Intel Efficiency Latency Control "ELC" Feature Slated For Linux 6.12

[phoronix]

Phoronix: Intel Efficiency Latency Control "ELC" Feature Slated For Linux 6.12

Last month I wrote about Intel Linux engineers working on a new Efficiency Latency Control feature for their uncore driver. This ELC option allows for adjusting the behavior of the Intel uncore for efficiency versus latency characteristics. Those Intel ELC patches to the TPMI uncore driver are now queued up for merging with the upcoming Linux 6.12 cycle...

https://www.phoronix.com/news/Intel-Uncore-ELC-For-Linux-6.12

[coder]

Hardware monitors the average CPU utilization across all cores at regular intervals. If the average CPU utilization is below a user defined threshold (elc_low_threshold_percent), the user defined uncore frequency floor frequency will be used (elc_floor_freq_khz), minimizing latency. Similarly in high load scenario where the CPU utilization goes above the high threshold value (elc_high_threshold_percent) instead of jumping to maximum uncore frequency, uncore frequency is increased in 100MHz steps until the power limit is reached."

Why doesn't it just watch the occupancy of queues feeding the interconnect fabric. As average queue occupancy increases, it can ramp up clockspeed to maintain fairly consistent communication latency.

[emansom]

"until the power limit is reached"

When will Intel and AMD learn that shooting so much voltage and amperage in such tiny traces will melt them eventually?

Keep the limits sane. 65W is plenty and 75C max temperature target is best (50C for SFF and mobile).

Need more processing power? Add more servers. KISS.

[coder]

"until the power limit is reached"

When will Intel and AMD learn that shooting so much voltage and amperage in such tiny traces will melt them eventually?

Keep the limits sane. 65W is plenty and 75C max temperature target is best (50C for SFF and mobile).

Need more processing power? Add more servers. KISS.

I guess you missed the part where the patch starts out:

"In the realm of high-performance computing, particularly with Xeon processors, ...

65 W is not a realistic target for a modern server CPU. They would have to reduce core count, significantly. By having multiple servers with 65 W CPUs, instead of fewer servers with higher-powered CPUs, it's actually less efficient, which is why the industry is pushing server CPU makers towards ever higher core counts.

Also, WTF are you talking about traces melting? These server CPUs usually have 5-year warranties and datacenter upgrade cycles are typically 3 to 4 years.

[chuckula]

I guess you missed the part where the patch starts out:

"In the realm of high-performance computing, particularly with Xeon processors, ...

65 W is not a realistic target for a modern server CPU. They would have to reduce core count, significantly. By having multiple servers with 65 W CPUs, instead of fewer servers with higher-powered CPUs, it's actually less efficient, which is why the industry is pushing server CPU makers towards ever higher core counts.

Also, WTF are you talking about traces melting? These server CPUs usually have 5-year warranties and datacenter upgrade cycles are typically 3 to 4 years.

In addition to all of that, notice how there's not a peep about GPU/AI accelerators that pull down power on a level that makes even a high-end CPU look like an LED light bulb. Somehow those don't explode all the time.