[News] NVIDIA Adds Samsung Foundry to NVLink Fusion Ecosystem for Custom Silicon Manufacturing

After the recent NVIDIA–Intel collaboration—which lets Intel build x86 CPUs that plug directly into NVIDIA platforms via NVLink Fusion—Team Green is expanding its NVLink ecosystem in a surprising move, now bringing in Samsung Foundry. According to NVIDIA, Samsung will help meet growing demand for custom CPUs and XPUs, providing full design-to-manufacturing support for tailored silicon.

At the 2025 OCP Global Summit in San Jose (Oct 13–16), Ian Buck, VP of HPC and Hyperscale, explained that beyond the existing Vera Rubin OCP rack, NVIDIA has been working to integrate other processors into the MGX and OCP infrastructure using a solution called NVLink Fusion — an IP and chiplet technology that lets CPUs and other accelerators seamlessly join the NVLink and OCP rack design.

At the event, NVIDIA showcased key players in the NVLink Fusion ecosystem, starting with Intel, which can now build NVLink Fusion-enabled CPUs that connect seamlessly to NVIDIA GPUs in data center environments. Fujitsu also joins the ecosystem, with its Monaka processor able to communicate directly with NVIDIA GPUs over NVLink Fusion using NVLink protocol, creating a tight CPU–GPU integration for high-performance workloads.

Notably, NVIDIA drew attention by revealing its custom silicon partners. According to the company, developing these chips requires a full ecosystem—including custom silicon designers, CPU, and IP partners. Its CPU partners already include Intel, Fujitsu, and Qualcomm, and NVIDIA is now adding Samsung to the lineup, joining Alchip, AsteraLabs, Marvell, and MediaTek as part of its custom silicon ecosystem.

NVIDIA notes that Samsung Foundry has partnered with the company to meet growing demand for custom CPUs and custom XPUs, offering design-to-manufacturing experience for custom silicon.

However, as TechPowerUp reported, the collaboration on NVLink Fusion comes with strict limits: any custom chip must link with an NVIDIA product, and NVIDIA controls the critical software that manages these connections. Partners cannot build fully independent, mix-and-match systems, as NVIDIA keeps control over the communication controller and PHY layers, and requires a license for third-party hardware to use its NVLink Switch chips, the report noted.

Vera Rubin Update

At the 2025 OCP Global Summit, NVIDIA also provided more details for Vera Rubin. According to Ian Buck, the platform is expected to go online in the second half of 2026. Unlike a single-chip design, the platform reportedly features two chips: a CPX processor for context processing and the Rubin GPU for AI workloads.

Buck noted that Vera Rubin could deliver over 8 exaflops of inference performance—about 7.5 times more than GB200—with increased memory capacity. The platform also supports 400G scale-up networking, providing up to 260 terabytes per second of aggregate bandwidth across all GPUs.

Read more

• [News] NVIDIA Unveils NVLink Fusion, Amplifying ASIC Push with MediaTek & Marvell Onboard

(Photo credit: TrendForce, at OCP Summit)