Quantum Benchmark Zoo

← All benchmarks

Layer fidelity

System-level

IBM's scale-friendly quality benchmark: the fidelity of one full layer of simultaneous two-qubit gates across an N-qubit chain, reported per gate as EPLG.

Measures
Fidelity LF of a full layer of simultaneous two-qubit gates on an N-qubit chain, normalized per gate as EPLG = 1 − LF^(1/N_2q) (N_2q = 99 on IBM's best 100-qubit chain)
Introduced by
McKay, Hincks & colleagues (IBM) (2023)
Also known as
LF, EPLG, Error per layered gate

Layer fidelity (LF) is IBM’s benchmark for gate quality at scale, introduced by McKay, Hincks, Pritchett, Carroll, Govia and Merkel in 2023. It scores the fidelity of one full layer of simultaneous two-qubit gates across an N-qubit chain (gates firing together, as they would inside a real circuit), so crosstalk and spectator errors that isolated gate benchmarks miss count against the score. Since late 2023 it has been IBM’s headline quality metric (with CLOPS for speed), displacing Quantum Volume in the company’s device reporting.

How it works

Pick a connected chain of N qubits. The full layer of two-qubit gates along the chain cannot run all at once (interior qubits belong to two gates), so it is split into disjoint sub-layers of non-overlapping gates. Each sub-layer is measured with direct randomized benchmarking run simultaneously on all of its gates, in the spirit of simultaneous RB, so the fitted error rates include crosstalk. The measured fidelities multiply into the layer fidelity LF. For comparison across devices and sizes, IBM reports the error per layered gate, EPLG = 1 − LF^(1/N_2q); the best 100-qubit chain has N_2q = 99 two-qubit gates, so EPLG = 1 − LF^(1/99).

Strengths and limitations

LF runs at full device scale with no classical simulation, inherits RB’s resilience to state-preparation and measurement errors, and is decomposable: the same data localizes which qubit pairs drag the chain down. Because gates run simultaneously, it exposes crosstalk that per-pair benchmarks understate (McKay et al.). Its scope is deliberately narrow: one layer type on a 1D chain exercises neither full connectivity nor the depth-and-permutation structure of Quantum Volume circuits, and quoting the best 100-qubit chain makes headline EPLG a best-case figure. It also inherits direct RB’s noise assumptions.

Notable results

The introducing paper demonstrated LF on the 127-qubit Eagle device ibm_sherbrooke (EPLG 1.7 × 10^−2) and a 133-qubit tunable-coupler Heron (EPLG 6.2 × 10^−3 to 1.2 × 10^−2). IBM has reported EPLG on the best 100-qubit chain for its fleet since late 2023, and IQM has adopted it in its iqm-benchmarks suite.

Key papers

Reference implementations

  • Direct randomized benchmarking : Benchmarks random layers of native gates directly instead of compiled multi-qubit Cliffords, measuring gates as they are actually used and scaling well beyond standard RB.
  • Simultaneous randomized benchmarking : Runs randomized benchmarking on several subsystems at once and compares with isolated runs, turning the difference in decay rates into a direct measure of crosstalk and addressability.
  • CLOPS : IBM's speed benchmark: the sustained number of circuit layers per second a quantum system and its classical stack execute while running a batch of parameterized model circuits.
  • Quantum Volume : Single-number, full-system benchmark that scores a device by the largest random square circuit it can run while still generating heavy outputs reliably.