Why The Heck Is Intel Struggling To Make Smaller, Faster CPUs?

It looks as though Intel has some issues with its next generation of CPUs. Yesterday during an earnings call, Intel CEO Brian Krzanich admitted that Intel would be delaying the highly anticipated Cannon Lake processor until 2019. The delay means Intel's CPUs won't see a very large jump in speed or power efficiency any time soon.

Instead, we'll be getting the just-announced Whiskey Lake, a new CPU architecture, as well as a body of booze I'm sure more than one CPU designer at Intel would like to leap into.

Photo: Alex Cranz (Gizmodo)

We don't have many details on Whiskey Lake as of yet. Its existence was first reported by Motley Fool back in December 2017, and Krzanich yesterday largely confirmed that initial report. CPUs based on the Whiskey Lake microarchitecture will have a minimum of four CPU cores and two GPU cores, and they will be implemented using a 14nm process. Intel first announced a 14nm CPU way back in 2013, and has effectively been on some iteration of the process since later 2014.

The process, or node, is highly critical to how fast and power efficient a CPU is. A 14nm process means the most important components of the CPU are on 14nm die. A 10nm die would be smaller and pack all the components of the processor closer together. That would mean electrical signals don't have to move as far, which means the CPU becomes faster, and more power efficient.

So moving to a smaller die is always a very good thing! But getting smaller also means getting more complex. You're having to pack all the same goodies on a much smaller plane!

Intel's having a lot of trouble doing that. That's why the move from 14nm to 10nm has now been delayed twice. Intel has currently been stuck on 14nm since the launch of Broadwell way back in 2014.

That's a really long time to stick to the same CPU process! For reference, Intel's competitor, AMD, moved from 14nm to 12nm in just one year, and has promised to be on 7nm by the end of next year. Back in February 2018, the CTO of the foundry that produces AMD's silicon, Dr Gary Patton of GlobalFoundries, told Anadtech he thinks Intel's goal of using 10nm at all is misguided to begin with. "I mean if you look at the scaling and the performance, it is a pretty weak node."

Intel clearly disagrees as it claims it's shipping some 10nm products now (just not consumer-facing), and Brian Krzanich insists the problem isn't with the node. Rather it's more a problem because Intel set very high targets for 10nm. "You're almost 10-per cent more aggressive" Krzanich said during the call.

But being that aggressive isn't a good thing if you can't get the product in the shape necessary to ship, and this problem is a long-standing one for Intel.

Even the move to 14nm in 2014 was delayed by a year, which might be one of the many reasons Intel announced the hiring of Jim Keller this week. Keller is leaving Tesla for Intel, but before that he worked at AMD, where he was one of the chief architects of the highly successful Zen architecture, and before that at Apple he did the same kind of magic to create Apple's first internally developed CPUs for the iPhone and iPad. The guy is good, and if anyone can help Intel with its die-shrink problems, it's Keller.

Whether that's actually the task Keller's been hired to accomplish remains to be seen. He doesn't join Intel officially until April 30, and because Cannon Lake has been in development since before 2014 (that was when it was officially added to Intel's roadmap), it's unlikely he'll play any kind of significant role in its design or development.

[Intel]


Comments

    As the gate pitch and interconnect pitch is almost identical between the 10nm format and proposed 7nm, Intel could also be taking the view that it's not really worth the additional complexity of going smaller.

    Why are they struggling? Because it's hard and they want to get it right?

    Anyway, something the article fails to point out is the Intel and AMD processes aren't an apples to apples comparison. From what I understand the 10nm process used by Samsung is roughly equivalent to the 14nm process used by Intel in terms of transistor density (the number transistors per square mm).

    So while it looks like AMD have a huge lead in process size they really don't. They probably do have a lead but it's not as dramatic as you might think (which is backed up by the relative performance of the chips).

    Interesting article (from last year) with more details; https://www.extremetech.com/computing/246902-intel-claims-three-year-advantage-10nm-process-wants-change-define-process-nodes

Join the discussion!

Trending Stories Right Now