it baffles me how people only consider what is currently “verifiable”. the verification tools of the future are not limited to the domains that exist today
At the end of the day chip design is just RTL + knowing how to use Synopsys/Cadence. Blackpillingly, it is also verifiable, so highly hill-climbable via RL.
Hardware engs might be ousted to analog or RF if they wanna keep up with the times. Claude can't solder yet :)
memory on logic will be really good for memory going forward imo. effectively couples the price of memory and compute since you have to buy them together now
Our first-gen @Qualcomm HBC silicon is validated in our San Diego lab and exceeding expectations. Placing compute closer to memory tackles the bandwidth constraint in AI inference, and it’s coming to Qualcomm Dragonfly AI250. Proud of this team.
Read more: https://t.co/ttrFKjBoCt
cerebras issues make a lot of sense to me intuitively when u think about cerebras power inefficiency (already known) and what we already know about electromigration, thermal expansion etc
@NTMonX@Busybissm scaling is limited, their advantage becomes smaller as more of the data runs between wafers. as we scale to larger models this will naturally happen
Introducing XPU Grasshopper, an AI chip co-designed with AI for the era of recursive self-improvement
It went from its first tokens to 816x faster in 13 weeks
Pre-order the FPGA devkit: https://t.co/9w28BYtO6D
so many uninformed investors nowadays; back when apps and such were the rage mayyybe you could read an article and understand roughly the value proposition. there’s no way you understand how the semiconductor industry will progress from a 2 minute article
@sirloinfurr@EkaterinaJanss1@beniduboss do you understand why there’s no bottleneck on sram supply? they also 100% do suffer that issue they literally said themselves the way they’re scaling in the future is by incorporating more dram