Mobile ZK: When a Timed Run Leaves No Proof
An Ethereum Foundation benchmark on ten Android phones separates timed iterations, saved proofs and the work a credential app still has to do.

Concept illustration: a crowded phone workshop and empty output tray represent memory pressure and an undelivered proof. This is not a benchmark screenshot or real phone internals.
Key Points
- The Ethereum Foundation reported credential-proof benchmarks on ten Android phones. Their measurements describe specific workloads and configurations, not production-service performance.
- Both Galaxy A06 Prepare attempts completed several timed iterations before Android stopped them; neither left a proof. Its presentation test used Prepare output made on a laptop.
- Cold setup, warm medians, excluded file reads and reused blinding values must be read separately. The benchmark did not test presentation unlinkability or acceptable user waiting time.
Inside this story
Chronological Timeline
OpenAC and coordination-network credential workloads run on ten physical Android phones, from budget models to flagships.
Separate hashing and signature-check workloads add measurements; these are distinct from the full credential workloads.
The report presents the phone results and names the measurements still missing for a deployed app.
What the failed run actually says
A zero-knowledge (ZK) credential flow may let a phone prove a selected fact, such as eligibility, without disclosing every field in its credential. The Ethereum Foundation’s November 3 report asks whether that work fits the phones people own. It measured ten physical Android devices inside applications, where the operating system could terminate a memory-hungry process. These are credential-proof workloads, not Ethereum transaction-throughput tests.[1]
The Galaxy A06 result needs more than the table’s failure label. In each of two OpenAC Prepare attempts, the cold iteration and three or four warm iterations completed before Android’s low-memory killer stopped the process. Neither attempt left a proof to retrieve. Recorded work therefore did not become a completed benchmark run with an output artifact. This describes the tested configuration, not every possible implementation on that phone.[2]
Prepare and presentation are different jobs
OpenAC’s Prepare step checks and processes a credential once and saves state. A later presentation combines a Show proof with re-randomization of the Prepare and Show proofs. To test that later stage on the A06, the team supplied a laptop-generated Prepare proof. Its 6.18-second warm presentation median is consequently not evidence that the phone completed the whole preparation-and-presentation flow on its own.[1][2]
The harness, a test runner that automates the measurements, configured a fresh app process for one cold iteration plus five warm iterations, ten seconds apart. Warm results summarize the five repeats; there are six planned iterations, not five in total. For completed runs, the last proof was pulled from the phone and verified on a laptop. The interrupted A06 Prepare attempts lacked that final artifact.[2][3]
Two clocks on the same device
The coordination-network credential proof has a different structure: eight checks share one proof. On the Galaxy A06, Infinix SMART 8 and moto g05, standard-mode warm medians were 7.85, 9.11 and 9.61 seconds, respectively. Their cold totals were 13.3, 15.7 and 16.9 seconds. Cold setup includes loading the reference parameters (SRS) and circuit, then deriving the verification key. Calling all of these results an 8–10-second wait would erase that distinction.[1][3]
The two systems use different credential formats and divide preparation differently. Even a fast individual Show step omits other work in a complete OpenAC presentation. A comparison should identify which stage was timed, what state was supplied beforehand and whether loading was included. Those boundaries matter before treating a benchmark row as a promise about an app.[1][2][3]
Work that the timer left outside
The OpenAC test runner skipped re-reading the 374 MB circuit description that the mobile app described in the cited repository revision reads on every Prepare. Presentations do not read that file. The results preserved in that revision disclose this difference explicitly: their Prepare timings do not include the same repeated file-reading work as that app. They cannot simply be substituted for its end-to-end latency.[2]
The runs also reused one set of shared blinding values across presentations, while the SDK draws fresh ones. The README says the computational work is the same, but the benchmark did not test whether two presentations could be linked. Performance evidence is not an unlinkability test. This limitation establishes neither a general privacy guarantee nor a demonstrated privacy failure.[2]
The application questions still open
The report covers one set of runs per phone and workload set on two dates. It did not measure battery use, integrated application size, integration effort or the wait people would accept. Five warm repeats provide a small timing sample; they do not establish a service’s reliability across everyday background workloads or repeated user sessions.[1]
Memory figures need their own label. The reported resident-memory high-water mark, VmHWM, includes mapped key files and excludes pages compressed into zRAM; it is not the quantity Android’s killer acts on. The Infinix memory peaks are treated as a floor, and its firmware changed between the two dates. Android’s own documentation describes process killing as a response to memory pressure. A phone’s advertised RAM is not a universal per-app allowance.[1][2][4]
The cited repository revisions preserve iteration records, session metadata, trimmed logs and outputs from completed runs. They say the full logs, host runner and predictions are in a repository not yet public. The useful result is a map of what completed under stated conditions. For a product claim, the next questions are which proof was actually delivered, what preparation remained outside the timer, and which privacy and user-experience properties were tested separately.[1][2][3]
Continue reading
Explore the topic through other cases and contexts.
Sources & References
- [1]Source 1: Benchmarking Zero-knowledge Credential Proofs on Low-end Mobile DevicesEthereum Foundation — Public Sector · 2026-11-03Accessed 2026-11-06
- [2]Source 2: OpenAC Android results — pinned zkID revision 591f49aethereum/zkIDAccessed 2026-11-06
- [3]Source 3: Coordination-network Android results — pinned revision 6c27308ethereum/coordination-networkAccessed 2026-11-06
- [4]Source 4: Android low memory killer daemonAndroid Open Source ProjectAccessed 2026-11-06Back to text1


