SemiAnalysisAI · Oseltamivir · Jul 20, 2026 · Jul 20, 2026 · Jul 20, 2026
@@ -1,6 +1,7 @@
 ---
 title: 'B200 NVFP4 vs H200 FP8 on GLM-5: Up to 3.65x Better Performance per Dollar with SGLang MTP'
 subtitle: 'Both SKUs run SGLang EAGLE MTP; the Blackwell generation lifts perf/$ by ~1.2x at the peak and the NVIDIA GLM-5-NVFP4 checkpoint on FlashInfer TRT-LLM sparse MLA stacks another ~2.4–3.0x on 8K/1K'
+seoDescription: 'B200 NVFP4 delivers up to 3.65x better perf/$ than H200 FP8 on GLM-5 with SGLang MTP, stacking Blackwell gains and a sparse-MLA checkpoint.'
 date: '2026-05-26'
 publishDate: '2026-05-26'
 tags:

@@ -1,6 +1,7 @@
 ---
 title: 'B200 NVFP4 vs H100 FP8 on MiniMax-M2.5: Up to 8.2x Better Performance per Dollar with vLLM'
 subtitle: 'vLLM PR #36307 unlocks the trtllm-gen FP8 MoE kernel for MiniMax on B200; combined with NVFP4, perf/$ scales from 4.0x at 22 tok/s/user to 8.2x at 110 on 8K/1K'
+seoDescription: 'B200 NVFP4 delivers up to 8.2x better perf/$ than H100 FP8 on MiniMax-M2.5 with vLLM — trtllm-gen FP8 MoE plus NVFP4 on 8K/1K.'
 date: '2026-05-26'
 publishDate: '2026-05-26'
 tags:

@@ -1,6 +1,7 @@
 ---
 title: 'B200 NVFP4 vs H200 INT4 on Kimi K2.5/K2.6: Up to 2.95x Better Performance per Dollar'
 subtitle: "On vLLM 8K/1K the NVFP4 path on B200 is 2.71x–2.95x cheaper per million tokens than H200 INT4 across the entire 30–90 tok/s/user serving band, and 2.45x–2.74x cheaper than B200 INT4 on the same silicon. Both factors decompose cleanly into B200's HBM bandwidth, HBM capacity, and NVFP4 tensor cores"
+seoDescription: 'B200 NVFP4 runs Kimi K2.5/K2.6 up to 2.95x cheaper per token than H200 INT4 on vLLM 8K/1K, across the full 30–90 tok/s/user serving band.'
 date: '2026-05-26'
 publishDate: '2026-05-26'
 tags:

@@ -1,6 +1,7 @@
 ---
 title: 'GB200 NVL72 vs B200 on DeepSeek R1 670B: Up to 4.4x Throughput per GPU at 125 tok/s/user'
 subtitle: "DeepSeek R1 FP4 1k/1k. NVL72's 72-GPU NVLink scale-up fabric lets decode run wide EP up to EP=32, where B200's 8-GPU NVLink island caps out at EP=8 over RoCEv2"
+seoDescription: 'GB200 NVL72 hits up to 4.4x more per-GPU throughput than B200 on DeepSeek R1 670B at 125 tok/s/user — 72-GPU NVLink enables wide EP=32.'
 date: '2026-05-23'
 publishDate: '2026-05-23'
 tags:

@@ -1,6 +1,7 @@
 ---
 title: 'GB300 NVL72 vs GB200 NVL72 Inference Performance & Perf per Dollar - on DeepSeek-V4-Pro 1.6T: Up to 2.83x Throughput'
 subtitle: "DSv4-Pro FP4 8K/1K, Dynamo+vLLM, disaggregated on both racks. GB300's 50% extra HBM (288 vs 192 GB/GPU) unlocks a wider prefill+decode recipe GB200 can't fit — lifting middle-of-curve perf/$ by 2.31x despite a 20% per-GPU TCO premium."
+seoDescription: 'GB300 NVL72 adds 50% more HBM, lifting DeepSeek-V4-Pro perf/$ up to 2.31x over GB200 NVL72 on vLLM FP4 8K/1K — up to 2.83x throughput per GPU.'
 date: '2026-05-27'
 publishDate: '2026-05-27'
 tags:

@@ -1,6 +1,7 @@
 ---
 title: 'InferenceMAX: Open Source Inference Benchmarking'
 subtitle: 'NVIDIA GB200 NVL72, AMD MI355X, Throughput Token per GPU, Latency Tok/s/user, Perf per Dollar, Cost per Million Tokens, Tokens per Provisioned Megawatt, DeepSeek R1 670B, GPTOSS 120B, Llama3 70B'
+seoDescription: 'InferenceMAX: open-source inference benchmarks across NVIDIA GB200 NVL72 and AMD MI355X — throughput, latency, cost per token and perf per watt.'
 date: '2025-10-09'
 publishDate: '2025-10-09'
 tags:

@@ -1,6 +1,7 @@
 ---
 title: 'MI355X DeepSeek-V4-Pro on SGLang: 110.5x Throughput per GPU in 26 Days'
 subtitle: 'The amd/deepseek_v4 side branch shipped TileLang attention indexer, Triton sparse MLA, fused RoPE/Hadamard, FlyDSL MoE, and FP4 weights across 31 performance optimizations PRs — lifting first-light 20 tok/s/GPU at 2.4 tok/s/user into 2,256 tok/s/GPU at 9.4 tok/s/user on 8K/1K, with both throughput and interactivity climbing together'
+seoDescription: 'MI355X ran DeepSeek-V4-Pro 110.5x faster per GPU in 26 days on SGLang — from 20 to 2,256 tok/s/GPU via TileLang, sparse MLA and FP4 on 8K/1K.'
 date: '2026-05-26'
 publishDate: '2026-05-26'
 tags:

@@ -1,6 +1,7 @@
 ---
 title: 'AMD MI355X GLM-5 Inference: Up to 40% Cheaper per Million Tokens than B200 on SGLang FP8'
 subtitle: '14 weeks after GLM-5 launched, AMD landed both MTP and non-MTP SGLang FP8 recipes on MI355X — fused MLA + FP8 KV cache via TileLang flips the single-node FP8 cost curve in AMD favor across most of the performance Pareto'
+seoDescription: 'AMD MI355X runs GLM-5 up to 40% cheaper per million tokens than B200 on SGLang FP8, with fused MLA and FP8 KV cache via TileLang.'
 date: '2026-05-25'
 publishDate: '2026-05-25'
 tags:

@@ -1,6 +1,7 @@
 ---
 title: 'AMD MI355X Qwen3.5 397B-A17B Inference: Up to 19x Throughput per GPU in 3 Months on SGLang FP8'
 subtitle: 'From v0.5.8 (Feb) → v0.5.10rc0 (Apr) → v0.5.12 (May), three AITER kernel landings on MI355X plus a TP=8 → TP=2/TP=4 retune push Qwen3.5 8k/1k peak from 1.3k to 6.4k tok/s/GPU and extend the curve out to 75 tok/s/user'
+seoDescription: 'AMD MI355X lifted Qwen3.5 397B-A17B up to 19x per-GPU throughput in 3 months on SGLang FP8 — 1.3k to 6.4k tok/s/GPU via AITER kernels.'
 date: '2026-05-25'
 publishDate: '2026-05-25'
 tags:

@@ -1,6 +1,7 @@
 ---
 title: 'B200 NVFP4 对比 H200 FP8 运行 GLM-5：SGLang MTP 下性价比提升高达 3.65 倍'
 subtitle: '两款 GPU 均运行 SGLang EAGLE MTP；Blackwell 世代在峰值处带来约 1.2 倍的性价比提升，NVIDIA GLM-5-NVFP4 检查点搭配 FlashInfer TRT-LLM 稀疏 MLA 在 8K/1K 场景下再叠加约 2.4–3.0 倍优势'
+seoDescription: 'B200 NVFP4 在 SGLang MTP 下运行 GLM-5，性价比比 H200 FP8 最高提升 3.65 倍——叠加 Blackwell 世代增益与稀疏 MLA 检查点。'
 date: '2026-05-26'
 publishDate: '2026-05-26'
 tags:

@@ -1,6 +1,7 @@
 ---
 title: 'B200 NVFP4 vs H100 FP8 运行 MiniMax-M2.5：vLLM 下每美元性能最高提升 8.2 倍'
 subtitle: 'vLLM PR #36307 为 MiniMax 在 B200 上解锁了 trtllm-gen FP8 MoE 模块化内核；结合 NVFP4，在 8K/1K 负载下性能/成本从 22 tok/s/user 时的 4.0 倍扩大到 110 tok/s/user 时的 8.2 倍'
+seoDescription: 'B200 NVFP4 在 vLLM 下运行 MiniMax-M2.5，每美元性能比 H100 FP8 最高提升 8.2 倍——trtllm-gen FP8 MoE 结合 NVFP4（8K/1K）。'
 date: '2026-05-26'
 publishDate: '2026-05-26'
 tags:

@@ -1,6 +1,7 @@
 ---
 title: 'B200 NVFP4 对比 H200 INT4 运行 Kimi K2.5/K2.6：性价比提升高达 2.95 倍'
 subtitle: '在 vLLM 8K/1K 工作负载下，B200 NVFP4 路径在 30–90 tok/s/user 推理区间内每百万 tokens 成本比 H200 INT4 低 2.71x–2.95x，比同一 B200 硬件上的 INT4 低 2.45x–2.74x。三个因素——B200 的 HBM 带宽、HBM 容量和 NVFP4 张量核心——可清晰分解该优势'
+seoDescription: '在 vLLM 8K/1K 下，B200 NVFP4 运行 Kimi K2.5/K2.6 每 token 成本比 H200 INT4 最多低 2.95 倍，覆盖 30–90 tok/s/user 全区间。'
 date: '2026-05-26'
 publishDate: '2026-05-26'
 tags:

@@ -1,6 +1,7 @@
 ---
 title: 'GB200 NVL72 对比 B200 运行 DeepSeek R1 670B：在 125 tok/s/user 下每 GPU 吞吐量最高达 4.4 倍'
 subtitle: 'DeepSeek R1 FP4 1k/1k。NVL72 的 72-GPU NVLink 扩展域允许解码使用最高 EP=32 的宽专家并行，而 B200 的 8-GPU NVLink 岛通过 RoCEv2 上限为 EP=8'
+seoDescription: 'GB200 NVL72 在 125 tok/s/user 下运行 DeepSeek R1 670B，每 GPU 吞吐量比 B200 最高提升 4.4 倍——72-GPU NVLink 支持 EP=32 宽专家并行。'
 date: '2026-05-23'
 publishDate: '2026-05-23'
 tags:

@@ -1,6 +1,7 @@
 ---
 title: 'GB300 NVL72 vs GB200 NVL72 推理性能与性价比对比 — DeepSeek-V4-Pro 1.6T：吞吐量最高提升 2.83 倍'
 subtitle: 'DSv4-Pro FP4 8K/1K，Dynamo+vLLM，两套机架均采用分离式部署。GB300 多出 50% 的 HBM（每 GPU 288 GB vs 192 GB）解锁了 GB200 无法容纳的更宽预填充+解码配方——尽管单 GPU TCO 溢价 20%，曲线中段性价比仍提升 2.31 倍。'
+seoDescription: 'GB300 NVL72 多出的 50% HBM 使 DeepSeek-V4-Pro 性价比比 GB200 NVL72 最高提升 2.31 倍（vLLM FP4 8K/1K），每 GPU 吞吐量最高提升 2.83 倍。'
 date: '2026-05-27'
 publishDate: '2026-05-27'
 tags:

@@ -1,6 +1,7 @@
 ---
 title: 'InferenceMAX：开源推理基准测试'
 subtitle: 'NVIDIA GB200 NVL72、AMD MI355X、每 GPU 吞吐量 Token、延迟 Tok/s/user、性价比、每百万 Token 成本、每配置兆瓦 Token 数、DeepSeek R1 670B、GPTOSS 120B、Llama3 70B'
+seoDescription: 'InferenceMAX：面向 NVIDIA GB200 NVL72 与 AMD MI355X 的开源推理基准测试——吞吐量、延迟、每 token 成本与每瓦性能。'
 date: '2025-10-09'
 publishDate: '2025-10-09'
 tags:

@@ -1,6 +1,7 @@
 ---
 title: 'MI355X 上 DeepSeek-V4-Pro 搭配 SGLang：26 天内每 GPU 吞吐量提升 110.5 倍'
 subtitle: 'amd/deepseek_v4 分支合入了 TileLang 注意力索引器、Triton 稀疏 MLA、融合 RoPE/Hadamard、FlyDSL MoE 以及 FP4 权重，历经 31 个性能优化 PR——将首次点亮时 20 tok/s/GPU、2.4 tok/s/user 的水平提升至 8K/1K 负载下 2,256 tok/s/GPU、9.4 tok/s/user，吞吐量与交互性同步攀升'
+seoDescription: 'MI355X 在 SGLang 上 26 天内将 DeepSeek-V4-Pro 的每 GPU 吞吐量提升 110.5 倍——8K/1K 下经 TileLang、稀疏 MLA 与 FP4 从 20 提升至 2,256 tok/s/GPU。'
 date: '2026-05-26'
 publishDate: '2026-05-26'
 tags:

@@ -1,6 +1,7 @@
 ---
 title: 'AMD MI355X GLM-5 推理：SGLang FP8 单节点每百万 token 成本比 B200 最高低 40%'
 subtitle: 'GLM-5 发布 14 周后，AMD 在 MI355X 上同时实现了 SGLang FP8 的 MTP 和非 MTP 方案 — 通过 TileLang 实现的融合 MLA + FP8 KV 缓存在大部分性能 Pareto 前沿上将单节点 FP8 成本曲线翻转为 AMD 占优'
+seoDescription: 'AMD MI355X 在 SGLang FP8 上运行 GLM-5，每百万 token 成本比 B200 最高低 40%——TileLang 融合 MLA 与 FP8 KV 缓存。'
 date: '2026-05-25'
 publishDate: '2026-05-25'
 tags:

@@ -1,6 +1,7 @@
 ---
 title: 'AMD MI355X Qwen3.5 397B-A17B 推理：SGLang FP8 三个月内每 GPU 吞吐量提升最高 19 倍'
 subtitle: '从 v0.5.8（2 月）→ v0.5.10rc0（4 月）→ v0.5.12（5 月），三次 AITER 内核合入 MI355X 加上从 TP=8 到 TP=2/TP=4 的重新调优，将 Qwen3.5 8k/1k 峰值从 1.3k 推高至 6.4k tok/s/GPU，并将曲线延伸至 75 tok/s/user'
+seoDescription: 'AMD MI355X 在 SGLang FP8 上 3 个月内将 Qwen3.5 397B-A17B 每 GPU 吞吐量最高提升 19 倍——经 AITER 内核从 1.3k 提升至 6.4k tok/s/GPU。'
 date: '2026-05-25'
 publishDate: '2026-05-25'
 tags:

@@ -16,6 +16,7 @@ import { ShareTwitterButton, ShareLinkedInButton } from '@/components/share-butt
 import { Card } from '@/components/ui/card';
 import { JsonLd } from '@/components/json-ld';
 import {
+  blogDescription,
   getAllPosts,
   getAdjacentPosts,
   extractHeadings,
@@ -43,10 +44,14 @@ export async function generateMetadata({ params }: Props): Promise<Metadata> {
   const result = getPostBySlug(slug);
   if (!result) return {};
   const { meta } = result;
+  const description = blogDescription(meta);
 
   return {
-    title: meta.title,
-    description: meta.subtitle,
+    // `absolute` keeps a short " | InferenceX" suffix instead of the long
+    // "%s | InferenceX by SemiAnalysis" root template, leaving more room for
+    // the headline before Google truncates the SERP title.
+    title: { absolute: `${meta.title} | ${SITE_NAME}` },
+    description,
     keywords: meta.tags,
     authors: [{ name: AUTHOR_NAME }],
     alternates: {
@@ -56,7 +61,7 @@ export async function generateMetadata({ params }: Props): Promise<Metadata> {
     },
     openGraph: {
       title: `${meta.title} | ${SITE_NAME}`,
-      description: meta.subtitle,
+      description,
       url: `${SITE_URL}/blog/${slug}`,
       type: 'article',
       publishedTime: `${meta.date}T00:00:00Z`,
@@ -67,7 +72,7 @@ export async function generateMetadata({ params }: Props): Promise<Metadata> {
     twitter: {
       card: 'summary_large_image',
       title: meta.title,
-      description: meta.subtitle,
+      description,
       site: AUTHOR_HANDLE,
       creator: AUTHOR_HANDLE,
     },
@@ -139,7 +144,10 @@ export default async function BlogPostPage({ params }: Props) {
     publisher: { '@type': 'Organization', name: AUTHOR_NAME },
     datePublished: `${meta.date}T00:00:00Z`,
     ...(meta.modifiedDate && { dateModified: `${meta.modifiedDate}T00:00:00Z` }),
-    description: meta.subtitle,
+    // Keep structured-data description in sync with the SERP/OG/Twitter meta
+    // (both go through blogDescription) so they never diverge for posts with a
+    // seoDescription or a long subtitle.
+    description: blogDescription(meta),
     url: `${SITE_URL}/blog/${slug}`,
     wordCount: raw.trim().split(/\s+/u).length,
     timeRequired: `PT${meta.readingTime}M`,

@@ -1,12 +1,7 @@
 import type { Metadata } from 'next';
 import { notFound, permanentRedirect } from 'next/navigation';
 
-import {
-  HW_REGISTRY,
-  SITE_NAME,
-  SITE_URL,
-  SUPPORTERS_LINE,
-} from '@semianalysisai/inferencex-constants';
+import { HW_REGISTRY, SITE_NAME, SITE_URL } from '@semianalysisai/inferencex-constants';
 
 import { JsonLd } from '@/components/json-ld';
 import { languageAlternates } from '@/lib/i18n';
@@ -15,11 +10,14 @@ import {
   canonicalCompareSlug,
   compareDisplayLabel,
   compareModelDisplayLabel,
+  compareModelSeoName,
+  compareSeoTitle,
   parseCompareSlug,
 } from '@/lib/compare-slug';
 import {
   buildBreadcrumbJsonLd,
   buildJsonLd,
+  compareMetaDescription,
   compareTableNarrative,
   computeCompareTableData,
   dateRangeForPair,
@@ -46,16 +44,30 @@ export async function generateMetadata({ params }: Props): Promise<Metadata> {
   if (!parsed) return {};
   const fullLabel = compareModelDisplayLabel(parsed.model, parsed.a, parsed.b);
   const gpuLabel = compareDisplayLabel(parsed.a, parsed.b);
-  const url = `${SITE_URL}/compare/${canonicalCompareSlug(parsed.model.slug, parsed.a, parsed.b)}`;
-  const description = `${gpuLabel} inference benchmark on ${parsed.model.label}: verified, reproducible head-to-head results from InferenceX, the independent open-source GPU benchmark by SemiAnalysis. ${SUPPORTERS_LINE} Compare latency, throughput & cost.`;
+  const modelSeoName = compareModelSeoName(parsed.model);
+  const canonical = canonicalCompareSlug(parsed.model.slug, parsed.a, parsed.b);
+  const url = `${SITE_URL}/compare/${canonical}`;
+
+  // Lead the SEO title with the GPU pair — that's the phrase people search
+  // ("b200 vs b300") and must survive Google's ~60-char SERP truncation. The
+  // `absolute` form bypasses the long "%s | InferenceX by SemiAnalysis" root
+  // template so the query isn't pushed off the end.
+  const title = `${compareSeoTitle(gpuLabel, modelSeoName)} | ${SITE_NAME}`;
+
+  // Stat-led meta description built from the interpolated head-to-head numbers
+  // at the slug's default operating point (falls back to boilerplate for
+  // sparse-data pairs). Fetch is blob-cached and shared with the page render.
+  const rows = await getCachedBenchmarks(parsed.model.dbKeys);
+  const { sequence, precision } = pickPairDefaults(rows, parsed.a, parsed.b);
+  const { ssrRows } = computeCompareTableData(rows, parsed.a, parsed.b, sequence, precision);
+  const description = compareMetaDescription(parsed.model, parsed.a, parsed.b, ssrRows);
+
   return {
-    title: `${fullLabel} Inference Benchmark`,
+    title: { absolute: title },
     description,
     alternates: {
       canonical: url,
-      languages: languageAlternates(
-        `/compare/${canonicalCompareSlug(parsed.model.slug, parsed.a, parsed.b)}`,
-      ),
+      languages: languageAlternates(`/compare/${canonical}`),
     },
     openGraph: {
       title: `${fullLabel} | ${SITE_NAME}`,

@@ -16,7 +16,13 @@ import { ReadingProgressBar } from '@/components/blog/reading-progress-bar';
 import { ShareTwitterButton, ShareLinkedInButton } from '@/components/share-buttons';
 import { Card } from '@/components/ui/card';
 import { JsonLd } from '@/components/json-ld';
-import { getAllPosts, getAdjacentPosts, extractHeadings, getPostBySlug } from '@/lib/blog';
+import {
+  blogDescription,
+  getAllPosts,
+  getAdjacentPosts,
+  extractHeadings,
+  getPostBySlug,
+} from '@/lib/blog';
 import { ZH_LANG_TAG, ZH_OG_LOCALE, zhAlternates } from '@/lib/i18n';
 import {
   AUTHOR_HANDLE,
@@ -38,16 +44,19 @@ export async function generateMetadata({ params }: Props): Promise<Metadata> {
   const result = getPostBySlug(slug, 'zh');
   if (!result) return {};
   const { meta } = result;
+  const description = blogDescription(meta);
 
   return {
-    title: meta.title,
-    description: meta.subtitle,
+    // Short " | InferenceX" suffix via `absolute` (mirrors the English page)
+    // so the headline keeps more of the SERP title before truncation.
+    title: { absolute: `${meta.title} | ${SITE_NAME}` },
+    description,
     keywords: meta.tags,
     authors: [{ name: AUTHOR_NAME }],
     alternates: zhAlternates(`/blog/${slug}`),
     openGraph: {
       title: `${meta.title} | ${SITE_NAME}`,
-      description: meta.subtitle,
+      description,
       url: `${SITE_URL}/zh/blog/${slug}`,
       type: 'article',
       locale: ZH_OG_LOCALE,
@@ -59,7 +68,7 @@ export async function generateMetadata({ params }: Props): Promise<Metadata> {
     twitter: {
       card: 'summary_large_image',
       title: meta.title,
-      description: meta.subtitle,
+      description,
       site: AUTHOR_HANDLE,
       creator: AUTHOR_HANDLE,
     },
@@ -131,7 +140,10 @@ export default async function ZhBlogPostPage({ params }: Props) {
     publisher: { '@type': 'Organization', name: AUTHOR_NAME },
     datePublished: `${meta.date}T00:00:00Z`,
     ...(meta.modifiedDate && { dateModified: `${meta.modifiedDate}T00:00:00Z` }),
-    description: meta.subtitle,
+    // Keep structured-data description in sync with the SERP/OG/Twitter meta
+    // (both go through blogDescription) so they never diverge for posts with a
+    // seoDescription or a long subtitle.
+    description: blogDescription(meta),
     url: `${SITE_URL}/zh/blog/${slug}`,
     inLanguage: ZH_LANG_TAG,
     wordCount: raw.trim().split(/\s+/u).length,