2026-08-30 · Cross-vendor disaggregated inference

NVIDIA H200 prefill → AMD MI300X decode

A 21-point Dynamo + vLLM production evaluation of heterogeneous prefill/decode using Mooncake TCP for KV-cache transfer. Qwen3-30B-A3B FP8 ran on one GPU per node with prefix caching disabled.

21/21 points passed381 requests · 0 failures202.40625 GiB KV payload
Peak · 1k / 8k
1,971.1 tok/s
concurrency 32
Peak · 8k / 8k
878.9 tok/s
concurrency 32
Peak · 8k / 1k
393.9 tok/s
concurrency 64
Model
Qwen3-30B-A3B
Instruct 2507 · FP8

Throughput scaling

Output token throughput across the requested concurrency ladder. Both long-output shapes peak at c=32; c=64 exceeds the efficient operating point for this public cross-cloud TCP topology.

Topology and runtime

client on H200 loopback → Dynamo 1.4.2 frontend → H200 Dynamo + vLLM 0.26 prefill → Mooncake 0.3.13 TCP · 8 lanes · public cross-cloud network → MI300X Dynamo + vLLM 0.26 decode
This verifies cross-vendor compatibility, but it is not a recommended production network design. The nodes were in separate clouds and VPCs, with no RDMA fabric; high-concurrency TTFT includes public-network KV transfer and queueing.

KV transfer verification

Exact token counts establish 98,304 KV bytes per input token. Independent socket-counter probes validate both input sizes; Mooncake reported zero transfer failures, failed receives, or expired KV requests.

1,024 tokens → 96 MiB

100,744,384 TCP payload bytes observed, including 0.0806% protocol overhead.

8,192 tokens → 768 MiB

805,896,192 TCP payload bytes observed, including 0.0732% protocol overhead.

217,332,056,064 bytes

Exact derived application KV payload across the complete 381-request matrix.

Complete matrix

Every row used exact input and output token lengths. TTFT and TPOT are milliseconds; throughput is output tokens per second.

ISL / OSLConcurrencyOutput tok/sMedian TTFTP99 TTFTMedian TPOTValidation
1k / 8k1125.2214.1214.17.96pass
1k / 8k2244.5355.1355.28.14pass
1k / 8k4449.7678.8679.28.81pass
1k / 8k8782.7845.02,838.69.86pass
1k / 8k161,306.61,766.55,687.011.51pass
1k / 8k321,971.15,905.711,831.114.69pass
1k / 8k641,788.618,775.634,779.631.54pass
8k / 8k1109.61,279.71,279.78.97pass
8k / 8k2209.31,911.12,442.49.25pass
8k / 8k4347.63,259.64,822.710.91pass
8k / 8k8569.56,661.614,824.912.25pass
8k / 8k16755.727,818.744,247.415.82pass
8k / 8k32878.977,376.2146,828.520.22pass
8k / 8k64809.1136,243.0324,594.145.67pass
8k / 1k1102.51,314.11,314.18.48pass
8k / 1k2180.61,904.62,431.88.70pass
8k / 1k4279.83,759.84,802.49.61pass
8k / 1k8304.68,848.916,782.010.02pass
8k / 1k16334.820,309.138,987.610.20pass
8k / 1k32351.044,365.583,639.110.18pass
8k / 1k64393.983,285.7156,061.610.78pass

Download evidence

Machine-readable performance data and the strict transfer-byte validation summary.