Hacker News Reader

Comment by ashtonkem | Hacker News Reader

Intel is literally on an older fabrication process right now, they’re on 10nm while AMD is on 7nm. They’re objectively behind.

View on HN

You're giving Intel a little too much credit. They're only "on 10nm" for tiny, low-power laptop chips. The yields are reportedly still too low to use it for desktop or server CPUs.

Meanwhile in a couple of months AMD will be on 7nm+

(However, it should be said that Intel 10nm is roughly equivalent to TSMC 7nm in terms of density. AMD is still ahead but not by as much as the numbers would imply).

Meanwhile AMD & TSMC are talking about 5nm for their Zen 4 architecture.

The nm numbers are marketing. Processes are too nuanced to boil down to a single number.

Process isn’t everything, but they’re not just marketing. The process has a huge impact on the heat & power characteristics of the core, and also controls how much stuff you can stuff into a die.

The point he's making is that 10nm, 7nm, etc don't actually denote anything that lets you do an apples-to-apples comparison anymore. At this point, they're not measures the way they were in legacy process nodes.

They haven't been apple to apples in over 20 years. But being stuck on one feature size for more than 2-3 years is still a major roadblock to performance improvement.

And if one insists on comparing, Intel 10nm is kind of TSMC 7nm. However Intel only manages to make a few of their laptop parts with their 10nm, and with not great frequencies IIRC. So they are behind now.

I thought that it used to measure gate width or something like that. Did that stop being true?

It never stopped being true, but a lot of what makes a processor faster these days is smarter layout, better branch prediction, and better microcode. Process size matters, but it’s probably less important than it was in the 1990s.

Also, there's a difference between the minimum feature size that can be fabricated (mostly a lithography challenge), the minimum size of a reliable device that works well (e.g. with enough doping atoms in the channel) and the size and shape a specific transistor needs to be for its particular requirements.

Yep

https://en.wikichip.org/wiki/technology_node

"Recent technology nodes such as 22 nm, 16 nm, 14 nm, and 10 nm refer purely to a specific generation of chips made in a particular technology. It does not correspond to any gate length or half pitch."

"At the 45 nm process, Intel reached a gate length of 25 nm on a traditional planar transistor. At that node the gate length scaling effectively stalled; any further scaling to the gate length would produce less desirable results. Following the 32 nm process node, while other aspects of the transistor shrunk, the gate length was actually increased."

That is literally not how process nodes work. They are objectively matched.

Nodes mean 3 things: density, elecrical performance (ion, max f, leakage, etc), and processing flexibility (how many supported feature primitives or designs).

AMD and Intel do not use the same node designs.

A generalization, but accurate enough. Disappointing to see it getting downvotes. Intel 10nm and TSMC 7nm have similar characteristics, as a brief Wikipedia search reveals.

Reply on news.ycombinator.com