Hello NVIDIA team and community,
I am looking for a small, bounded benchmark on a crypto-enabled BlueField-2 using the public Mellanox/NVIDIA libPKA 2.0 API (GitHub - Mellanox/pka: Mellanox BlueField PKA support · GitHub), preferably an MBF2M516A-CEEO or another confirmed crypto-enabled model.
The objective is to measure secp256k1 point-addition and scalar-multiplication performance, not key search or processing private keys. A public-vector compatibility test has returned the expected 2G coordinates with curve coefficient a represented as p (equivalent to zero modulo p). This establishes compatibility for that tested setup only; it does not establish latency or throughput, or make a=p a documented vendor-supported representation.
Could NVIDIA or a community member provide either existing measurements or a single bounded run covering:
- ECC_ADD and ECC_MULT separately, using fixed public inputs and verifying returned coordinates.
- One hardware ring, 100 warmup operations, then 1,000 measured operations with one request in flight for latency.
- One hardware ring, 100 warmup operations, then 10,000 measured operations with up to 16 independent requests in flight for throughput, if supported by the installed API.
- Completed and verified operations per second, elapsed time, error counts, and latency statistics if available. Please distinguish host-observed latency from hardware execution latency.
- BlueField model, installed libPKA/BSP versions, execution location (DPU Arm or host), and actual ring/queue configuration. No serial numbers, credentials, private logs, or confidential details are needed.
No system patches, firmware changes, crypto-enablement changes, persistent access, or extended benchmark campaign are requested. Please review any test source before execution; if the API/configuration is unsuitable, a version note or explanation is sufficient. This post does not include a source attachment.
I saw the older closed discussion on PKA performance (Public Key Accelerator performance on Bluefield 2), but I would like to distinguish secp256k1 ECC_ADD/ECC_MULT measurements from P-256 or full ECDSA performance.
If a benchmark is not possible, is there a supported evaluation service or paid third-party lab that explicitly offers crypto-enabled BlueField-2 libPKA testing? I am not assuming that a general GPU instance or a BlueField-3 system is equivalent.
Thank you for any guidance.