Generative AI meets RF circuit design = game changer. • Passive networks tailored by diffusion models. • Specify stop-band/pass-band; AI does the rest. • Pixel patterns are not intuitive to electrical response. Designs getting more abstract. Prepare for a cognitive shift.

I quite like this table. Probably misses a lot of nuance tho.

Process node is 180nm. Just like me. Old.

Guess India named their first indigenous chip after me. Lol.

After wathcing @highyieldYT's latest YT video, I'm beginning to think I should look into the technical side of energy as much as I look at chips and semi on my substack.

Who has heard of this law?

All this slander about Substack and how the in-app purchases are raising cost to subscribers by 30%. Don’t wanna pay the apple tax? Just go on your laptop and upgrade. Simple. Smh.

Read here:.

The memory wall is the real bottleneck in AI hardware. GPUs today can deliver massive TFLOPS, but they sit idle waiting for data. What bridges the gap is High Bandwidth Memory (HBM) — the only memory tech that can keep up with modern AI training. HBM is so critical that it.

Very long post drops this Sunday ⚓

I got too nerdy on HBM. From DRAM memory cells to hybrid bonding. Gory detail you didn't know you wanted to know. Publishing on Sunday on my newsletter. 👇🏽.

d-Matrix announced their 3D In Memory Compute chip. → Logic die sits on top of DRAM (likely custom).→ The interconnect is so short between memory and compute that bandwidth is supposedly 10x that of HBM4. → Energy efficiency is also supposedly better (<1 pJ/bit?). There is no

AI context windows need to get large enough just so that people can prompt self help books into 2,000 word essays.

latest model but data from a year ago? what use is this?

guys what am I doing wrong

RT @dylan522p: Feel like I'm taking crazy pills. We are just back at step one. Don’t store KV cache, just recompute it.

Read more:.