麻省理工学院研究人员利用Ferveret冷却平台推进AI核材料工作

AI 模型和平台

麻省理工学院研究人员利用Ferveret冷却平台推进AI核材料工作

在 Google 上的首选来源中添加 Unite.AI

AI研究可以在想法耗尽前很长一段时间就进入物理极限：一个实验室的计算量可以提供动力并保持冷却。 麻省理工学院的Ferveret 称研究人员正在利用其高性能计算平台开发用于研究核材料的AI方法，将其适应性相冷却技术用于持续的科学工作量。

10月22日的公告将两个工程挑战联系起来。 研究人员需要更多的计算能力来理解暴露在极端条件下的材料。 数据中心操作员需要从有限的电力能力中提取更有用的工作。 Ferveret的主张是，更好的除热可以帮助两者，而不需要新一代的处理器。

为什么核材料研究需要更多的AI计算

这项工作涉及麻省理工学院极端环境实验室的材料，由核科学与工程与电气工程与计算机科学助理教授Ericmoore Jossou领导。 麻省理工学院的院系简介描述了将先进的显微镜和连贯的X射线成像相结合的研究，以在纳米尺度分辨率下检查结构性质和缺陷。

Ferveret利用它的平台确定两个项目。 一是制定AI方法，分析对核材料进行一致的X射线测量。 其他训练机器学习的原子间潜力，模型可以预测原子如何相互作用，帮助研究人员在要求的条件下模拟物质行为。

第一个项目的研究方向得到了独立的支持。 麻省理工学院2026年的核科学与工程研究博览会列举了东亨隆的演讲，题为“BCDI阶段检索的物理-成型深基因模型 ”。 BBCDI代表了布拉格连贯的疏松成像，乔苏的院系研究描述也确定了这一技术。 尽管事件上市并不评价Ferveret的硬件

原子间潜力解决了不同的计算瓶颈。 训练有素的模型不是在模拟的每个阶段进行昂贵的电子结构计算，而是估计原子安排、能量和力量之间的关系。 麻省理工学院寄存处进行的研究描述了机器学习力场如何可以延长原子模拟的时间和长度尺度，同时保留所训练的参考计算精度。

这种准确性仍然取决于模型及其所了解的情况。 Ferveret的公告称支持的研究旨在模拟接近密度-功能-理论精确度；它没有提供错误表、公布的模式或数字比较，以确定这些项目的结果。

适应性冷却如何消除热量

Ferveret的技术描述以次冷沸腾为中心，这是由核反应堆热转移研究参考的方法。 芯片表面的热量形成蒸气泡，但周围液体仍然低于沸腾温度。 公司称小气泡迅速分解，并凝结回冷却液中，刷新与热表面的接触。

区别在于蒸汽的流向。 在饱和-沸腾的浸泡方法中，Ferveret比较一下，气泡在凝固之前升入蒸汽空间。 它的方法试图在液态内将它们在当地凝固。 称公司支持在环境压力和紧凑的封闭装置附近运行，服务于单个服务器。

其数据中心产品页描述了一个名为Lambert的架式4U冷却装置，这是一个自定义的热交换器和控制软件，可以调整操作条件。 预定的部署是服务器规模的冷却系统，与设施的防热基础设施相连，而不是一个大型的共用浸水罐。

对AI操作员来说，目的是在延长工作量期间将处理器保持在热限度内，同时减少与冷却有关的间接费用。 较冷的硬件可以在热量否则会限制它时保持更高的操作性能。 效益的规模仍然取决于用于比较的处理器，工作量，电源设置和冷却系统。

15%和35%的实绩数字是多少？

该公告的两条头条数字描述了不同的测量。 Ferveret报告，与领先的直达芯片冷却方法相比，服务器的计算量增加了15%，如果包括设施一级的改进，则增加了35%的可用计算量。

Ferveret的3月公告提供了第一个数字的更多细节。 它描述了加州大学洛杉矶分校使用NVIDIA H200 GPU的智能连接实验室的基准，报告每千瓦服务器功率的TFLOP计算效率提高了15%。 发布内容明确排除了该服务器衡量标准中的设施级功率使用效能。

这是有用的背景，但它仍然是公司对合作的叙述。 释放不能提供完整的基准测试协议或足够复制比较的原始结果。 每瓦浮点吞吐量提高15%，也没有自动确定培训时间或应用产出提高15%。

较大的数字结合了服务器效率与设施模型。 在Ferveret的例子中，比较从1.20的功率使用效率，即PUE，转移到1.03，然后应用15%的服务器效率改进。 PUE是将总设施功率除以IT设备功率；较低比例意味着特定IT负载的设施间接费用较少。

利用上述投入，综合收益约为34%:1.15乘以1.20/103。 Ferveret将结果描述为约35%. 算术解释了索赔的规模，但它是基于假设的设施效率，而不是在麻省理工学院的核材料实验中公布的35 % 速度的测量。

证据的终结和机会的开始

在所提供的公告中，Jossou 称访问该平台可以实现更大的模型，更快的迭代和连续的实验。 这些意见支持额外计算访问的实际价值。 它们并不孤立冷却带来的改进程度，硬件分配产生的改进程度，或某项科学任务运行的速度。

这一区别同样适用于可持续性。 Ferveret)描述了一种闭锁-闭锁系统，通过干冷却器可以拒绝加热而无需蒸发水消耗。 这是关于冷却配置的说法。 不应将其理解为消除发电、设备生产或各种可能设施设计的水足迹。

该公告还预期称未来的研究出版物将承认Ferveret对基础设施的贡献。 公布的方法和应用层面的结果将有助于确定平台如何影响每个完成的实验的重建准确性、模拟尺度、培训时间和能量。

目前最强烈的结论是重点突出：费尔韦雷特正在为一项研究方案提供基础设施，而麻省理工学院证实了这项研究方案的AI和核材料重点，而其更广泛的效率声称需要注意测量界限。 更长期的机会是使稀缺的计算能力更进一步——将热传动的改进转化为同一电力预算范围内的更科学的工作。

本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。

AI Models & Platforms

MIT Researchers Use Ferveret Cooling Platform to Advance AI Work on Nuclear Materials

Add Unite.AI to your preferred sources on Google

AI research can run into a physical limit long before it runs out of ideas: the amount of computing a laboratory can power and keep cool. Ferveret says researchers at MIT are using its high-performance computing platform to develop AI methods for studying nuclear materials, putting its Adaptive Phase Cooling technology to work on sustained scientific workloads.

The October 22 announcement connects two engineering challenges. Researchers need more computational capacity to understand materials exposed to extreme conditions. Data-center operators need to extract more useful work from limited electrical capacity. Ferveret’s proposition is that better heat removal can help both, without requiring a new generation of processors.

Why Nuclear Materials Research Needs More AI Compute

The work involves MIT’s Materials in Extreme Environments Lab, led by Ericmoore Jossou, an assistant professor in Nuclear Science and Engineering and Electrical Engineering and Computer Science. MIT’s faculty profile describes research combining advanced microscopy and coherent X-ray imaging to examine structural properties and defects at nanoscale resolution.

Ferveret identifies two projects using its platform. One develops AI approaches for analyzing coherent X-ray measurements of nuclear materials. The other trains machine-learning interatomic potentials, models that predict how atoms interact and help researchers simulate material behavior under demanding conditions.

There is independent support for the first project’s research direction. MIT’s 2026 Nuclear Science and Engineering Research Expo lists a presentation by DongHun Ryu titled “Physics-Informed Deep Generative Models for BCDI Phase Retrieval.” BCDI stands for Bragg coherent diffraction imaging, a technique also identified in Jossou’s faculty research description. That establishes a concrete research connection beyond the cooling company’s announcement, although the event listing does not evaluate Ferveret’s hardware.

Interatomic potentials address a different computational bottleneck. Rather than performing an expensive electronic-structure calculation at every stage of a simulation, a trained model approximates the relationship between atomic arrangements, energies and forces. Research held in MIT’s repository describes how machine-learning force fields can extend the time and length scales of atomistic simulations while retaining the accuracy of the reference calculations on which they are trained.

That accuracy remains conditional on the model and the situations it has learned to represent. Ferveret’s announcement says the supported research aims for simulations approaching density-functional-theory accuracy; it does not provide an error table, a published model or a numerical comparison establishing that result for these projects.

How Adaptive Phase Cooling Removes Heat

Ferveret’s technology description centers on subcooled boiling, an approach informed by nuclear-reactor heat-transfer research. Heat at a chip’s surface forms vapor bubbles, but the surrounding liquid remains below its boiling temperature. The company says small bubbles detach rapidly and condense back into that cooler liquid, refreshing contact with the hot surface.

The distinction is where the vapor goes. In the saturated-boiling immersion approach Ferveret compares against, bubbles rise into a vapor space before being condensed. Its approach seeks to condense them locally within the liquid. The company says this supports operation near ambient pressure and compact enclosures serving individual servers.

Its data-center product page describes a rack-mounted 4U cooling unit called Lambert, a custom heat exchanger and control software that adjusts operating conditions. The intended deployment is a server-scale cooling system connected to the facility’s heat-removal infrastructure, rather than a large shared immersion tank.

For AI operators, the purpose is to keep processors within thermal limits during extended workloads while reducing cooling-related overhead. Cooler hardware can sustain higher operating performance when heat would otherwise be limiting it. The size of the benefit still depends on the processors, workload, power settings and cooling system used for comparison.

What the 15% and 35% Performance Figures Actually Mean

The announcement’s two headline figures describe different measurements. Ferveret reports up to 15% more server-level compute compared with leading direct-to-chip cooling approaches, and up to 35% more usable compute when facility-level improvements are included.

A March announcement from Ferveret provides more detail on the first figure. It describes benchmarking with UCLA’s Intelligent Connectivity Laboratory using NVIDIA H200 GPUs, reporting a 15% improvement in computational efficiency measured in TFLOPs per kilowatt of server power. The release explicitly excludes facility-level power usage effectiveness from that server metric.

That is useful context, but it remains a company-issued account of the collaboration. The release does not provide a full benchmark protocol or raw results sufficient to reproduce the comparison. Nor does a 15% improvement in floating-point throughput per watt automatically establish a 15% improvement in training time or application output.

The larger figure combines server efficiency with a facility model. In Ferveret’s worked example, the comparison moves from a power usage effectiveness, or PUE, of 1.20 to 1.03, then applies the 15% server-efficiency improvement. PUE is total facility power divided by IT-equipment power; a lower ratio means less facility overhead for a given IT load.

Using those stated inputs, the combined gain is approximately 34%: 1.15 multiplied by 1.20/1.03. Ferveret describes the result as approximately 35%. The arithmetic explains the scale of the claim, but it is a scenario based on assumed facility efficiencies, rather than a published measurement of a 35% speedup in MIT’s nuclear-materials experiments.

Where the Evidence Ends—and the Opportunity Begins

In the supplied announcement, Jossou says access to the platform enables larger models, faster iteration and continuous experiments. Those observations support the practical value of additional compute access. They do not isolate how much improvement came from cooling, how much came from the hardware allocation, or how much faster a particular scientific task ran.

The same distinction applies to sustainability. Ferveret describes a closed-loop system that can reject heat through dry coolers without evaporative water consumption. That is a claim about the cooling configuration. It should not be read as eliminating the water footprint of electricity generation, equipment production or every possible facility design.

The announcement also says future research publications are expected to acknowledge Ferveret’s infrastructure contribution. Published methods and application-level results would help establish how the platform affects reconstruction accuracy, simulation scale, training time and energy per completed experiment.

For now, the strongest conclusion is a focused one: Ferveret is supplying infrastructure for a research program whose AI and nuclear-materials focus is corroborated by MIT, while its broader efficiency claims require attention to measurement boundaries. The longer-term opportunity is to make scarce computing capacity go further—turning improvements in heat transfer into more scientific work within the same electrical budget.

Auto-fetched and translated by AI Atlas. All rights belong to the original author.