Nvidia has tapped TSMC to build its next-generation Rosa CPUs, which are expected to arrive in 2029.
The chip was detailed only a day earlier, and reports are now turning to which process technology Nvidia wants for the beast.
According to Commercial Times Nvidia is considering two TSMC process options, either 2nm or A16.
TSMC’s N2P is already lined up for chips including AMD’s EPYC Venice and Intel Nova Lake. Chipzilla, Samsung and TSMC are all trying to make 2nm sound like the only process worth having.
For Rosa, Nvidia is more interested in the A16 because of its Super Power Rail feature, which includes backside power delivery.
The trick gives nanosheet transistors a cleaner split. The front side handles signal and clock distribution, while the back side deals with power distribution.
Industry insiders reckon Rosa is likely to use TSMC’s 2nm process family and may include A16 back-end power supply technology.
That would push up demand for chemical mechanical polishing, or CMP, along with the consumables needed to keep that manufacturing circus moving.
Super Power Rail on A16 is designed to improve logic density and performance, with an 8-10 per cent speed bump compared with N2P.
It is said to offer 15 to 20 per cent power savings at the same speed and a 10 per cent increase in chip density, all within the same layout footprint.
Using A16 would lift demand for CMP processes and benefit Taiwanese supply chain outfits such as Suntech Power Semiconductor and Suntech Power Technology. Carrier wafer demand is expected to rise too.
Rosa is the follow-up to Nvidia’s leading-edge AI CPU push. Nvidia has said the chip will use a new core architecture called “Rigel”, based on the Arm v9.2 CPU core.
Like Vera, Rosa is expected to be a fast AI-processing CPU, chasing even more “max single-threaded” performance at scale than its predecessor.
Vera uses Olympus cores, custom Armv9.2-A designs that offer twice the throughput of Grace. Rosa is supposed to push single-core performance further without gobbling more silicon area.
Grace has been shipping since 2023 with 72 Arm Neoverse V2 cores. Vera is due in 2026 with 88 custom Olympus cores and 176 threads through spatial multithreading.
Rosa is expected to arrive with Nvidia’s Feynman GPUs, with custom Rigel cores, larger L2 cache than Olympus and a focus on single-thread performance.
The memory side is still less clear. Grace uses LPDDR5X with ECC; Vera moves to LPDDR5X with ECC and SOCAMM, or to LPDDR6; while Rosa is expected to move towards LPDDR6 or LPDDR6X.
Grace offers up to about 480GB/s to 512GB/s per CPU. Vera is expected to reach up to 1.2TB/s, while Rosa’s numbers are secret.
The wider design direction is obvious enough. Nvidia wants CPUs that can feed AI systems faster, with less latency and better single-thread performance for agentic AI loops.
Rosa should reach data centres by 2029, followed by PC-specific variants in Rosa Feynman Spark systems by 2030.
