You are currently viewing SemiWiki as a guest which gives you limited access to the site. To view blog comments and experience other SemiWiki features you must be a registered member. Registration is fast, simple, and absolutely free so please, join our community today!
I do enjoy the Anastasi In Tech Youtube channel but something just does not look right. They are very professionally produced videos and very informative but they are highly edited and she does get some things wrong. Is she really a semiconductor design engineer? Maybe, but she does not know implementation or manufacturing even though she pretends she does. Thoughts?
Uh, her LinkedIn profile has no company name mentioned, at the top it says Austria, but then it also says Munich area. Her education was in Russia. A more credible person would certainly list the name of their company,, and the list of chips and foundry nodes that they have designed. She does know how to promote her YouTube channel, that's for sure.
I do enjoy the Anastasi In Tech Youtube channel but something just does not look right. They are very professionally produced videos and very informative but they are highly edited and she does get some things wrong. Is she really a semiconductor design engineer? Maybe, but she does not know implementation or manufacturing even though she pretends she does. Thoughts?
In recent years, compute-intensive artificial intelligence tasks have prompted creation of a wide variety of custom hardware to run these powerful new systems efficiently. Deep learning models, suc…
She said Blackwell is on TSMC N4 versus TSMC N3 due to low yield. No, when Blackwell was started the N4 PDK was the only 1.0 PDK available for high performance design. Intel 4 was not ready and Samsung 4 had serious yield problems. Nvidia is using TSMC N3 for their next generation of chips due out in 2025/26 and will also use TSMC N2. And her explanation of yield is adorable....
She said Blackwell is on TSMC N4 versus TSMC N3 due to low yield. No, when Blackwell was started the N4 PDK was the only 1.0 PDK available for high performance design. Intel 4 was not ready and Samsung 4 had serious yield problems. Nvidia is using TSMC N3 for their next generation of chips due out in 2025/26 and will also use TSMC N2. And her explanation of yield is adorable....
While it wouldn't surprise me if N3 HPC wasn't ready in time, and N3E will almost certainly not launch products until the back half of the year. However we need to divorce the idea from our minds that "bad yields" on an 800+ mm^2 die means bad defect density. If TSMC can do 10% yield on a die that big with blazing fast DRAM drivers I think that would be very impressive given how fast it is after the lead product. If TSMC could achieve good yields (by the standards of maximum die sizes) on N3, then there is no reason that apple couldn't have used N3 on A series SOCs 1 or even 2 years ago. NVIDIA sticking to a older node with mature yield is normal. Look at their prior generations of GPUs. Ampere N7 DC chips launched in 2021 (like 11mo after the Samsung manufactured gaming chips), and I don't think anyone in their right mind would have said that N7 had sucky defect densities (in an absolute sense) in 2020. I looked back to their 16FF products and this trend holds for every datacenter GPU. For kicks and giggles I also checked intel's lead product on the rapidly yielding 22nm; Ivy Bridge. The client products started launching in 4/12, while the larger Xeons didn't launch until around a year later.
Now to be absolutely clear, my point is not to poo-poo N3 yields. My point is that yield on big dies start dropping off exponentially with die size. "Good yields" (god I hate saying "yield" instead of defect density if we are talking about a process rather than a product) are not good enough for something as big as BWL. It needs mature yields. The fastest way to get there is for TSMC to launch small easy to yield dies as fast as possible with Apple to drive faster volume ramp and as a result faster yield learning with more data turns. This is NOT something exclusive to TSMC, this is how things have worked and will always work since the very first ICs at Fairchild.
I watched some of her videos a few years ago; she gets information from the usual sources that tech news tubers often do, including the misunderstanding of facts that come with that. It’s OK “lazy” news..
I mean you could argue Nvidia Blackwell also isn’t available on Intel 14A because of low yields..
engineering.fb.com
