NVIDIA 3nm GPUs expected to launch in 2025 via TSMC, marking next-generation roadmap.

NVIDIA's next-generation GPUs based on TSMC's 3nm process node are likely to be delayed until 2025, according to DigiTimes citing industry sources. The pushback is attributed to a slowed PC market and global economic tensions. The company will not be introducing new GeForce GPUs this year and possibly not the next, as it completes the transition to its Ada Lovelace architecture. Mainstream parts are expected in the coming month with potential refreshes such as the RTX 4090 Ti and RTX 4080 Ti, while the laptop segment may see updates at the CES 2024 show floor.

NVIDIA faces significant inventory challenges that are delaying consumer GPU production. The company is holding substantial stock and does not intend to produce many gaming GPUs until existing supply begins moving. Recent price hikes on GeForce RTX GPUs have made consumers reluctant to purchase mainstream cards, and the availability of older GeForce RTX 30 series GPUs at lower prices in the used market further dampens demand for new products.

At GTC 2023, NVIDIA announced new Hopper products and volume production of its 4N-powered data center chips now shipping to supercomputers, HPC customers, and AI segments. With Hopper demand remaining strong—these chips cost 10-20 times more than average consumer GPUs—the company is focused on producing and delivering these products before transitioning to new GPU architectures. Currently, gaming GPUs use the 5/4nm family while AI GPUs use 7nm and 5/4nm families, with the industry estimated to enter the 3nm generation by 2025.

NVIDIA's next-generation architecture, potentially codenamed Blackwell, is already deep in development. According to leaker Kopite7kimi, Blackwell is likely to be announced at next year's GTC but will not see production until close to 2024 or early 2025, mirroring Hopper's timeline from initial announcement to volume shipments. Blackwell may not utilize the 3nm process node and is expected to feature a monolithic design, potentially launching on TSMC's 4nm process instead, with the generation after Blackwell possibly utilizing 3nm or sub-3nm technology.

NVIDIA is leveraging its AI capabilities to accelerate next-generation GPU development. Using machine learning, the company aims to make chips "Better Than Humans," and with cuLitho technology, chips can be designed and manufactured 40x faster than traditional designs. NVIDIA is collaborating with ASML and TSMC to adopt this technology.