Samsung Cuts Chip Simulation 18 h to 41 s on Nvidia GPUs, ~1,600x

Nvidia says Samsung's chip thermal and mechanical analysis fell from about 18 hours on a CPU to 41 seconds on GPUs; SK hynix reports 30x on full-chip runs. Gains cover test cases, not chip design overall.

Chips & Power DeskEdited by Ege Erten

Nov 11, 2026, 04:36 UTC2 min readAI-assisted

X in r/HN

Key-facts card: Samsung chip simulation falls from 18 hours on CPU to 41 seconds on Nvidia GPUs, roughly 1,600x, a vendor claim — Samsung's reported chip-simulation result on Nvidia GPUs: about 18 hours to 41 seconds, a roughly 1,600x vendor claim for one case.
Illustration (AI-generated)

Samsung Electronics cut a chip-level thermal and mechanical simulation from about 18 hours on a 32-core CPU to 41 seconds on Nvidia GPUs, a speedup of roughly 1,600x, according to an Nvidia blog post published Nov. 9.

SK hynix said at the same event, Nvidia AI Day Seoul 2026 in Seoul's COEX on Nov. 9, that full-chip thermal and deformation analysis ran more than 30 times faster than its CPU-based method, according to E-Today's report of the session. Both companies built their models with Nvidia's PhysicsNeMo framework.

Bar chart of company-stated chip-simulation speedups: SK hynix more than 30x, Samsung in an Nvidia case 1,600x, a Samsung presenter nearly 4,000x — Company-stated speedups run from 30x to nearly 4,000x, measured against different CPU baselines.
Illustration (AI-generated)

Samsung's own presenter gave a wider range. Lee Sung-yeop of Samsung's semiconductor research institute said a job that took about a day on a multi-core CPU ran in tens of seconds to about a minute on single or multiple GPUs, a gain of as little as 1,000-fold and nearly 4,000-fold at most, iNews24 reported. DigiTimes headlined the result as up to 3,800-fold. The 1,600x figure is the one Nvidia itself states for a single case.

How the speedup was reached

Nvidia's own description shows the gain is not a plain CPU-to-GPU port. Samsung combined three changes: AI-predicted effective material properties for each small region of the chip, a coarser computational grid, and GPU acceleration. The baseline was a fine-grid solve on the CPU, so the comparison changes both the method and the hardware. Lee said the result needed both an improved algorithm and good hardware such as Nvidia GPUs.

SK hynix split full-chip runs across several GPUs and reused results computed for small regions. Its team lead Kim Jung-han said performance rose almost in proportion to the number of GPUs, and an AI model predicting etch profiles ran dozens of times faster than conventional simulation, E-Today reported.

Why it matters for the supply chain

The analysis targets defects caused by heat and warping, which matter more as devices shrink and advanced packaging stacks several chips together. Faster solves let engineers test more design and process variants before committing wafers. That is a development-cycle gain for HBM and logic makers, not a performance claim about any chip they sell.

The figures cover single calculations under test conditions. Neither company gave data on yield, time to market or cost, and the 1,600x claim should not be read as a speedup of chip development as a whole. The work follows the multiyear Nvidia and SK hynix partnership announced on June 7, which named PhysicsNeMo and CUDA-X for TCAD (technology computer-aided design) workflows. For earlier Samsung supply-chain coverage, see our report on its reported 2nm HBM base dies at Taylor and SK hynix's 6F2 DRAM plan.

What is not confirmed yet

The 1,600x, 3,800x and 4,000x figures use different baselines and GPU counts, and Samsung has not published a paper or the full benchmark conditions. The DigiTimes report, which gives the up-to-3,800-fold figure, was behind a paywall and was not read in full for this article. Samsung and SK hynix have not said when the methods will enter production design flows.