# Analysis of the 100 most popular hardware setups on Hugging Face
**作者**: clem
**日期**: 2026-05-06T13:38:12.000Z
**来源**: [https://x.com/ClementDelangue/status/2052020105328890188](https://x.com/ClementDelangue/status/2052020105328890188)
---
Just dropped a fun new dataset: the 100 most popular hardware setups for AI builders on Hugging Face, based on 297,135 users who voluntarily filled in the local hardware section of their HF profile: https://huggingface.co/datasets/clem/100_most_popular_hardware_setups_on_HF
A note on methodology, h/t @JordanNanos @SemiAnalysis_ who rightly pointed out that the cleanest way to read this is to scope each row into exactly one of four mutually exclusive buckets: Discrete GPU, SoC/APU, CPU-only, or CPU+GPU combo. That avoids comparing Apple SoCs to standalone GPU SKUs, and avoids double-counting vendors across combos. The 140,141 top-100 user-reports break down as 60,120 Discrete GPU users (43%), 50,077 SoC/APU users (36%), 17,841 CPU-only (13%), and 12,103 CPU+GPU combos (8.6%).
The long tail is enormous. The top 100 setups only covers 47% of all 297,135 reporters. More than half of HF builders are running something that didn't even crack the top 100. Hardware fragmentation in the local-AI world is real.
Each vendor owns its own category. NVIDIA dominates discrete GPUs at 97.3%, Apple dominates SoCs/APUs at 95.6%, Intel dominates pure CPU rows at 82.5%. The naive "Apple vs NVIDIA" comparison is misleading because they actually compete in different buckets.
The CPU world has a striking inversion. Among CPU-only users, Intel leads 82.5% to 17.5%. But among users who explicitly build a CPU+GPU rig, AMD Ryzen leads 65% to 35%. The enthusiast / DIY local-AI builder crowd has clearly moved to Ryzen, even though Intel still has the bigger broad installed base.
VRAM beats raw power. The single most popular discrete GPU isn't the 4090 or the 5090, it's the RTX 3060 at 4,737 users. Specifically the 12GB version, which has more VRAM than the 3060 Ti, 3070, 4060, and 4060 Ti, and the same as the 3080. The 12GB RTX 3060 has roughly 4x the users of the 8GB RTX 3060 Ti, even though they share a name. AI builders care about memory size, not benchmark scores.
AI builders skew hard toward Pro and Max chips. For the M3 family, only 22% of users are on the base M3. The rest are on M3 Pro, M3 Max, or M3 Ultra. The M3 Pro alone (3,141 users) is significantly more popular than the M3 base (1,968). Same pattern for M1: more people run the M1 Pro (4,815) than the base M1 (4,499). Higher unified-memory tiers matter way more for local AI than they do for general computing.
The 10-series refuses to die. GTX 1660, 1650 Mobile, 1070 Ti, 1060, 1080 Ti, and 1050 Ti, six to nine-year-old Pascal-era cards, collectively account for roughly 5,000 users in the top 100. Roughly 1 in 13 discrete-GPU users on HF is still running silicon from before the RTX era.
The M4 family is already bigger than the M1 family. M4 totals 16,639 users, 34% more than M1 (12,444). The M5, which just shipped, is already at 2,907. The M2 generation (6,917) sits oddly low, fewer users than the older M3 (8,967), suggesting M2 had a short adoption window before users jumped to M3.
The 50-series is the fastest gen ramp in the data. RTX 50-series already has 13 SKUs in the top 100 totaling 15,341 users. Faster gen-on-gen adoption than the 30 or 40 series saw at the same age.
Datacenter / pro silicon is real but smaller than you'd expect. H100, A100, H200, V100, T4, L4, L40s, RTX 6000 Ada, RTX PRO 6000 WS, A6000, and GB10 together total 10,792 users (~7.7% of top-100 reports). Almost certainly under-counted because researchers don't usually list shared-cluster GPUs on their personal profile.
New AI-specific silicon is already showing up. NVIDIA's GB10 (DGX Spark) sits at #36 with 1,241 users. AMD's Ryzen AI Max+ 395 (Strix Halo) is at #49 with 962. Both are recent, and both are already meaningful in the rankings.
The most popular enthusiast combos: Ryzen 9 Zen 5 + RTX 5090 (1,504), then Intel i9 13th-gen + RTX 4090 (1,424).
Caveats worth keeping in mind: the data is self-reported and opt-in (biased toward HF-engaged local-AI builders); users typically list one machine even if they own several; cloud and shared-cluster work is almost certainly under-represented; and label granularity isn't uniform across vendors.
Dataset (Apache 2.0): https://huggingface.co/datasets/clem/100_most_popular_hardware_setups_on_HF
Haven't added your hardware yet? Takes 30 seconds: https://huggingface.co/settings/local-apps
Let's go local AI! 🤗
## 相关链接
- [Hugging Face reposted](https://x.com/huggingface)
- [clem](https://x.com/ClementDelangue)
- [@ClementDelangue](https://x.com/ClementDelangue)
- [9.7K](https://x.com/ClementDelangue/status/2052020105328890188/analytics)
- [local hardware section of their HF profile](https://huggingface.co/settings/local-apps)
- [https://huggingface.co/datasets/clem/100_most_popular_hardware_setups_on_HF](https://huggingface.co/datasets/clem/100_most_popular_hardware_setups_on_HF)
- [@JordanNanos](https://x.com/@JordanNanos)
- [@SemiAnalysis_](https://x.com/@SemiAnalysis_)
- [https://huggingface.co/datasets/clem/100_most_popular_hardware_setups_on_HF](https://huggingface.co/datasets/clem/100_most_popular_hardware_setups_on_HF)
- [https://huggingface.co/settings/local-apps](https://huggingface.co/settings/local-apps)
- [Upgrade to Premium](https://x.com/i/premium_sign_up)
- [9:38 PM · May 6, 2026](https://x.com/ClementDelangue/status/2052020105328890188)
- [9,751 Views](https://x.com/ClementDelangue/status/2052020105328890188/analytics)
---
*导出时间: 2026/5/6 23:00:04*
---
## 中文翻译
# Hugging Face 上最受欢迎的 100 种硬件配置分析
**作者**: clem
**日期**: 2026-05-06T13:38:12.000Z
**来源**: [https://x.com/ClementDelangue/status/2052020105328890188](https://x.com/ClementDelangue/status/2052020105328890188)
---
刚刚发布了一个有趣的新数据集：Hugging Face 上 AI 构建者中最受欢迎的 100 种硬件配置，基于 297,135 名自愿填写其 HF 个人资料中本地硬件部分的用户：https://huggingface.co/datasets/clem/100_most_popular_hardware_setups_on_HF
关于方法论的说明，感谢 @JordanNanos 和 @SemiAnalysis_，他们正确地指出，解读这些数据最清晰的方式是将每一行归入四个互不相同的类别之一：独立显卡、SoC/APU、仅 CPU，或 CPU+GPU 组合。这避免了将 Apple 的 SoC 与独立显卡 SKU 进行比较，也避免了在组合中重复计算供应商。这 140,141 份前 100 名的用户报告细分为：60,120 名独立显卡用户（43%），50,077 名 SoC/APU 用户（36%），17,841 名仅 CPU 用户（13%），以及 12,103 名 CPU+GPU 组合用户（8.6%）。
长尾效应非常巨大。前 100 种配置仅覆盖了所有 297,135 名报告者的 47%。超过一半的 HF 构建者正在使用甚至未能进入前 100 的设备。本地 AI 领域的硬件碎片化是真实存在的。
每个供应商都占据了自己的类别。NVIDIA 以 97.3% 的比例主导独立显卡，Apple 以 95.6% 主导 SoC/APU，Intel 以 82.5% 主导纯 CPU 行。那种简单的“Apple 对决 NVIDIA”的比较具有误导性，因为它们实际上在不同的赛道上竞争。
CPU 领域呈现出显著的反转。在仅 CPU 用户中，Intel 以 82.5% 对 17.5% 领先。但在明确组装 CPU+GPU 主机的用户中，AMD Ryzen 以 65% 对 35% 领先。热衷者/ DIY 本地 AI 构建群体显然已经转向 Ryzen，尽管 Intel 仍然拥有更广泛的装机基础。
显存胜过原始算力。最受欢迎的独立显卡不是 4090 或 5090，而是拥有 4,737 名用户的 RTX 3060。特别是 12GB 版本，其显存比 3060 Ti、3070、4060 和 4060 Ti 更多，并与 3080 持平。12GB 的 RTX 3060 用户数量大约是 8GB RTX 3060 Ti 的四倍，尽管它们名字相似。AI 构建者关心的是内存大小，而不是基准测试分数。
AI 构建者明显倾向于 Pro 和 Max 芯片。对于 M3 系列，只有 22% 的用户使用基础款 M3。其余用户使用 M3 Pro、M3 Max 或 M3 Ultra。仅 M3 Pro（3,141 名用户）就比基础款 M3（1,968 名用户）更受欢迎。M1 也有同样的模式：使用 M1 Pro 的人（4,815）比使用基础款 M1 的人（4,499）更多。对于本地 AI 而言，更高的统一内存容量级别比对于通用计算重要得多。
10 系列显卡拒绝退役。GTX 1660、1650 Mobile、1070 Ti、1060、1080 Ti 和 1050 Ti，这些已有六到九年历史的 Pascal 时代的显卡，在前 100 名中总计约占 5,000 名用户。在 HF 上，大约每 13 个独立显卡用户中就有一个仍在使用 RTX 时代之前的硅片。
M4 系列的总用户数已经超过 M1 系列。M4 总计有 16,639 名用户，比 M1（12,444）多 34%。刚刚出货的 M5 已经达到 2,907。M2 代（6,917）处于奇怪的低位，用户比更老的 M3（8,967）还少，这表明在用户跳转到 M3 之前，M2 的采用窗口期很短。
50 系列是数据中增长最快的一代。RTX 50 系列已经有 13 个 SKU 进入前 100 名，总计 15,341 名用户。代际采用速度比同期的 30 或 40 系列更快。
数据中心/专业级芯片确实存在，但规模可能低于你的预期。H100、A100、H200、V100、T4、L4、L40s、RTX 6000 Ada、RTX PRO 6000 WS、A6000 和 GB10 总计 10,792 名用户（约占前 100 名报告的 7.7%）。几乎肯定是被低估了，因为研究人员通常不会在他们的个人资料中列出共享集群的 GPU。
新型 AI 专用芯片已经开始出现。NVIDIA 的 GB10 (DGX Spark) 位居第 36 位，拥有 1,241 名用户。AMD 的 Ryzen AI Max+ 395 (Strix Halo) 位居第 49 位，拥有 962 名用户。两者都是近期发布的，并且都已经在排名中占据了一席之地。
最受欢迎的发烧友组合：Ryzen 9 Zen 5 + RTX 5090（1,504），其次是 Intel i9 13 代 + RTX 4090（1,424）。
值得注意的注意事项：数据是自我报告和自愿参与的（偏向于积极参与 HF 的本地 AI 构建者）；用户通常只列出一台机器，即使他们拥有多台；云服务和共享集群工作几乎肯定被低估了；且不同供应商的标签粒度并不统一。
数据集 (Apache 2.0): https://huggingface.co/datasets/clem/100_most_popular_hardware_setups_on_HF
还没添加你的硬件配置？只需 30 秒：https://huggingface.co/settings/local-apps
让本地 AI 走起！🤗
## 相关链接
- [Hugging Face reposted](https://x.com/huggingface)
- [clem](https://x.com/ClementDelangue)
- [@ClementDelangue](https://x.com/ClementDelangue)
- [9.7K](https://x.com/ClementDelangue/status/2052020105328890188/analytics)
- [local hardware section of their HF profile](https://huggingface.co/settings/local-apps)
- [https://huggingface.co/datasets/clem/100_most_popular_hardware_setups_on_HF](https://huggingface.co/datasets/clem/100_most_popular_hardware_setups_on_HF)
- [@JordanNanos](https://x.com/@JordanNanos)
- [@SemiAnalysis_](https://x.com/@SemiAnalysis_)
- [https://huggingface.co/datasets/clem/100_most_popular_hardware_setups_on_HF](https://huggingface.co/datasets/clem/100_most_popular_hardware_setups_on_HF)
- [https://huggingface.co/settings/local-apps](https://huggingface.co/settings/local-apps)
- [Upgrade to Premium](https://x.com/i/premium_sign_up)
- [9:38 PM · May 6, 2026](https://x.com/ClementDelangue/status/2052020105328890188)
- [9,751 Views](https://x.com/ClementDelangue/status/2052020105328890188/analytics)
---
*导出时间: 2026/5/6 23:00:04*