 3201 ^ AMD Ryzen 5 2600: 256 ^
AMD A10-7870K Radeon R7 12 Compute Cores 4C: 436 ^
AMD FX-8370E: 533 ^
AMD FX-6300: 658 ^

I challenge anyone to beat that!

If you insist.

ctx_clock:
pts/ctx-clock-1.0.0
Test 1 of 1
Estimated Trial Run Count: 3
Estimated Time To Completion: 1 Minute [17:44 CEST]
Started Run 1 @ 17:44:07
Started Run 2 @ 17:44:12
Started Run 3 @ 17:44:17

Context Switch Time:
119
119
119

Average: 119 Clocks
Deviation: 0.00%

The chip is a rather ordinary mobile Zen 2 with no "secret" tweaks and tricks applied. Even all Spectre mitigations are switched on.

PROCESSOR: AMD Ryzen 7 PRO 4750U @ 1.70GHz
Core Count: 8
Thread Count: 16
Extensions: SSE 4.2 + AVX2 + AVX + RDRAND + FSGSBASE
Cache Size: 512 MB
Microcode: 0x8600106
Core Family: Zen 2
Scaling Driver: acpi-cpufreq schedutil (Boost: Enabled)

OPERATING SYSTEM: Arch Linux
Kernel: 5.12.15-arch1-1 (x86_64)
Desktop: KDE Plasma 5.22.3
Display Server: X Server 1.20.12 + Wayland
Compiler: GCC 11.1.0 + Clang 12.0.1 + LLVM 12.0.1
Security: itlb_multihit: Not affected
+ l1tf: Not affected
+ mds: Not affected
+ meltdown: Not affected
+ spec_store_bypass: Mitigation of SSB disabled via prctl and seccomp
+ spectre_v1: Mitigation of usercopy/swapgs barriers and __user pointer sanitization
+ spectre_v2: Mitigation of Full AMD retpoline IBPB: conditional IBRS_FW STIBP: conditional RSB filling
+ srbds: Not affected
+ tsx_async_abort: Not affected

[perpetually high]

If you insist.

ctx_clock:
pts/ctx-clock-1.0.0
Test 1 of 1
Estimated Trial Run Count: 3
Estimated Time To Completion: 1 Minute [17:44 CEST]
Started Run 1 @ 17:44:07
Started Run 2 @ 17:44:12
Started Run 3 @ 17:44:17

Context Switch Time:
119
119
119

Average: 119 Clocks
Deviation: 0.00%

The chip is a rather ordinary mobile Zen 2 with no "secret" tweaks and tricks applied. Even all Spectre mitigations are switched on.

PROCESSOR: AMD Ryzen 7 PRO 4750U @ 1.70GHz
Core Count: 8
Thread Count: 16
Extensions: SSE 4.2 + AVX2 + AVX + RDRAND + FSGSBASE
Cache Size: 512 MB
Microcode: 0x8600106
Core Family: Zen 2
Scaling Driver: acpi-cpufreq schedutil (Boost: Enabled)

Not cool, dude.

edit: “And I took that personally”

[stormcrow]

I hear you, no I've actually done some unofficial benchmarks and they back up what my eyes are seeing also.

I even looked up many resources online, acpi_cpufreq performance is the business. Give it a shot! I've been very pleased. I know I throw around "this tweak made X faster" but all of these little damn tweaks have made my machine faster than your nephew's new Ryzen or Intel chip. And I'm not even trying to be funny.

ctx_clock:
pts/ctx-clock-1.0.0
Test 1 of 1
Estimated Trial Run Count: 3
Estimated Time To Completion: 1 Minute [08:12 PDT]
Started Run 1 @ 08:12:11
Started Run 2 @ 08:12:16
Started Run 3 @ 08:12:20

Context Switch Time:
142
142
142

Average: 142 Clocks
Deviation: 0.00%

Comparison to 3,645 OpenBenchmarking.org samples since 18 February 2019; median result: 255. Box plot of samples:
[ |*-----------------------------------------------####*#*#*##!*#* ]
This Result (83rd Percentile): 142 ^
Intel Celeron J4115: 3201 ^ AMD Ryzen 5 2600: 256 ^
AMD A10-7870K Radeon R7 12 Compute Cores 4C: 436 ^
AMD FX-8370E: 533 ^
AMD FX-6300: 658 ^

I challenge anyone to beat that!

edit: $ mkdir tmp && cd tmp && git clone https://github.com/phoronix-test-sui...test-suite.git && cd phoronix-test-suite && ./phoronix-test-suite benchmark pts/ctx-clock-1.0.0

You do realize you're functionaly comparing apples to oranges then asking people to come up with better lemonade, right?

[perpetually high]

You do realize you're functionaly comparing apples to oranges then asking people to come up with better lemonade, right?

Absolutely. Controversy is my middle name

only mentioned it though because of the p_state mention on the first comment. But yes this is not to be taken serious about the benchmarking. Just bragging rights amongst real men (and real women, too)