At GTC 2025, NVIDIA unveiled its future product roadmap, announcing that the next-generation GPU following Blackwell and Rubin is named Feynman. This cutting-edge GPU, set to debut in 2028, will incorporate a new generation of high-bandwidth memory (HBM), further advancing AI and high-performance computing efforts. Named after Richard Feynman, the revered American theoretical physicist renowned for his monumental contributions to quantum mechanics, quantum electrodynamics, and particle physics, the name continues NVIDIA's tradition of honoring scientific luminaries.

One of Feynman's standout features is its memory technology. Although specific details remain under wraps, NVIDIA has confirmed it will utilize next-generation HBM memory, potentially HBM4e or HBM5. Known for its high bandwidth and energy efficiency, HBM is ideal for AI training and inference. The upcoming HBM4 standard aims to double the number of channels per stack compared to HBM3, with expected bandwidth increases to 6.4Gbps or more. As a potential successor, HBM5 may overcome performance bottlenecks, offering robust support for Feynman. Such a memory enhancement will boost data throughput and help GPUs tackle increasingly complex computational tasks.

Complementing Feynman will be the Vera CPU, previously debuted with the Rubin and Rubin Ultra platforms. The Vera CPU boasts a custom 88-core ARM architecture supporting 176 threads, alongside a high-speed 1.8TB/s GPU connection via the NVLink interface. This CPU-GPU co-design optimizes AI workloads, enhancing computational efficiency. NVIDIA plans to integrate 8th-generation NVSwitch NVL-Next network technology within Feynman, along with Spectrum7 and CX10 network components. Spectrum7 offers 204T of throughput while utilizing CPO (co-packaged Optical) technology, and the CX10 enhances data center interconnectivity. These advancements, slated for release with Feynman in late 2028, promise to push NVIDIA further ahead in the data center arena.

Looking at NVIDIA's recent technology trajectory, the unveiling of Feynman is no major surprise. Since the introduction of Blackwell GPUs in 2024, NVIDIA has maintained a steady annual update cadence. By late 2025, Blackwell Ultra promises 288GB of HBM3e memory and 15 petaflops of FP4 compute performance; in 2026, the Vera Rubin platform is expected to feature HBM4 memory, delivering 50 petaflops of inference per chip; and by 2027, Rubin Ultra will further expand to 1TB of HBM4e memory and 15 exaflops of inference performance on a single rack. Feynman, a continuation of this sequence, will inherit and surpass its predecessor's technologies, targeting the AI and HPC market in 2028.

While Feynman's precise specifications remain unannounced, its predecessors provide some insights. For instance, Rubin Ultra features a 4-reticle-sized GPU design paired with 12 stacks of HBM4e memory, offering total bandwidth of 13 TB/s. Feynman could adopt a similar multi-chip packaging approach or scale-up to meet larger AI model training needs. Additionally, NVIDIA's strategic investment in advanced packaging technologies should not be underestimated. Feynman may leverage TSMC's 3nm process, combined with CoWoS-L packaging, for greater transistor density and energy efficiency.

Beyond hardware, Feynman's launch will benefit from NVIDIA's robust software ecosystem. The CUDA platform and NVIDIA Inference Microservices (NIM) are set to continue optimizing AI development processes, ensuring hardware performance is maximized. By 2024, NVIDIA's GPUs held nearly 90% of the global data center market share, and Feynman's arrival will undoubtedly reinforce this lead, especially as generative AI and robotics rapidly evolve.

Importantly, Feynman isn't the roadmap's final stop. NVIDIA has hinted at a subsequent Feynman Ultra, potentially debuting around 2030, which could further widen the gap between NVIDIA and its competitors. Currently, AMD's Instinct MI series and Intel's Gaudi processors are gaining traction in the AI sphere, yet NVIDIA remains at the forefront, thanks to its technological foresight and ecological integration.

From Blackwell to Feynman, each step of NVIDIA's journey has centered around AI computing's core needs. Feynman not only symbolizes hardware evolution but also foreshadows future computing paradigms. The year 2028 promises to bring thrilling developments for tech enthusiasts with this GPU's introduction, propelling the industry to new heights.