feat(blog): B200 NVFP4 vs H200 FP8 on GLM-5 — up to 3.65x better perf/$

[functionstackx]

Summary

• New blog post comparing B200 NVFP4 + MTP vs H200 FP8 + MTP on GLM-5 8K/1K with SGLang v0.5.12 (GHA run 26381101926). Headline: up to 3.65x better performance per dollar at 80 tok/s/user, stable in a 3.24x–3.65x band across H200's 25–84 tok/s/user range.
• Decomposes the peak 3.65x into a 1.22x generation step (Blackwell vs Hopper at FP8 + MTP) and a 2.98x precision step (B200 FP8 → B200 NVFP4).
• New ## On-Paper Specs section with the /gpu-specs radar + an absolute-values table sourced from packages/app/src/lib/gpu-specs.ts. Bridges from raw silicon ratios to the measured perf/$ lift via a "compute-bound ceiling / HBM-bound floor" bracket.
• Iso-interactivity table uses the bundled Pareto helper at .claude/skills/write-inferencex-blog/iso_interactivity.py; cells outside H200's range render as _unreachable_.
• Anchors: SGLang tracking #19380, kernel PRs #21783 / #21405, FlashInfer #2726 / #2836, InferenceX recipes #1087 (B200 NVFP4+MTP) and #1480 (H200 FP8+MTP).
• SKILL.md updates: codifies the new On-Paper Specs section pattern, adds the /gpu-specs radar image to the Step 0 ask list, and adds a banned-phrases banner so future iso-interactivity intros stop leaking algorithm names into reader-facing prose.

Test plan

• Vercel preview renders at /blog/b200-glm5-nvfp4-vs-h200-fp8-3-6x-perf-per-dollar
• Hero benchmark <Figure> and radar <Figure> render in both themes (currently the same image is in both light + dark slots — drop a real dark radar export later if desired)
• Iso-interactivity table values match the live cost view for spot-checked iv rows (25 / 50 / 80 / 84)
• OG image generates
• Sitemap + RSS pick up the new slug
• FAQ JSON-LD passes a schema validator

🤖 Generated with Claude Code

---

Note

Low Risk
Content-only change (MDX blog post and authoring skill docs); no application runtime, auth, or data-path code.

Overview
Adds a new InferenceX benchmark post comparing B200 NVFP4 + MTP vs H200 FP8 + MTP on GLM-5 8K/1K (SGLang v0.5.12): headline up to 3.65× perf/$ at 80 tok/s/user, decomposed into ~1.22× generation (FP8+MTP) and ~2.98× precision (NVFP4 on B200). The post follows the standard blog shape—dashboard CTAs, hero/repeat figures, per-concurrency tables, iso-interactivity cost table with _unreachable_ cells, upstream PR narrative, FAQ JSON-LD—and introduces an ## On-Paper Specs block (radar figure + table from gpu-specs.ts + perf/$ bracket vs measured lift).

Updates write-inferencex-blog SKILL.md: Step 0 now asks for a /gpu-specs radar screenshot; documents the full On-Paper Specs section template; renumbers upstream/recipe inputs; and adds an explicit banned-phrases list so published posts must not name the dashboard’s iso-interactivity spline/Hermite implementation (approved copy: “interpolated along each SKU’s Pareto frontier”).

^{Reviewed by Cursor Bugbot for commit f479a2d. Bugbot is set up for automated code reviews on this repo. Configure here.}

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
inferencemax-app	Ready	Preview, Comment	May 26, 2026 3:14am

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit f479a2d. Configure here.}

[cursor]

"Three gaps" listed but only one bullet present

High Severity

The "What's Next" section promises "Three gaps still narrow the headline number" but only one bullet point (Disaggregated serving on NVL72) follows. The FAQ JsonLd at the bottom lists all three: (1) disaggregated serving, (2) piecewise CUDA graph for B200 FP8 Agg prefill (#23351/#24276), and (3) H200 recipe improvements (disagg, trtllm-mha, KV FP8). Two bullet points were dropped from the published prose. The sibling blog post (mi355x-glm5-fp8-sglang-40-cheaper-than-b200.mdx) correctly matches its "Two gaps" count with two bullets.

Additional Locations (1)

• packages/app/content/blog/b200-glm5-nvfp4-vs-h200-fp8-3-6x-perf-per-dollar.mdx#L210-L211

 

^{Reviewed by Cursor Bugbot for commit f479a2d. Configure here.}