Hacker News Reader

Comment by kllrnohj | Hacker News Reader

Dr. Lisa Su took over as CEO in 2014 and seems to have driven a re-focusing at AMD, devoting the company's relatively few R&D resources effectively. Starting with large growth in non-PC markets, actually, mostly the consoles.

Jim Keller was certainly key in getting Zen up & going, but he left in 2015 so his impact on Zen 3 would be rather minimal to say the least.

Combine that with the timing of Intel struggling hard with new process nodes and there you go. A bit of AMD got better, a bit of Intel got worse, and now you've got this landslide result.

View on HN

Yeah, it's worth noting that Intel's planned microarchitecture iteration on the desktop was delayed by over three years - from 2017 to 2021 - and in 2021 will be released on a larger process than originally planned.

AMD's execution looks meticulous by comparison, and they deserve all the praise and profits resulting from it. But their win is inflated by Intel completely imploding. Until next year they're merely iterating on 2015 technology - their Skylake architecture - on the desktop and in servers at least. This is not to say that Intel's 2021 chips will be competitive with this (they won't be) but if Intel with better management would have been at this point in, say, 2018 instead of 2021 this would have played out very differently.

The 10nm delay is of course legendary at this point, but 14nm was itself also about a year behind schedule. In early 2013 Intel stated Broadwell 14nm would be out that year. It didn't come out until Q4 2014. And Intel's Fab 42, which was originally scheduled to open in 2013 as their first 14nm fab, only just opened this year. Intel ended up needing to re-fit existing fabs to be 14nm instead.

So before they faceplanted with 10nm Intel first stumbled with 14nm. 2 generations in a row of issues.

Proper leadership would have taken this risk seriously, and would not have tied up microarchitecture development so much to process improvement. Basing your strategy so fully on Moore's Law keeping up is a recipe for disaster. Clearly Sunny Cove works on 14nm (that's what they're releasing next year!), and they had all the technology available to release that in 2018 too - maybe with lower clocks and core counts.

Seems like a classic textbook case of a large coorporation not understanding how to change their ways, ending up being beaten by a more agile smaller competitor. Oh well, hindsight is 20/20.

> Jim Keller was certainly key in getting Zen up & going, but he left in 2015 so his impact on Zen 3 would be rather minimal to say the least.

Highly doubt that he did so in a technical role, maybe in a PM like role, but not a design lead. The team lead was Suzanne Plummer, architect Michael Clark.

I was saying many times, if you hire a JK level cadre, it's best to not spend his very expensive time just doing technical work.

Apple also helped a lot indirectly. They have been spending billions on R&D for their SoC's and outsourced the production of the waffers to TSMC which used Apples R&D money to contribute to their 7nm and 5nm node (though Zen 3 is "still" on 7nm).

Although AMDs and Dr. Lisa Su's achievement are not to be underestimated, I doubt the 7nm processes would be as mature as it is right now without Apple.

I think that's vastly overstating Apple's importance. TSMC has had a significant, steady cashflow from far more than just Apple over the years. Nvidia would if anything be the more likely one pushing TSMC's top end, as Nvidia makes by far the largest, highest performance dies out of anyone on TSMC. Apple's latest & greatest A14 only just hits the transistor count on TSMC's 5nm that Nvidia was pushing 4 years ago on TSMC's 16nm.

Apple's extra cash certainly didn't hurt, but TSMC wasn't struggling before Apple came along, either.

And prior to 7nm the other major fab, Global Foundries, as perfectly competitive with TSMC despite not having any Apple money.

If Nvidia was such a key partner for TSMC then why did they get priced out of 7nm? Nvidia made some huge chips on 16nm and 12nm, no doubt. I’m not so sure the transistor count comparison is all that meaningful though. Maybe compare the wafer allocation instead.

Clearly TSMC is heavily diversified and gets to pick their clients. Apple has also been the primary (volume) launch partner on both 7nm and 5nm, which to me indicates how much TSMC values that partnership. Imagine the slam dunk Nvidia would have had if the RTX 3000 series was on 5nm.

> If Nvidia was such a key partner for TSMC then why did they get priced out of 7nm?

They wanted more margins? But note that Nvidia does still use TSMC's 7nm for their largest & most expensive dies. The A100 is TSMC 7nm at a staggering 54 billion transistors on 826mm² of silicon. Nvidia retains the largest die manufactured on TSMC's 7nm. By a lot. The next largest would I think be Navi 21 at 536 mm² and 26.8 billion transistors.

> Apple has also been the primary (volume) launch partner on both 7nm and 5nm

Apple's die sizes & transistor counts are comparatively tiny. The first use of a new fab is pretty much always a small, low-power die. That's what yields best when yields are the lowest.

See also why Snapdragons are also among the first to launch on a new TSMC node, despite those SoCs having probably the lowest margins of the high-volume stuff TSMC makes.

I’m not convinced that Nvidia having the biggest die size is a particularly important metric here. It’s obvious to me that TSMC would rather produce tiny Apple chips and AMD Zen chiplets to for yield and wafer space efficiency. We all know that smaller rectangles pack better into circular wafers than big rectangles.

Apple could have made the A14 much bigger and faster than the A13, but they chose to cram more chips on the wafer, which I speculate they did to have more wafers available for their bigger upcoming iPad and Mac chips.

I think you have it backwards on margins. Per square mm it’s clearly more expensive to produce an A100 die than an A13 or Snapdragon SoC due to the wasted die space and need for golden samples on the Nvidia side of things where they’re not selling any cut down dies in GeForce cards.

> TSMC would rather produce tiny Apple chips and AMD Zen chiplets to for yield and wafer space efficiency.

TSMC sells wafers, not functioning dies. They don't care how you slice it up or how bad you yield as a result.

I was going to mention this. The point is that Apple and AMD buy more wafers.

It’s rumoured that Nvidia’s deal with Samsung was for working dies, not total wafers.

They did make a habit of badmouthing TSMC whenever they had problems. That probably played a role in TSMC allocating their scarce 7nm capacity away from NVidia when it was in short supply.

A cpu needs multiple years from their idea to the tape-out. So its possible Zen 3 is still based on ideas by Keller. But only AMD empoyees would know...

Reply on news.ycombinator.com