←back to thread

132 points cl3misch | 4 comments | | HN request time: 0.706s | source
Show context
hu3 ◴[] No.40718245[source]
I slightly undervolted an i7 11800H by -100mV.

That resulted in a -10C average CPU temperature. Massive!

I don't remember last time I heard the fans. Not even with Docker + Jetbrains IDEs.

I wonder why doesn't it come like this by default. It runs faster too because it's no longer frequently thermal throttled.

replies(4): >>40718594 #>>40718806 #>>40719032 #>>40733881 #
1. Out_of_Characte ◴[] No.40718806[source]
Because not all chips will be stable at -100mV Intel is providing the chip with warranty/guarantee that the chip performs 'up to' spec.

Is your chip completely stable in all workloads? Would it be stable at 90c? when all Dram slots are occupied? When driving all pcie lanes at max?

I bet that you could make any chip unstable with the right tools and knowledge.

replies(1): >>40718960 #
2. hu3 ◴[] No.40718960[source]
At the time, I ran a CPU stress test which put all cores at full load for a couple of hours.

Then I ran RAM tests.

Then I ran NVMe stress tests.

Then I ran GPU stress tests (NVIDIA GPU is underclocked).

Then I ran all of these tests together.

Worked butter smooth. Not a crash to this day.

To be fair, -100mV is on the safe side according to the articles I read at the time. Some folks run at -200mV or less. I don't need that kind of tweak.

replies(2): >>40720377 #>>40743323 #
3. thot_experiment ◴[] No.40720377[source]
chips do increasingly come like that as intel gets better at binning, they don't want to leave money on the table and are creating more SKUs in order to mark up good chips like yours (this is why silicon lottery is no longer in business)

the reason your chip didn't come like that is because intel plays it very safe when it comes to stability and their margins for error are likely broader than yours, the stochastic nature of the failures means that the voltage margin between one crash for every 2 hours of stress test and one crash for every day of stress test can be hundreds of milivolts, but they're definitely working on it because pat doesn't want you getting free real estate

4. RealStickman_ ◴[] No.40743323[source]
When undervolting with an offset, most of the crashes I had were when the CPU state changed. Opening a browser and playing a video from idle, starting a transcode, stuff like that.

The only way to test is, unfortunately, to use it until it crashes.