Competitive Benchmarks
The 42µs Advantage.
We benchmarked ZeroCopy against the leading MPC and KMS providers. The results expose a fundamental limit of networked signing.
The Physics of Signing
We measured End-to-End Signing Latency: the time from SignRequest(hash) to receiving Signature(r,s,v).
- MPC (Multi-Party Computation): Requires 3+ network roundtrips. Latency is bound by the slowest node and the speed of light.
- Cloud KMS (HSM): Requires an HTTP request to an HSM fleet. Latency is dominated by TLS handshakes and datacenter queuing.
- ZeroCopy (In-Process TEE): No network. The key lives in the same CPU memory space as the application logic. Latency is bound by the Elliptic Curve Math itself.
Benchmark Results (p99)
CRITICAL
Why p99 matters more than Avg: In trading, your "average" latency doesn't kill you—your worst latency does. During market crashes, network congestion spikes. This is exactly when MPC networks fail (timeout), while ZeroCopy remains constant because it doesn't leave the CPU.
Visualizing the Difference
Speed is abstract. Let's visualize a "Signing Race". Be warned: on a linear scale, ZeroCopy is so fast it might look like it didn't even run.
Latency Race Simulator
Watch the speed difference in real-time (scaled)
ZeroCopy Systems--
Turnkey--
AWS KMS--
Privy (Server)--
STARTFINISH →
Detailed Analysis
| Provider | Architecture | Trust Model | Bottleneck |
|---|---|---|---|
| ZeroCopy Systems | In-Process TEE (Nitro) | Hardware + Code | EC Math (CPU) |
| Turnkey | Cloud TEE | Provider + Hardware | HTTPS RTT |
| AWS KMS | Cloud HSM | AWS Trust | HTTPS + Fleet Queue |
| Fireblocks | MPC Network | MPC Nodes | Network RTT x 3 |
| Lit Protocol | Decentralized MPC | Threshold Network | Consensus + Net RTT |
Methodology
- Hardware: AWS c5.4xlarge (Intel Xeon Platinum 8000 series).
- OS: Amazon Linux 2 (Nitro Enclaves enabled).
- Measurement: CPU cycles (RDTSC) converted to microseconds.
- Sample Size: 1,000,000 signatures per provider.
Was this page helpful?