As for PC desktops and laptops, the one thing this announcement doesn't touch on is what it means for LPCAMMs. 24 bits/die does not make for a very wide DIMM. So are we going down to 96b CAMMs? Or are the Big 3 going to make double-wide (48b) chips with 2 dies?.

If it feels like the server industry has infiltrated the LPDDR industry, you aren't wrong. There's a ton of RAS (reliability) features here that are in DDR5 for server hardware. Just in time for things like SOCAMM. Expect (a lot) more LPDDR use by servers later this decade.

In some respects, the new spec is a grab-bag from different technologies. Further refinements of LPDDR power savings. The return of LPDDR4's sub-channels. The ECC and reliability features of DDR5, including Per-Row Activation Counting (PRAC).

LPDDR6!. JEDEC has finally released the full spec. Lots of fun, weird stuff in this one: chips are now only 24b wide instead of the usual 32b, split up into two 12b sub-chans. 32B of data per packet is meta data. On die ECC. And more power efficiency.

So GlobalFoundries is acquiring MIPS, eh? I can't say I saw that one coming. On the surface it looks like an odd marriage: an outdated fab and one of so very many RISC-V vendors. But I'm interested to see what GloFo can do with the IP. There is potential.

RT @ChipsandCheese9: Hello you fine Internet folks,. Today's article is on Nvidia's RTX PRO 6000 Blackwell and diving into the Blackwell ar….

So after almost 20 years, Comcast has finally eliminated data caps on their network nationwide. I'm not sure I ever thought I'd see the day. Of course, they cut speeds and raised prices - so there's that classic Comcast twist. But still, no caps!.

"We look forward to changing the world together. Thank you for joining us today".And that's a wrap! Please see our full piece on AMD's announcements on the More Than Moore substack:

AMD is moving faster than ever before to deliver the best AI solutions for the market. It's been a pace unlike anything before in modern computing. "The future of AI is not going to be built by any one company"

And that's Sam Altman. Back to Lisa, here's the rack roadmap for 2026 and 2027. Helios will be followed by a next-gen rack with Verano CPUs and MI500 GPUs

Thanks in part to newer hardware, Altman also thinks they can maintain the same pace of model improvement in the second half of the decade as in the first half.

Altman is extremely excited for the MI400 series. They expect it will be awesome for inference, and could be great for training as well.

Enterprises are seeing "lots and lots of value" from AI.

Lisa; "Everyone listens to what Sam Altman has to say about AI".

And one last partner plug. Sam Altman, CEO of OpenAI

With Helios, AMD is bending the performance curve further. Up to 10x more performance for "frontier" models

It's NIC counterpart will be Velcano, an 800G NIC with UAL support

MI400 itself will have 20PFLOPS FP8, 432GB HBM4 memory, and nearly 20GB of memory bandwidth.

And feeding MI400 will be EPYC Venic. Zen 6 built on TSMC 2nm with up to 256 cores. AMD just got Venice back from the labs and it is "looking fantastic"

AMD wants to go toe-to-toe with NVIDIA's next-gen Vera Rubin racks