Skip to main content

Research & Benchmarks

Research, DemosWhitepapers

Technical deep-dives, product demonstrations, and research on enterprise AI infrastructure.

How we benchmark →

Product Demos


Video

Streamlined Agentic Inference with KV Cache Offload | Intel × HPE × Metrum AI Solution Demo

Live Demos


Screenshot of the Quantum-Safe Payments on Dell PowerEdge R770 demo
Livepqc-demo.metrum.ai

Quantum-Safe Payments on Dell PowerEdge R770

Quantum-safe payment signing measured on a Dell PowerEdge R770 with Intel Xeon 6 and Intel QuickAssist Technology: the story, how it works, and every result.

Screenshot of the KV Cache over RoCE demo
Livekvcache.metrum.ai

KV Cache over RoCE

Prefill and decode on two Dell PowerEdge XE9680 servers, with the KV cache sent over Broadcom 400G RoCE or 400G TCP: time to first token, throughput and network load per scenario.

Screenshot of the Institutional Portfolio Risk Agent: ROCm Gen-on-Gen demo
Livefsi-demo.metrum.ai

Institutional Portfolio Risk Agent: ROCm Gen-on-Gen

An institutional portfolio risk agent compared rung against rung on ROCm 10.1.0 and ROCm 7.14.1, on the same Dell PowerEdge R7725 with AMD Instinct MI350P GPUs.

Screenshot of the Heterogeneous Inference: llm-d and round robin on Dell PowerEdge R7725 demo
Liveht-infer-demo.metrum.ai

Heterogeneous Inference: llm-d and round robin on Dell PowerEdge R7725

Measured results for llm-d and standard round robin routing on one pool of NVIDIA RTX PRO 6000 and AMD Instinct MI350P GPUs.

Screenshot of the AgenticBench: Coding Agent Capacity demo
Livecoding-agents-demo.metrum.ai

AgenticBench: Coding Agent Capacity

Capacity and throughput for fleets of autonomous coding agents running sandboxed on a single AMD-powered Dell PowerEdge R7725, on a Terminal-Bench 2.1 coding loop.

Screenshot of the Enterprise Agentic Bench demo
Liveenterprise-bench-demo.metrum.ai

Enterprise Agentic Bench

Contract review, document redaction and email threat analysis run as enterprise agent workloads, with concurrent agents per node and the workload profile behind each number.