How to Configure CPU Energy-Performance Bias in Linux Using x86_energy_perf_policy

Modern Intel and AMD processors feature hardware-level power-management registers that define the trade-off between execution throughput and electrical power draw. While standard CPU frequency governors (such as performance or powersave) adjust clock speeds via P-states, the Energy-Performance Bias (EPB) register directly controls internal CPU throttling algorithms, turbo boost aggressiveness, uncore interconnect frequencies, and autonomous power gating. In Linux systems, the kernel provides access to Model-Specific Register (MSR) 0x1B0 (IA32_ENERGY_PERF_BIAS) through the user-space utility x86_energy_perf_policy. Configuring EPB allows administrators to fine-tune low-latency workloads or maximise energy efficiency on server and workstation hardware.

Understanding Energy-Performance Bias (EPB) Architecture

The Energy-Performance Bias architecture operates on a scale from 0 (maximum raw performance) to 15 (maximum energy conservation). In modern processors supporting Intel Speed Shift (HWP – Hardware P-States), the EPB policy acts as a hint to the embedded Power Control Unit (PCU) microcontroller. The PCU uses this value to balance frequency ramp-up latency against package thermals:

• Performance (0): Uncompromised execution speed. The PCU maintains maximum uncore ring bus speeds, aggressively scales up turbo frequencies on light-thread bursts, and delays downclocking idle logical cores.
• Balance-Performance (4 – 6): Recommended default for enterprise compute nodes. Optimises throughput for bursty computational loads while allowing power gates to engage when threads stall on memory I/O.
• Normal / Default (6 – 7): Factory default balance between energy efficiency and responsive computing.
• Balance-Power (8): Favours thermal dissipation and reduced energy consumption; limits sustained turbo durations to minimise fan noise and power draw.
• Power (15): Maximum power conservation. Disables opportunistic turbo headroom and accelerates sleep-state transitions.

Installing x86_energy_perf_policy on Linux

The x86_energy_perf_policy binary is distributed as part of the Linux kernel source tools (linux-tools) and requires the msr kernel driver to access hardware MSR registers.

Install the required packages across standard distributions:

# Ubuntu and Debian
sudo apt update && sudo apt install -y linux-tools-common linux-tools-generic linux-tools-$(uname -r)

# Fedora, RHEL, and Rocky Linux
sudo dnf install -y kernel-tools

Ensure the MSR driver is loaded into the kernel ring buffer:

sudo modprobe msr
lsmod | grep msr

Inspecting Current Core Bias Policies

Run the utility without parameters to inspect the current EPB values across every physical and logical core in the system:

sudo x86_energy_perf_policy -r

The command reports per-package, per-die, and per-core energy policies alongside HWP autonomous settings:

cpu0: EPB 6
cpu0: HWP_REQ: min 1 max 48 des 0 epp 128 window 0 pkg 0
cpu1: EPB 6
cpu1: HWP_REQ: min 1 max 48 des 0 epp 128 window 0 pkg 0

To inspect a concise summary of the system-wide policy without core-by-core verbosity, omit the raw flag:

sudo x86_energy_perf_policy

Configuring System-Wide Energy-Performance Bias

To change the bias policy across all CPU cores simultaneously, specify one of the standard policy profiles or pass an explicit integer value (0–15):

# Apply maximum performance bias for low-latency compute
sudo x86_energy_perf_policy performance

# Apply balanced performance bias
sudo x86_energy_perf_policy balance-performance

# Apply power saving bias
sudo x86_energy_perf_policy balance-power

# Pass a custom numerical register value (e.g., bias level 2)
sudo x86_energy_perf_policy 2

Verify that the new values are active across all cores:

sudo x86_energy_perf_policy | head -n 10

Configuring Per-Core Asymmetric Energy Policies

On hybrid microarchitectures (such as Intel Alder Lake, Raptor Lake, or Arrow Lake) combining Performance-cores (P-cores) and Efficient-cores (E-cores), administrators can enforce asymmetric policies. For example, configure P-cores (CPU 0–7) for peak performance while setting background E-cores (CPU 8–15) to balance-power:

# Assign maximum performance to primary execution cores
sudo x86_energy_perf_policy -c 0,1,2,3,4,5,6,7 performance

# Assign power conservation to background efficiency cores
sudo x86_energy_perf_policy -c 8,9,10,11,12,13,14,15 balance-power

Persisting EPB Policies Across System Reboots

Because MSR values reset to hardware firmware defaults upon system reboot or ACPI resume, persist custom energy policies using a dedicated systemd one-shot unit:

sudo nano /etc/systemd/system/cpu-energy-bias.service

Insert the following service specification:

[Unit]
Description=Apply CPU Energy-Performance Bias Policy
After=multi-user.target

[Service]
Type=oneshot
ExecStartPre=/sbin/modprobe msr
ExecStart=/usr/bin/x86_energy_perf_policy performance
RemainAfterExit=yes

[Install]
WantedBy=multi-user.target

Reload systemd, enable the service, and activate it immediately:

sudo systemctl daemon-reload
sudo systemctl enable --now cpu-energy-bias.service
systemctl status cpu-energy-bias.service

Employing x86_energy_perf_policy gives Linux system administrators deterministic control over hardware MSR registers, ensuring CPU cores deliver optimal performance or thermal efficiency tailored to specific workload profiles.