Physical AI · Physical AI Simulation
Samsung Cut a Chip Simulation From About 18 Hours to 41 Seconds on Nvidia GPUs
The roughly 1,600x speedup came in one case, using AI-predicted material properties and a coarser grid.
· 2 min read
- NVDA
Nvidia said Samsung, Kakao Mobility and LG are moving physical AI toward production in Korea, and only Samsung’s result carries a number: one chip heat and mechanical analysis that took about 18 hours ran in 41 seconds.
The 1,600x is not a hardware speedup. The 18 hours was a fine-grid solve on a 32-core CPU; the 41 seconds was a coarse-grid solve on a GPU, with AI estimating the materials’ properties first. A manufacturer budgeting GPUs for simulation cannot expect that multiple from the chips alone, because the hardware, the grid and the method all changed together.
How far does the number travel?
Samsung’s figure comes from one case. Nvidia, which sells the GPUs and software involved, did not say how the coarse-grid answer compares in accuracy with the fine-grid one, or how many of Samsung’s simulations could follow the same path.
The only other Samsung number is smaller and covers different work. Samsung’s materials platform, sMOL, ran selected quantum chemistry calculations up to 5x faster with Nvidia software. That result is a ceiling for one kind of calculation, not a second reading of the same gain.
What are Kakao Mobility and LG doing?
Kakao Mobility plans to develop a level 4 self-driving system on Nvidia’s vehicle computing platform, starting with Kia purpose-built vehicles. It will pair its own driving software and its experience operating in Seoul’s Gangnam district with Nvidia’s technology, and it is aiming at future services in Korea. Nothing in the post suggests the vehicles are on the road yet.
LG demonstrated its CLOiD robot performing manipulation and manufacturing tasks in an actual manufacturing environment, using Nvidia’s robot model, with real-world data collected and augmented for training. The teams are using the work to build a repeatable way to train, test and improve factory robots. Neither company gave a figure for how much shorter robot training gets, which is the number a plant manager would need.
Samsung’s case suggests where to look in a company’s own engineering and robot-training cycle: which step takes longest, and whether a faster GPU, a coarser model or better data would shorten it?
