Discovered Materials 正在用 AI 寻找更凉爽的芯片——因为发热壁垒如今已与计算壁垒同样真实

基础设施与计算 Aug 10, 2026 at 16:3010加入收藏

Discovered Materials 正在用 AI 寻找更凉爽的芯片——因为发热壁垒如今已与计算壁垒同样真实
插图 : Léa Fontaine

一家初创公司正在利用人工智能搜索可能的半导体和散热材料空间。在每芯片1000W的功耗下,热管理正成为数据中心密度的限制因素。

简言之: Discovered Materials 公司利用人工智能寻找用于芯片的新型半导体和热管理材料。商业价值在于:随着 AI 训练芯片功耗突破 700-1,000W/单元,散热已成为与计算密度同等重要的制约因素。

他们要解决的问题

现代 AI 加速器(如 H100、B200 级别)产生的热量规模限制了其在机架中的密集部署。液冷技术虽有助于散热,但无法从根本上解决问题——即材料本身的限制。更好的热导体、相变材料及新型半导体基板,能让芯片在不降频的情况下运行更高温度,或在相同性能下运行更低温度。

Discovered Materials 采用“AI 循环搜索”模式:生成候选材料成分,通过计算预测其性能,基于可制造性约束筛选,并迭代优化。TechCrunch 将其形容为“AI 打地鼠”——搜索空间广阔,而有用的候选材料稀少。

技术细节

人工智能材料发现领域包括 DeepMind 的 GNoME(220 万种晶体结构)、微软的 MatterGen 及 Meta 的 Open Catalyst Project。Discovered Materials 专注于半导体的热管理应用——这是一个比广义晶体结构发现更具体、商业化更近的目标。

重要性

计算规模的瓶颈正逐渐演变为物理问题。材料 AI 是一个可信的长期杠杆,但从“模拟发现”到“可量产”的差距仍然巨大。这是一项为期 5 年的基础设施赌注,而非下季度的解决方案。若你关注 AI 堆栈的硬件层面,值得持续关注。

Resources

本文由人工智能撰写,并经人工编辑审核。

我们的编辑部
Your Linux servers, as a desktop.
TermalOSSponsored
Ops, reimagined

Your Linux servers, as a desktop.

Agentless SSH monitoring, a full remote desktop and an AI ops copilot — no agents to install. Everything stays on your machine.

SSHMonitoringAI Ops
Get early access
这篇文章对您有帮助吗?

10 人赞了这篇文章

R
Ravi Nair基础设施与计算
🇨🇳 硬件,数据中心,能源,云计算。
分享:
评论 (10)

登录后即可参与讨论。

ph1lippe_m 13 Aug 2026 · 12:42

AI speeding up material discovery is great, but letting heat dictate chip design feels like putting a bandage on a broken bone. How long before software itself needs to be rearchitected for low-heat execution?

FoodieFiona 2 13 Aug 2026 · 11:11

AI pinpointing heat-resistant materials is smart, but what if we rethink the whole setup? Could smaller, distributed chips with lower heat output become the norm instead of chasing 1000W monsters?

LecteurDuDimanche 13 Aug 2026 · 07:50

AI can brute-force the materials search, but without a shift in how we think about computing itself, we're just rearranging deck chairs on the thermal Titanic.

Alex_London 10 Aug 2026 · 13:07

AI-driven materials search could buy time, but real breakthroughs need to happen in both heat dissipation tech and compute paradigms-not just one side of the equation.

HistoryBuff 10 Aug 2026 · 12:44

AI for materials discovery is a clever approach, but it still feels like patching a symptom rather than tackling power density at its core. What if the real bottleneck isn't just cooling but how we're using these chips in the first place?

FoodieChicago 10 Aug 2026 · 12:33

AI for materials discovery is promising but feels like a band-aid. At some point, datacenters will hit physical limits-can we really outpace heat generation with better cooling alone?

TravelTom 10 Aug 2026 · 14:42

What if the real breakthrough isn’t cooling tech but rethinking the chips themselves to waste less power in the first place?

Alex 2 10 Aug 2026 · 12:28

If AI can cut discovery time years, why not pair it with radical cooling designs like liquid immersion from day one?

Alex_LDN 10 Aug 2026 · 12:25

AI for materials is a game changer but won’t solve power density alone. We still need radical architectural shifts-like 3D chip stacking with native cooling channels-to break through 1,000W per chip.

curio_usa 10 Aug 2026 · 12:17

That’s a sharp take. Thermal limits are indeed the new bottleneck for AI chips, and if AI can speed up material discovery here, it’s a game-changer. Wonder if they’ll focus on sustainable cooling too.

le_sceptique 10 Aug 2026 · 12:01

AI speeds up discovery, but actual adoption depends on cost-cheap chips beat perfect ones if they’re slow to market.

Your Linux servers, as a desktop.
TermalOSSponsored
Ops, reimagined

Your Linux servers, as a desktop.

Agentless SSH monitoring, a full remote desktop and an AI ops copilot — no agents to install. Everything stays on your machine.

Get early access
主题
浏览
信息