> For clean Markdown content of this page, append .md to this URL. For the complete documentation index, see https://docs.nvidia.com/dynamo/llms.txt. For full content including API reference and SDK examples, see https://docs.nvidia.com/dynamo/llms-full.txt.
# Feature Benchmarks
Feature Benchmarks evaluate Dynamo features, topologies, and feature stacks under controlled traffic. Each page states the question, compares deployable configurations, shows how to reproduce the run, and links to the [Recipe](/dynamo/v1.4.0/recipes/browse) target when one deployment should be used directly.
## Features Under Test
Serving techniques and topology changes benchmarked across the comparisons below.
**KV routing**
Routes traffic to workers with reusable KV cache so TTFT, ITL, and goodput can improve on prefix-heavy workloads.
**Prefill/decode split**
Separates prompt prefill and token decode into specialized worker pools for long-context latency and throughput tests.
**WideEP**
Spreads MoE experts across a wider GPU set so expert-heavy requests get more parallel capacity.
**Embedding cache**
Reuses multimodal embeddings, especially repeated images, instead of recomputing them for every request.
**Speculative decoding**
Drafts candidate tokens and verifies them with the target model; Eagle3 is the speculative path used here.
**KV offload**
Moves colder KV blocks to a host-memory tier so longer context can fit without keeping all KV on GPU.
**Frontend decoding**
Moves decode coordination into the Dynamo frontend so routing and cache policy can act before backend execution.
**Multi-node topology**
Runs serving workers across node boundaries to compare aggregate, single-node P/D, and multi-node P/D shapes.
Benchmark Features Model Target
Feature composition
## Agentic coding throughput stack
How much do KV routing, speculative decoding, P/D split, and KV offload gain when composed?
KV routing Spec decoding P/D split KV offload 
**Kimi-K2.5 NVFP4**
**Traffic****Agentic coding trace** **Hardware****24x GB200** Open [Open](/dynamo/v1.4.0/recipes/benchmarks/kimi-k2-5-feature-stack)
Feature composition
## Frontend decoding plus embedding cache
How do Dynamo frontend decoding and embedding cache change a single-GPU multimodal benchmark versus vanilla vLLM serve?
Frontend decoding Embedding cache 
**Qwen3.6-35B-A3B FP8**
**Traffic****Multimodal sliding window** **Hardware****1x H100 or GB200** Open [Open](/dynamo/v1.4.0/recipes/benchmarks/qwen3-6-35b-feature-stack)
A/B test
## Multimodal embedding cache
How much does enabling the vLLM multimodal embedding cache improve repeated-image traffic on one GB200 worker?
Embedding cache 
**Qwen3-VL-30B**
**Traffic****80% image reuse** **Hardware****1x GB200** Open [Open](/dynamo/v1.4.0/recipes/benchmarks/qwen3-vl-embedding-cache)
A/B test
## KV-aware routing + WideEP + P/D split
Does disaggregated KV-aware routing with WideEP improve latency and goodput against a GB200 control?
KV routing WideEP P/D split 
**DeepSeek V3.2 NVFP4**
**Traffic****Long-context coding trace** **Hardware****32x GB200** Open [Open](/dynamo/v1.4.0/recipes/benchmarks/deepseek-v3-2-wideep-routing)
A/B test
## KV-aware routing + prefill/decode split
Does disaggregated KV-aware routing reduce TTFT and ITL compared with aggregated round-robin routing?
KV routing P/D split 
**Qwen3-32B**
**Traffic****Mooncake prefix reuse** **Hardware****16x H200** Open [Open](/dynamo/v1.4.0/recipes/benchmarks/qwen3-32b-kv-routing)
Topology benchmark
## Aggregate vs single-node P/D vs multi-node P/D
How do vLLM topologies compare when normalized by GPU?
P/D split Multi-node 
**Llama-3.3-70B FP8**
**Traffic****8K ISL / 1K OSL** **Hardware****4-16 GPUs** Open [Open](/dynamo/v1.4.0/recipes/benchmarks/llama-3-70b-topology)
> Start with the feature or performance claim you want to inspect.