The Master Comparison Table
This table compares every video API vendor we could find published performance data for. We distinguish between different types of latency measurements because they are not equivalent. An API response time measures control plane speed. Protocol-level latency measures data plane speed. End-to-end latency includes network transit, encoding, and decoding.
| Vendor | Published Latency | Metric Type | Architecture | Source |
|---|---|---|---|---|
| V100 | 0.01ms (10µs) | Server processing (Server-Timing) | 20 Rust microservices | Gateway benchmark (Apple Silicon) |
| V100 | 220,661 RPS | Max throughput (gateway) | Apple Silicon (~1M+ Graviton4) | Gateway benchmark |
| V100 | 2.1ms p50 @ 50 conc. | Gateway latency distribution | 20 Rust microservices | Gateway benchmark (Apple Silicon) |
| V100 | 263.1ns per op | Protocol-level (pipeline tick) | 20 Rust microservices | Graviton4 benchmark |
| V100 | 68.4ns | STUN binding parse | 20 Rust microservices | Graviton4 benchmark |
| V100 | 3.63M ops/sec | Sustained throughput | 64 vCPUs, c8g.16xlarge | Graviton4 benchmark |
| Twilio | ~50ms | API response time | Proprietary | Twilio documentation |
| Daily | <100ms | Join time | Proprietary | Daily marketing materials |
| Agora | ~200ms | End-to-end latency | Proprietary, global edge | Agora case studies |
| LiveKit | Not published | No per-op latency data | Go SFU, open source | N/A |
| Zoom | Not published | No protocol-level data | C++ backend, proprietary | N/A |
| Mux | Not published | Focused on VOD/streaming | Proprietary | N/A |
| coturn | Not published | No public benchmarks | C, open source | N/A |
These numbers are not directly comparable. V100's 263.1ns is a server-side per-operation measurement. Twilio's ~50ms is an HTTP API response. Agora's ~200ms is end-to-end including network transit. Comparing them directly would be misleading. What this table shows is what each vendor chooses to publish and at what granularity.
V100 Detailed Benchmark Numbers
V100 is the only vendor in this comparison that publishes per-operation protocol-level latency. Here is the complete breakdown, now including gateway-level numbers from the March 2026 optimization pass:
Gateway Benchmarks (March 2026)
| Metric | Value | Category |
|---|---|---|
| Server Processing | 0.01ms (10µs) | Server-Timing header |
| Single Request (warm) | 0.38ms (380µs) | End-to-end localhost |
| p50 @ 50 Concurrent | 2.1ms | Latency distribution |
| p95 @ 50 Concurrent | 5.3ms | Latency distribution |
| p99 @ 50 Concurrent | 13.4ms | Latency distribution |
| p50 @ 200 Concurrent | 12.1ms | Latency distribution |
| Max RPS | 220,661 | Apple Silicon (~1M+ Graviton4) |
| Error Rate | 0% | Under full load |
| L1 Cache (DashMap) | sub-nanosecond | In-process cache |
| L2 Cache (Cachee) | 31ns | Distributed cache |
| Rate Limiter (local) | 0ns (95% hit rate) | Local token bucket |
Protocol-Level Benchmarks (Graviton4)
| Operation | Latency | Category |
|---|---|---|
| XOR Mapped Address (IPv4) | 34.5ns | Address encoding |
| STUN Binding Parse | 68.4ns | Protocol parsing |
| XOR Mapped Address (IPv6) | 125.8ns | Address encoding |
| Full Pipeline Tick | 263.1ns | End-to-end operation |
| TURN Channel Binding | 526.9ns | State management |
| STUN Integrity (HMAC-SHA1) | 664.2ns | Cryptographic |
| TURN Credential Validation | 863.0ns | Authentication |
Throughput Numbers
| Metric | Value | Hardware |
|---|---|---|
| Sustained Throughput | 3.63M ops/sec | c8g.16xlarge, 64 vCPUs |
| Pipeline Throughput | 3.61M ops/sec | c8g.16xlarge, 64 vCPUs |
| Per-Op Latency (sustained) | 0.3µs | c8g.16xlarge, 64 vCPUs |
| PQ Crypto Tests | 17/17 pass | All platforms |
| Total Test Suite | 542/542 pass | All platforms |
Categories Where V100 Has No Published Data
Credibility requires acknowledging gaps. Here are the benchmark categories where V100 does not yet have published numbers, and where competitors may have an advantage:
| Category | V100 Status | Who Leads |
|---|---|---|
| Global Edge PoP Count | Not published | Agora (200+ PoPs), Twilio (global edge) |
| Mobile SDK Latency | Not published | Agora, Twilio (mature mobile SDKs) |
| Time-to-First-Frame | Not published | Daily (<100ms join claim) |
| Max Concurrent Participants | Not published | Zoom (1,000+ in webinar mode) |
| 99th Percentile Tail Latency | 13.4ms p99 @ 50 conc. | V100 now publishes this; no other vendor does |
| Adaptive Bitrate Convergence | Not published | Mux (streaming-focused expertise) |
| Codec Support Breadth | Not published | Zoom (proprietary codecs), Twilio (wide support) |
V100's advantage is narrow and deep. We have the fastest published protocol-level latency and the highest published single-instance throughput. But we do not yet have published data for global edge performance, mobile SDK latency, or large-scale concurrent participant benchmarks. Agora, Twilio, and Zoom have years of production data in these categories that we have not yet matched.
Understanding the Different Types of Latency
Video API latency is not a single number. It is a stack of measurements, each important for different reasons:
1. Protocol-Level Latency (V100: 263.1ns)
The time for the server to process a single STUN/TURN operation from packet arrival to response ready. This determines server-side processing overhead, jitter contribution, and how many concurrent operations a single server can handle. Only V100 publishes this number.
2. API Latency (Twilio: ~50ms)
The time for an HTTP REST API call to complete. This measures control plane responsiveness — how fast you can create rooms, add participants, or query status. Important for developer experience but does not directly impact media quality during a call.
3. Join Time (Daily: <100ms)
The time from calling "join" in the SDK to the first media flowing. This includes signaling, ICE negotiation, DTLS handshake, and codec initialization. It is what users experience when they click "Join Meeting" and wait for video to appear.
4. End-to-End Latency (Agora: ~200ms)
The total time from one participant's camera capturing a frame to another participant's screen displaying it. This includes capture, encoding, network transit, server relay, decoding, and rendering. It is the most user-visible metric but includes many components outside the server's control.
5. Tail Latency (V100 now publishes this)
The 99th or 99.9th percentile latency. Average latency is less useful for real-time media because a few slow operations cause visible glitches. V100 now publishes tail latency: p99 at 50 concurrent: 13.4ms. No other video API vendor we surveyed publishes equivalent tail latency numbers.
Architecture Comparison
| Vendor | Language | Type | Open Source | GC |
|---|---|---|---|---|
| V100 | Rust (20 microservices) | Full platform | TURN crate (partial) | None |
| Twilio | Proprietary | Full platform | No | Unknown |
| Daily | Proprietary | Full platform | No | Unknown |
| Agora | Proprietary | Full platform | No | Unknown |
| LiveKit | Go | SFU | Yes | Go GC |
| Zoom | C++ | Full platform | No | None |
| Mux | Proprietary | VOD/Streaming | No | Unknown |
| coturn | C | TURN/STUN | Yes | None |
Methodology: How V100 Benchmarks Are Produced
Every V100 number in this post comes from the same benchmark methodology:
- Hardware: AWS c8g.16xlarge (Graviton4, 64 vCPUs, ARM Neoverse V2)
- Framework: Criterion.rs with 5-second warmup, minimum 100 iterations, 95% confidence intervals
- Isolation: Dedicated instance, no co-located workloads, performance CPU governor
- Reproducibility: Benchmarks are part of the CI pipeline and run on every release
- Honesty: We report median values with confidence intervals, not cherry-picked best runs
We use Graviton4 (ARM) specifically because it does not have turbo boost. x86 processors dynamically adjust clock frequency, which makes nanosecond benchmarks non-deterministic unless frequency is pinned. Graviton4 provides consistent, predictable performance that matches production conditions.
Request for the industry: We call on every video API vendor to publish per-operation protocol-level benchmarks using a standardized methodology. The current state — where some vendors publish API latency, some publish join time, some publish end-to-end, and many publish nothing — makes informed comparisons impossible for developers evaluating video infrastructure.
How to Use This Page
If you are evaluating video infrastructure for your product, here is how to read this data:
- If protocol-level performance is your priority (high-scale, latency-sensitive, many concurrent operations): V100's numbers are unmatched by any published competitor data.
- If global edge coverage matters most (users worldwide, diverse network conditions): Agora and Twilio have more published evidence of global deployment.
- If developer experience and ecosystem matter most (fast integration, many SDKs, large community): Twilio and Daily have mature ecosystems. LiveKit has a strong open-source community.
- If you need VOD and streaming alongside conferencing: Mux specializes in video-on-demand and streaming infrastructure.
- If post-quantum security is a requirement: V100 is the only vendor with published post-quantum crypto tests (17/17 passing).
- If you want open source you can self-host: LiveKit and coturn are fully open source. V100's TURN crate is partially open.
Future Updates
This page will be updated as new benchmark data becomes available from any vendor. Planned additions from V100:
- DONE: p99 tail latency numbers (13.4ms @ 50 concurrent, gateway benchmark)
- DONE: Gateway throughput (220,661 RPS on Apple Silicon, 0.01ms server processing)
- p999 tail latency numbers
- Multi-region benchmark results
- Mobile SDK end-to-end latency
- Large-scale concurrent participant benchmarks (100+, 1000+)
- Simulcast/SVC layer switching latency
If you represent a vendor listed in this comparison and would like to submit corrected or additional benchmark data, contact engineering@v100.ai. We will verify the methodology and update the table.
For deeper dives into V100's performance, read Inside V100's 263ns Pipeline Tick, Fastest WebRTC Server 2026, or Rust vs C++ for Video Servers.
Build on the Numbers, Not the Marketing
0.01ms server processing. 220K+ RPS. 263ns per protocol op. 3.63M ops/sec. 542 tests passing. Real benchmarks on real hardware. Start building with V100.
Get Started with V100