By manufacturer
Dell PowerEdge HPE ProLiant Lenovo ThinkSystem Fujitsu Primergy Supermicro IBM System x / Power Acer AltosCPU and thermal paste: recovering the performance lost to throttling.
On a server three to four years old, the factory thermal paste has almost always passed the end of its useful life. The symptom is not a crash — it is thermal throttling: the CPU runs below its nominal frequency to stay within thermal limits. A targeted intervention typically recovers 10-20% of the lost performance.
Thermal throttling is not visible from CPU% alone.
A system showing 100% CPU utilisation but with its dynamic frequency pulled down by thermal protection is doing far less real work than it appears. Diagnosis:
- BMC sensor readings (CPU temp, DIMM temp, VRM, ambient inlet): values close to the throttle threshold point to the problem.
- MSR readings (
turbostat,perf stat): core frequency consistently below base frequency under load = throttling active. - PROCHOT events in the SEL: explicit markers that the CPU has engaged thermal protection.
Below a certain severity level the work is done preventively; above a certain threshold it becomes urgent (risk of transient errors, reduced CPU service life).
Thermal paste, heatsinks, fans, fan curves.
- Thermal paste renewal with an enterprise-grade compound (e.g. Honeywell PTM7950, Arctic MX-6, Thermal Grizzly Kryonaut). Die cleaned with high-concentration isopropyl alcohol, measured application.
- Heatsink replacement where the base is oxidised, the heatpipes have degraded, or the system was configured with a heatsink unsuited to its current TDP.
- Fan replacement where fans are inadequate (noisy, drifting, stuck in "fail to high" mode).
- Fan curve calibration via BMC: default curves are often too conservative and trade performance for quiet operation. Dedicated curves for data centre workloads.
AI/GPU clusters and high-density systems.
On AI/GPU clusters (NVIDIA H100, H200, A100 on Dell PowerEdge XE, HPE Apollo, Lenovo SR670, Supermicro HGX platforms) the thermal margin is structurally tighter. GPU thermal paste, the thermal pads on VRAM and HBM, and chassis airflow are all variables that bear directly on training times. We cover this segment too, working in agreed windows so that no compute days are lost.
The questions we get most often.
How often should thermal paste be renewed on a server?
It depends on ambient temperature, load and the quality of the original compound. As a general rule, it is worth checking every 3-4 years. On systems under constant heavy load (databases, AI training, dense virtualisation), every 2-3 years. On light office workloads you can go 5+ years.
Can I replace the CPU with a more powerful one on the same socket?
Often yes, subject to: BIOS compatibility (updated microcode), power compatibility (the new CPU's TDP within the motherboard VRM limits), cooling compatibility (the existing heatsink must handle the new TDP). Dedicated page: server CPU upgrade.
On a Dell or HPE server, does renewing the thermal paste void the warranty?
On an active vendor warranty: typically yes, opening the system outside the official channel invalidates coverage for that component. On systems out of warranty there is nothing to invalidate. On systems under warranty we assess case by case before any work: it often makes sense to wait until expiry, unless the problem is serious.
How long does a thermal paste renewal take?
On single-socket systems: 1-2 hours of actual labour, plus 2-4h of stress testing afterwards. On dual-socket or GPU servers: 3-6 hours of labour. The server downtime window is agreed with the customer — typically overnight or at the weekend.