VideoGenX

20-clip C1 cohort · ≥10 valid per hardware

Serving latency, cost and energy for one matched 20-clip video workload across hardware. Click a point to pin its measurements and play its clips below; hollow points have no clips. View telemetry opens that deployment.

Benchmark config

Loading video results…

Unofficial run #44198185716Loading video results…

VideoGenX · Video-seconds per $1 TCO vs. Median latency per clip

Shift+Scroll to zoom · Drag to pan · Double-click to reset

Cost tier
GPU basis
Participating GPUs
Source
SemiAnalysis InferenceX™ + unofficial overlays

VideoGenX

No retained clips share a verified prompt, seed, take and generation configuration.

How to read these results

  • One cohort is 20 scheduled clips at concurrency 1 (C1) per hardware. A hardware appears only with at least 10 valid clips of the same workload.
  • Latency runs from request submit to downloaded media.
  • Power is GPU-board power of the participating GPUs only. Host, chassis and facility power are excluded.
  • TCO is modelled from registry $/GPU-hr and measured output, for both participating and allocated GPUs. It is not a measurement or an API bill.
  • Not measured yet: server-side stage timing, host memory and GPU memory where not recorded, SSIM, VBench, CLIP and VLM quality, sustained load, model load, first-request latency, LoRA switch time and best topology per GPU count.