SemiAnalysis · Launch Video Breakdown: Hook, Pacing & Motion Design

Scene-by-scene timeline & spoken transcript

1. 0:00 – 0:08

   The Hook

   Evolution of Silicon

   “(No spoken dialogue — ambient electronic synth music)”

   On screen

   4004 intel 71 GPC1 GPC2 GPC3 GPC4 GPC16

   Camera

   Rapid sequence of macro silicon die shots with slight zoom-ins

   Motion

   Hard jump-cuts synchronized to the synthesizer beat

2. 0:08 – 0:15

   Problem Agitation

   Measuring Thinking Machines

   “We now have machines that can think, but how do we measure their performance?”

   On screen

   We now have machines that can think... but how do we measure their performance?

   Camera

   Center lock framing on processor package transitioning to dual-pane terminal

   Motion

   Typewriter kinetic typography overlaid on dark UI recordings

3. 0:15 – 0:25

   Product Reveal

   Introducing InferenceX

   “InferenceX is the definitive benchmark for AI inference performance. The new agentic inference benchmark showcases real-world chip performance across many of the most popular frontier open source models.”

   On screen

   InferenceX by SemiAnalysis Compare Realistic Agentic Inference Perf

   Camera

   Smooth browser scroll and cursor navigation into model select

   Motion

   Cursor hover triggering active UI selection states

4. 0:25 – 0:42

   Feature Teaser

   Pareto Frontiers & Latency Metrics

   “Select from a wide range of hardware and inference engine configurations to view and compare their performance frontiers. Choose between common latency metrics such as interactivity, time to first token, and end-to-end latency.”

   On screen

   Total Tokens per $1 USD (Owning - Hyperscaler) vs. P90 Interactivity GB300 NVL72 MI355X H200 B200

   Camera

   Focused screen pan and zoom across scatter plot Pareto curves

   Motion

   Animated curve rendering and instant SVG data point toggling

5. 0:42 – 1:01

   Feature Teaser

   Custom Cost Metrics & Optimization Changelogs

   “Select a y-axis metric to compare what matters most to you. Track a chip configuration's performance over time to see how inference optimizations drive measurable performance gains.”

   On screen

   Token Throughput per All in Utility MW Config Changelog (3 dates with changes)

   Camera

   Dropdown focus zooming into selection menus and historical diffs

   Motion

   Fast UI modal pop-up and table expansion

6. 1:01 – 1:28

   Feature Teaser

   Granular Telemetry & Profiling

   “Each point on the Pareto frontier represents a single benchmark run. Hover over any point to explore detailed performance data. Click View Charts to view server metrics: throughput, KV cache metrics such as token source, KV working set size, prefix cache hit rate, and utilization over time.”

   On screen

   B300 (vLLM) Telemetry Throughput (input & decode) KV cache utilization over time

   Camera

   Hover tooltip zoom followed by smooth page transition into dashboard

   Motion

   Interactive tooltips, multi-series graph rendering, full-screen chart focus

7. 1:28 – 1:50

   Feature Teaser

   Agent Waterfall Timeline & Logs

   “View the actual timeline of requests received by the server, including realistic sub-agent rollouts. Click any request to inspect its source and token breakdown from the underlying dataset. Inspect logs from every part of the benchmark, including servers, routers, KV transfer engines, and more.”

   On screen

   Request timeline Subagent 1 Turn 2 Log Files vLLM router.log

   Camera

   Horizontal pan across Gantt chart timeline to code log viewer

   Motion

   Dynamic bar chart expansion and interactive stack-trace expansion

8. 1:50 – 2:03

   Feature Teaser

   Reproducibility on Real Hardware

   “Each benchmark runs on real hardware, managed by SemiAnalysis through self-hosted GitHub Actions, with the full run linked for every data point.”

   On screen

   Run Sweep SemiAnalysis/inferencex GitHub Actions launch-agent.sh

   Camera

   Direct capture zoom on GitHub workflow execution tree and terminal script

   Motion

   Tree-view expand animation and shell script syntax highlight

9. 2:03 – 2:19

   Call to Action

   100% Open Source Call to Action

   “InferenceX is 100% open source. Every result is reproducible, no private hyper-optimized containers. InferenceX is the authoritative index for real-world inference performance. Check it out today for yourself at inferencex.com.”

   On screen

   InferenceX by SemiAnalysis Compare Realistic Agentic Inference Perf inferencex.com

   Camera

   Slow downward scroll through industry endorsements to URL display

   Motion

   Smooth card scroll and subtle fade-to-black