Nvidia uses MCM packaging on GB100

At last year’s illustrious Arete Technology Symposium, NVIDIA’s Vice President and Chief Director of Accelerated Computing, Ian Buck, reaffirmed NVIDIA’s unwavering commitment to biennially refreshing their principal GPGPU architecture, officially announcing the debut of the Blackwell architectural GPU in 2024. It is widely anticipated that the GTC of 2024 may serve as the grand curtain-raiser for the Blackwell architecture, earmarked for products within the data center and artificial intelligence domains, while the consumer-oriented GeForce graphics cards may linger until 2025 for their grand unveiling.

Recent whispers among the digital cognoscenti suggest that the Blackwell architecture’s GB100 might embrace a more diminutive chip design, incorporating the MCM (multi-chip packaging), heralding a significant leap in NVIDIA’s product lineup. Additionally, sources intimate that while there may not be a discernible augmentation in the number of GPCs or TPCs for the Blackwell architectural GPU, profound metamorphoses are anticipated at the unit architecture level.

From the initial murmurings of the Blackwell architecture, it was postulated to be NVIDIA’s inaugural foray into the realm of smaller chip designs. However, subsequent conjectures posited that NVIDIA might remain steadfast in its allegiance to the monolithic design approach. In truth, both monolithic and smaller chip designs each bear their unique merits and drawbacks. Yet, weighing the equilibrium of performance enhancement against cost and efficiency, industry titans like Intel and AMD are veering toward the smaller chip designs, synergized with state-of-the-art packaging methodologies.

It warrants mention that the Blackwell architectural GPUs, crafted with the smaller chip design in mind, predominantly cater to the data center and artificial intelligence sectors, with known models being GB100 and GB102. For the consumer-grade GeForce graphics cards employing the Blackwell architecture, the trajectory seems inclined toward preserving the monolithic design, falling under the GB200 series.

Rumors are rife that NVIDIA is currently appraising Samsung’s 3nm GAA process, with projections aligning for mass production by 2025. However, it seems this technology might not be the cornerstone for the Blackwell architectural GPUs. At least for products destined for data centers and the AI sector, the allegiance appears to remain with TSMC’s foundry capabilities, while Samsung’s might be tapped for other product endeavors.