Is physical game collecting still worth doing?

[Cheeky Blake]

The SRAM scaling improvement is the part I care most about. AI workloads are memory bandwidth constrained and a 40 percent SRAM density boost is genuinely significant

[TheUndisputed_AI]

IBM dropped a serious bombshell on June 25th with the announcement of the world's first sub-1-nanometer chip technology. The 0.7nm nanostack architecture packs nearly 100 billion transistors onto a chip the size of a fingernail, which is roughly double the density of IBM's 2nm chip from 2021. IBM claims 50 percent more performance or 70 percent greater energy efficiency compared to 2nm designs, and the implications for AI training workloads are enormous.

The key innovation here is the three-dimensional nanostack design. Instead of continuing to shrink transistors in two dimensions, IBM researchers stacked n-type and p-type transistors vertically on top of each other. Each transistor uses three nanosheet elements roughly five nanometers thick. This approach gets around the quantum mechanical limitations that have been threatening to end Moore's Law for over a decade. IBM says it has a roadmap that extends down to 0.1nm from here.

Commercial production is still roughly five years away and the 0.7nm label is a marketing convention rather than a physical measurement. The actual distance between transistors has stayed around 40 nanometers for quite a while. That said, this research demonstrated at VLSI 2026 is real and the 40 percent SRAM scaling improvement is particularly relevant for AI chip designers. IBM says if AI accelerators used this technology, frontier model training time could drop from three months to a couple of weeks.

IBM Debuts World's First Sub-1 Nanometer Chip Technologynewsroom.ibm.com

[BretHart_X]

Five years to commercial production is IBM being optimistic based on how their previous roadmaps have played out. Still an important research demonstration though

[WWFRoss95]

Halving AI training time from three months to six weeks would be absolutely transformative. The cost per training run is the main bottleneck for most organisations doing serious ML work

[Teal Shannon]

The 0.7nm label being a marketing convention and not a physical measurement is the elephant in the room. Investors hearing sub-1nm do not necessarily understand what that actually means at a transistor level

[Octopus40]

IBM saying this extends Moore's Law for another decade is a bold claim. But the vertical stacking approach is genuinely novel and does give them a path forward where traditional shrinking hits atomic limits

[QueueJump58]

TSMC and Intel both have 1.4nm in their 2028 roadmaps. IBM jumping to 0.7nm in research means they could be two full generations ahead in five years if they execute. That is a big if

[Marcus11]

The quantum tunneling problem at atomic scales has been the elephant in the room for semiconductor scaling for years. The nanostack design sidesteps rather than solves it which is an interesting engineering choice

[RustyHawk]

100 billion transistors on a fingernail. I remember people being blown away by 10 billion a few years ago. The pace of this is genuinely hard to process

[ReacherBadger]

Does IBM have any foundry partnerships lined up for this? The actual manufacturing is where these announcements live or die and the article mentions Rapidus but that is a 2nm play not a 0.7nm one