The future of AI Computing and HBM: a new era of performance

NVIDIA – GTC 2025 highlighted that AI is at a transformative juncture, expanding beyond traditional data centers into edge applications. This shift is driving unprecedented demand for HBM, a trend highlighted in Yole Group’s latest memory and computing market reports: Next Gen DRAM 2025: Focus on HBM and 3D DRAM and Gen AI 2025 – Computing and AI for Data Center.

The stars of GTC 2025 were undoubtedly Nvidia’s upcoming flagship GPUs, Rubin and Rubin Ultra, set to launch in H2-2026 and H2-2027, respectively. These GPUs mark a substantial evolution as they are manufactured using TSMC’s 3nm process, compared to 4nm for Blackwell and Hopper. Rubin series GPUs introduce Nvidia’s second generation of GPUs based on a chiplet architecture, featuring from four chiplets per package for Rubin to eight chiplets per package for Rubin Ultra.

Rubin Ultra stands out with its impressive computational capabilities, delivering 100 PFLOPS in FP4 workloads, about five times the performance of the preceding Blackwell generation. Even when assessing FP16 dense workloads, Rubin Ultra provides around 16,000 TFLOPS, over seven times the capability of Blackwell. However, considering the significant difference in chiplet count between Rubin Ultra (4-GPU chiplets) and Blackwell (2-GPU chiplets), the single-die-to-single-die performance improvement remains remarkably substantial, estimated at 3-4 times superior to Blackwell.

In parallel, Nvidia introduced Vera, its new ARM-based CPU that also leverages TSMC’s 3nm node. Vera is Nvidia’s first CPU to feature multi-threading, offering two threads per core, significantly enhancing performance and efficiency compared to its predecessor, Grace. Vera comprises 88 customized ARM cores, resulting in 176 threads, a considerable evolution from Grace, which was based on the Neoverse V2 architecture with single-threaded cores.

Moreover, Nvidia continues to assert its dominance through its sophisticated rack-scale solutions, notably ahead of AMD and Intel, who are only expected to debut competing solutions in 2026. Nvidia’s latest rack-scale system, the NVL576, combines 144 Rubin Ultra GPUs (576 GPU chiplets) and 144 Vera CPUs, delivering unmatched processing density and efficiency.

Nvidia’s announcements suggest a GPU market that continues to intensify around performance per single chip, leading to a strong concentration of computing to maximize overall AI system performance. As a result, the total amount of silicon dedicated to AI is expected to continue increasing. However, if Rubin Ultra GPUs are widely adopted by late 2027 or early 2028, it would not be surprising to see a shift in the overall GPU volume.

The rapid adoption of AI workloads, particularly in data centers and edge computing, is fueling the need for high-performance memory solutions. The transition from AI model training to inference workloads—such as agentic AI and AI-powered PC workstations—further underscores the importance of optimized memory architectures.

Leading the HBM market are key players SK hynix, Samsung, and Micron.

Explore a selection of key announcements compiled by Yole Group’s memory and computing teams, set to significantly shape the industry’s future:

Other news

This year, the market research and strategy consulting company Yole Group once again offers a comprehensive collection of market and technology reports as well as reverse engineering and costing analyses and dedicated market monitors, all related to the memory and computing domain. They continue their methodical investigation of this interconnected ecosystem to better understand the strategies of semiconductor companies and assess their remarkable advancements.

Engage with Yole Group’s experts at any time to gain deeper insights into the semiconductor memory industry. Ask for a meeting.

Josephine Lau (Jo) is Senior Technology & Market Analyst, Memory at Yole Group.

Josephine is engaged in in-depth analysis of the dynamic memory market and co-produces Yole Group’s NAND and DRAM market research.

Prior to joining Yole Group, Josephine had 10 years of experience in the memory & storage industry with expertise in business strategy planning, financial analysis, and business development. This includes roles in investor relations and strategic sourcing at Seagate’s headquarters in the San Francisco Bay Area.

Josephine holds a Bachelor of Business Marketing with distinction from the Royal Melbourne Institute of Technology and has passed Chartered Financial Analyst (CFA) level 1.

Hugo Antoine is a Technology & Market Analyst, Computing and Software at Yole Group.

Hugo develops technology & market analyses covering computing hardware, software, and Artificial Intelligence (AI).

He holds a master’s degree from École des Mines de Saint-Étienne (France), with a focus on microelectronics and computing at the Centre de Microélectronique in Provence, France. In addition, he pursued an AI specialization at the École Polytechnique de Montréal in Canada. Furthermore, Hugo completed a dual-degree program in innovation management at EM Lyon Business School, highlighting his expertise at the intersection of technology and business.

Market & Technology reports

Technology & Cost reports

Comparison reports

As well as the Market Monitors:

i

i

Source: www.yolegroup.com