CPU Overclocking: A Performance Assessment of Air, Cold Plates, and Two-Phase Immersion Cooling

Overview

Computing capacity has always climbed with improvements in semiconductor manufacturing and packaging, usually accompanied by a steep rise in the heat flux density of the electronic component (CPU or GPU). State-of-the-art air cooling throttles after a certain heat flux level and would need bigger heat sinks driving enormous airflow. The study measured the highest stable frequency, junction temperature and junction-to-ambient thermal resistance at various coolant temperatures, boiler-plate surfaces and TIMs.

Compared with air cooling, cold plates gave up to 41% and 2P immersion up to 51% higher CPU clock rates. The 2P fluid was 3M HFE-7000, with a saturation temperature of 34 °C at atmospheric pressure and a density of 1400 kg/m³, versus 1.14 kg/m³ for air and 994 kg/m³ for water at 34 °C. About 90% of the CPU's heat came from 50% of the die area, so core heat flux could exceed 220 W/cm² at peak load. With air at 34 °C inlet, maximum stable frequencies were 3.3, 4 and 3.6 GHz for Prime95, XTU and Cinebench respectively, while cold plates and 2P immersion consistently held stable frequencies above 5 GHz.

At maximum workload, the 2P immersion heat-transfer coefficient (HFE-7000 with indium TIM2) was about 110% greater than air and 50% greater than cold plates, giving the lowest junction-to-ambient resistance of 0.25 °C/W. For well-optimised systems, typical thermal resistances of cold plates and 2P immersion are comparable; cold plates are a quick enabler for hybrid-cooled high-power chips, while 2P immersion could enable 3-D packaging. Both liquid options sustained much higher clock rates than air, and two-phase immersion gave the lowest junction temperatures.

The authors stress that once cooling is good the next limit is packaging and interface resistance, and that local heat flux at the cores, not TDP, is the coming constraint; they recommend adopting liquid cooling in early chip design and architecture phases. Limits: a single consumer-class CPU (i9-9900K) and HPC-style workloads, so results do not transfer directly to server-class chips; the authors provide a generic overclocking methodology.

Key points

1. Compared with air cooling, cold plates gave up to 41% and 2P immersion up to 51% higher CPU clock rates.
2. The 2P fluid was 3M HFE-7000, with a saturation temperature of 34 °C at atmospheric pressure and a density of 1400 kg/m³ (air 1.14, water 994 kg/m³ at 34 °C).
3. About 90% of the CPU's heat came from 50% of the die area, so core heat flux could exceed 220 W/cm² at peak load.
4. With air at 34 °C inlet, maximum stable frequencies were 3.3, 4 and 3.6 GHz for Prime95, XTU and Cinebench, while cold plates and 2P immersion consistently held stable frequencies above 5 GHz.
5. At maximum workload, the 2P immersion heat-transfer coefficient (HFE-7000 with indium TIM2) was about 110% greater than air and 50% greater than cold plates, with the lowest junction-to-ambient resistance of 0.25 °C/W.
6. For well-optimised systems, typical thermal resistances of cold plates and 2P immersion are comparable; cold plates are a quick enabler for hybrid-cooled high-power chips, while 2P immersion could enable 3-D packaging.

Using it in design

• For high-heat-flux chips, compare the point at which air cooling throttles (here 3.3-4 GHz) with the clock-rate gain from cold plates or 2P immersion (41-51%) in cooling technology selection.
• Use local core heat flux (above 220 W/cm²) rather than TDP, and packaging and interface (TIM) resistance, as inputs for junction temperature targets.
• Treat cold plates as the quick option for hybrid-cooled high-power chips and 2P immersion as the future option for 3-D packaging.

Who should read it

Thermal engineers choosing among air, cold-plate and two-phase immersion cooling for high-power CPUs and GPUs.

Document contents

1. I. Introduction
2. II. Overclocking (A. Advantages of Overclocking, B. Implications of Overclocking)
3. III. Methodology (A. Benchmarking Methodology, B. Computing Workloads)
4. IV. Cooling Methods (A. Air Cooling, B. Cold Plates, two-phase immersion)
5. V. Results (A. CPU Frequency Versus Voltage, B. Junction Temperature Versus CPU Power, C. Thermal Resistance, D. Thermal Resistance Breakdown)
6. VI. Conclusion

Read the document

Publication details

Authors

Bharath Ramakrishnan; Husam Alissa; Ioannis Manousakis; Robert Lankston; Ricardo Bianchini; Washington Kim; Rich Baca; Pulkit A. Misra; Inigo Goiri; Majid Jalili; Ashish Raniwala; Brijesh Warrier; Mark Monroe; Christian Belady; Mark Shaw; Marcus Fontoura

Publisher

IEEE Transactions on Components, Packaging and Manufacturing Technology (Institute of Electrical and Electronics Engineers (IEEE))

Year

2021

Volume

11

Issue

10

Pages

1703-1715

Document language

English

Source type

Scientific

Access

Open licence — downloadable

DOI

10.1109/tcpmt.2021.3106026

Licence

CC BY 4.0 — share with attribution

Licence evidence

doi.org

Citation