Skip to content

cross_provider_comparison.md

Latest commit

 

History

History
93 lines (66 loc) · 5.01 KB

cross_provider_comparison.md

File metadata and controls

93 lines (66 loc) · 5.01 KB
Raw

Cross-Provider Transport Performance Comparison

Generated: 2026-02-11

Platforms

AWS m6i.xlarge Azure DC4ads_v5 GCP c3-standard-4
CPU Intel Xeon 8375C (Ice Lake) AMD EPYC 7763 (Milan) Intel Sapphire Rapids (4th Gen Xeon)
vCPUs 4 4 4
Memory 15 GiB 14 GiB 14 GiB
Security Standard VM (none) SEV-SNP Confidential VM TDX Confidential VM
Kernel 6.1.159 (Amazon Linux 2023) 6.8.0-1044-azure-fde 6.14.0-1021-gcp
Rust 1.93.0 1.93.0 1.93.0
Date 2026-02-07 2026-02-07 2026-02-11

1. Handshake Latency (3-message, mock attestation)

Metric AWS m6i Azure DC4ads_v5 GCP c3-standard-4
p50 139.0 us 168.2 us 142.9 us
p95 139.7 us 174.7 us 148.0 us
p99 143.0 us 189.1 us 155.2 us
95% CI [138.9 .. 139.1] us [168.1 .. 168.4] us [142.7 .. 143.0] us

Takeaway: AWS and GCP within 3% of each other (both Intel). Azure ~21% slower (AMD EPYC Milan vs Intel Xeon). This measures X25519 keygen + HKDF — CPU crypto performance, not memory encryption.

2. Steady-State AEAD Latency (per send/recv round-trip)

1536B embedding (384-dim F32, MiniLM-L6-v2 output)

Metric AWS m6i Azure DC4ads_v5 GCP c3-standard-4
Plaintext duplex 20.2 us 101.1 us 84.4 us
SC p50 33.2 us 123.8 us 108.0 us
Overhead 64.1% 22.4% 27.9%

4KB activation tensor

Metric AWS m6i Azure DC4ads_v5 GCP c3-standard-4
Plaintext duplex 20.9 us 103.7 us 85.4 us
SC p50 42.1 us 139.5 us 124.8 us
Overhead 100.8% 34.5% 46.1%

384KB hidden state ([128, 768] F32)

Metric AWS m6i Azure DC4ads_v5 GCP c3-standard-4
Plaintext duplex 148.4 us 199.6 us 214.3 us
SC p50 1.85 ms 1.62 ms 1.86 ms
Overhead 1148% 709% 766%

Takeaway: AWS has the lowest absolute latency for small messages but the highest crypto overhead %. Azure/GCP have higher baseline duplex latency (memory encryption overhead from SEV-SNP/TDX) but the crypto overhead % is lower because the baseline is already higher. For large payloads (384KB), Azure is fastest — AMD EPYC Milan's memory subsystem handles large AEAD operations well.

3. Sustained Throughput (SecureChannel, unidirectional)

Payload AWS m6i Azure DC4ads_v5 GCP c3-standard-4
1536B embedding 488 MB/s 411 MB/s 403 MB/s
4KB activation 751 MB/s 618 MB/s 604 MB/s
384KB hidden 825 MB/s 725 MB/s 918 MB/s
1MB large 765 MB/s 519 MB/s 785 MB/s

Takeaway: AWS wins on small-message throughput (no memory encryption overhead on plaintext path). GCP TDX wins on large-message throughput — Sapphire Rapids' memory bandwidth advantage with TDX is smaller than SEV-SNP's on Milan. All platforms sustain >400 MB/s encrypted throughput even for the smallest payload.

4. Real-World Impact (MiniLM-L6-v2, ~100ms inference)

Component AWS m6i Azure DC4ads_v5 GCP c3-standard-4
Handshake 0.14% 0.17% 0.14%
AEAD RTT (1536B) 0.03% 0.12% 0.11%
Total overhead <0.17% <0.29% <0.25%

Bottom line: Transport crypto overhead is <0.3% of inference time on all three platforms, including real confidential VMs with hardware memory encryption (SEV-SNP, TDX). The encryption cost is negligible compared to model computation.

5. Key Observations

  1. Confidential VM memory encryption is measurable but small. SEV-SNP (Azure) adds ~4-5x to baseline duplex latency for small messages vs non-confidential (AWS). TDX (GCP) adds ~4x. This is the cost of AES-XTS memory encryption at the hardware level.

  2. Our AEAD overhead is consistent across platforms. ChaCha20-Poly1305 adds 22-64% on top of the already-encrypted baseline (for small messages). The crypto is CPU-bound and behaves predictably.

  3. Large messages are dominated by memory bandwidth, not crypto. At 384KB, all platforms converge to 700-920 MB/s encrypted throughput. The AEAD seal/open becomes a smaller fraction of total time.

  4. No platform is a clear overall winner. AWS is fastest for small messages (no memory encryption overhead). GCP TDX is fastest for large messages. Azure SEV-SNP is in the middle. Choose based on workload shape.

Notes

  • All measurements over tokio::io::duplex (in-process, no network)
  • Mock attestation (real attestation adds ~1-5ms to handshake, amortized over session)
  • Same Rust toolchain (1.93.0) on all platforms
  • AWS is a non-confidential baseline; Azure and GCP are real confidential VMs
  • Throughput differences partly reflect burst count changes between benchmark runs