feat(blog): add MI355X vs B200 GLM-5 FP8 SGLang post

[functionstackx]

Summary

• New blog post: AMD MI355X SGLang FP8 undercuts NVIDIA B200 SGLang FP8 per million tokens on GLM-5, 14 weeks after the model's 2026-02-11 release
• Peak gap: 1.41x at 18 tok/s/user with MTP (40% cheaper) and 1.36x at 10 tok/s/user without MTP on the 8k/1k workload, single-node
• Walks through sgl-project/sglang#21511 (HaiShaw): FP8 KV cache + FP8 attention via TileLang, reusing fused_qk_rope_cat_and_cache_mla for both Q and KV quant on MI355
• Covers GLM-5 architecture (744B/40B active, 256 experts top-8, glm_moe_dsa, DSA + MLA, 200K ctx)
• All tables sourced from InferenceX 2026-05-20 run (g_runid=26187777287); chart preset linked from both DashboardCTA blocks

Test plan

• Visual check on local dev server (pnpm dev → /blog/mi355x-glm5-fp8-sglang-40-cheaper-than-b200)
• Verify chart preset link resolves to the correct GLM-5 FP8 view with i_metric=y_costh and the four series active
• Confirm OG image renders and RSS feed picks up the post
• Sanity-check the GLM-5 parameter counts (744B/40B, 256 experts) against the official ZAI announcement
• Confirm the 2026-05-20 chart numbers in the iso-interactivity table still match once a newer dump publishes

🤖 Generated with Claude Code

---

Note

Low Risk
Content-only additions (documentation skill and static MDX); no application logic, auth, or data pipeline changes.

Overview
Adds a Claude skill (.claude/skills/write-inferencex-blog/SKILL.md) that documents how to draft InferenceX benchmark posts—source-of-truth priority (CSV vs chart), TCO/cost formulas, slug/frontmatter, MDX sections (DashboardCTA, Figure, FAQ JsonLd), and commit/PR workflow—and points at this post as the AMD-vs-NVIDIA single-node cost template.

Publishes a new MDX article at packages/app/content/blog/mi355x-glm5-fp8-sglang-40-cheaper-than-b200.mdx claiming MI355X SGLang FP8 on GLM-5 8k/1k is up to 40% cheaper per million tokens than B200 (peak 1.41x with MTP at 18 tok/s/user), tied to SGLang PR #21511 and InferenceX PR #1440, with per-concurrency tables, iso-interactivity comparisons (including where B200 wins above ~90 tok/s/user), preset dashboard links, and five FAQ JSON-LD entries.

^{Reviewed by Cursor Bugbot for commit c2f98a5. 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 25, 2026 11:22pm

[cursor]

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

There are 3 total unresolved issues (including 2 from previous reviews).

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

^{Reviewed by Cursor Bugbot for commit e8a9524. Configure here.}