BCLK Overclocking Attempt 1

Because my "Any Old PC" has the lowest-spec AM4 chipset (A320) I have very few options for increasing the single-thread throughput on my 5700X CPU. For this week's experiment I decided to see what I could achieve by increasing the motherboard's base clock (BCLK) above the standard 100MHz.

BCLK is a very crude way of overclocking a PC because it will force most components attached to the motherboard to operate at the increased speed. The overclocking might be a failure because even if the CPU and memory are stable at the increased speed, other components like SSDs might fail. When I read of people with a stable BCLK overclock, they are normally only running at a very small increase such as 103MHz.

However, for my experimentation I wanted to have a clear demonstration of the potential for speed improvement via BCLK overclock, so I ran my benchmarks under four configurations, two of them with a quite unsafe 10% overclock to 110MHz BCLK:

1. Configuration 1 was at 110MHz BCLK "Fullspeed", allowing the CPU and RAM to also run 10% faster.

2. Configuration 2 was at 110MHz BCLK but with the CPU and RAM speeds "Moderated" back down by 10% (or as close as I could get) such that they would operate at their normal speeds despite the higher BCLK.

   1. CPU Multiplier was reduced from default 3.4 -> 3.1 and memory multiplier was reduced from 32 -> 29.33.

3. Configuration 3 was the default 100MHz BCLK but with just the CPU and memory multipliers increased by 10%. This configuration was chosen to try and isolate whether any improvements from the 110MHz BCLK was simply due to the CPU and memory running at higher speeds.

   1. CPU Multiplier was increased from default 3.4 -> 3.7 and memory multiplier increased from 32 -> 35.33

4. Configuration 4 was everything default (100MHz BCLK, default CPU multiplier, memory at standard 3200MT) to provide a performance baseline.

Part 1 - CPU Utilization

For this experiment I also performed a separate run of the mypy benchmark on each configuration and recorded CPU frequency and temperature data for the duration of each run. This was to help identify A) whether the CPUs are attempting to run at an increased speed; and B) whether the increased CPU speeds were generating more heat than my under-spec CPU heatsink could handle.

CPU Frequencies

Unfortunately we are reaching the upper limits of my matplotlib skills, but if you look carefully you should be able to see that Configuration 1 with 10% higher BCLK is maxing out at about 3.7GHz on its fastest core, and Configuration 2 with 10% higher BCLK but -10% CPU frequency is maxing out at 3.4GHz on its fastest core - its non-boosted maximum core speed.

It seems that by increasing BLCK we are activating a kill-switch that prevents the CPU's core boost features. The 5700X should be able to boost at least one core up to 4.65GHz. Indeed, both of the Configurations with default BCLK show constant frequencies of 4.65GHz on at least one core.

CPU Temperatures

This chart is especially hard to read, but essentially what you can see here is that both the 110MHz BCLK configurations are running so slowly that they barely warm themselves above 50°C, while the 100MHz BCLK configurations are working hard and warm up to normal operating temperatures of 70°-80°C.

Part 2 - Benchmark Results

mypy

The outcome of having core boost deactivated is clear in these benchmark results. The 110MHz BCLK configurations are almost 30% slower than the 100MHz configurations. Perhaps unsurprisingly, the 100MHz configuration with manually-increased CPU and Memory speeds performed about 4% better than all default speeds. This is possibly just the result of increasing the memory frequencies which were previously shown to cause incremental performance improvements on this benchmark.

(Update: in a more recent experiment increasing the memory speed only achieved a 1% performance improvement, which indicates there was somehow an improvement here from the increased CPU base frequency even though the single core boost remained the same.)

Other Benchmarks

The remaining benchmarks really just show that deactivating core boost causes massive performance degradation across all benchmarks. If I can't retain core boost while changing BCLK then it will never be worthwhile.

Also, the "Single Unit Test" benchmark is missing from these results because the 110MHz "Full speed" system rebooted itself before it completed this benchmark. It seems likely that it became unstable and either Linux or the motherboard decided to reboot automatically. This is not surprising running at such a high BCLK. It is probably more surprising that the 110MHz configurations remaining stable long enough to complete so many other benchmarks.

Conclusion

This BCLK overclock experiment has been a failure, but really just because it caused the deactivation of core boot features. If I can work out how to increase BCLK without deactivating core boost then it may become worthwhile in the future.