We always knew the limits to Moore‘s Law are first and foremost economic. Given an industry used over decades to predictable lowering of price per compute function and thus swallowed any advance for new user functions and overhead when the limits are reached there is going to be a squeeze. AI scaled up at the time the production capacity became more inelastic.
Maybe it is time not just shrink transistors but also software bundles. I can see decades of possible progress hiding in plain sight behind a browser screen.
AI data centers are eating like 80% of memory.
Making user space applications more memory efficient is not even going to be a rounding error on memory demand.
I am with you that it needs to happen, but it's not going to solve a memory shortage.
It would make memory-poor phones more viable. Like why can't we have a 512MB, or even 256MB RAM phone. Although I doubt that the software effort would be cheaper than just buying the extra RAM. It's definitely much more uncertain.
Those exist, they're just feature phones. Smartphones have to run general purpose apps and that takes RAM.
That was my point, make it so that apps can run on that RAM budget. At least some apps.
That's newly fabbed memory vs. an existing stock. The stock is quite massive, so optimizing existing use and enabling it to be repurposed can be meaningful.
Not really...
If the new memory is needed for AI data centers, it doesn't matter if your existing MacBook doesn't need as much memory anymore.
If the existing MacBook needs less memory, it can use its current memory spec for higher-level uses that formerly required a MacBook Pro. That meaningfully affects the market.
You don't seem to understand supply and demand.
That does not help the demand side for AI data centers - which is the vast majority of the market...
If your point is purely about supply and demand for datacenter HBM and LPDDR, you're probably right. Local model inference (using the existing memory stock) can make a dent in current use, but not in projected future uses that will plausibly involve much larger models.