Samsung trims a chip thermal analysis from 18 hours to 41 seconds on NVIDIA GPUs
Summary
Samsung Electronics reported that one chip-scale thermal and mechanical analysis, which previously ran for about 18 hours on a 32-core CPU, now finishes in 41 seconds on a GPU. That works out to a speedup of roughly 1,600x. The gain came from three changes working together: material properties predicted by AI, a coarser simulation grid, and GPU acceleration.
The result was presented by NVIDIA at AI Day Seoul as part of a wider set of physical AI collaborations with Korean manufacturers. Kakao Mobility is working with NVIDIA on a Level 4 autonomous driving system. LG Electronics is applying the work to factory robots, with its CLOiD robot named in the headline.
Samsung is broadening its NVIDIA collaboration across materials research, process development and device engineering. In technology computer-aided design (TCAD), it is using NVIDIA's cuISS and cuDSS libraries to speed up the sparse linear systems that sit at the core of device simulation. StorageReview's account is cut off at that point, so the source gives no deployment scale, hardware counts, timelines or financial terms for any of the three partnerships.
Why it matters
The Samsung figure shows engineering simulation, not only model training, moving onto GPUs inside a major chipmaker. For global infrastructure players in Korea, that points to a second source of accelerated-compute demand from manufacturers running design, robotics and autonomy workloads. Such work may be hosted on-premises or in local facilities. Simulation of this kind needs enough GPU capacity to rerun analyses within an engineering workflow. Suppliers of capacity, power and cooling in Korea should track this demand alongside hyperscale training clusters, because it can shape where and how much compute those manufacturers buy.
Expert Lens
A speedup this large usually changes how a workload is run, not just how fast it finishes. Engineers who can rerun a thermal model in seconds tend to explore many more design variants, so total GPU hours may rise rather than fall. That pattern would push manufacturers toward steady, reserved capacity instead of occasional bursts. Proprietary chip and factory data also likely favors sovereign or on-premises hosting in Korea, which ties these projects to local grid connections and to cooling built for dense racks. The next signal to watch is whether Samsung, Kakao Mobility or LG disclose the size or location of the GPU clusters behind these programs.
Sources
This brief was drafted with AI assistance from public sources; every number is attributed to the sources above. See the AI-assisted Disclosure and Methodology.