Intel’s 14nm milkshake: It’s better than yours

At yesterday’s  Analyst  Day,  Intel  broke out some details on its upcoming 14nm process, the state of that technology, and just how far along the  ramp-up  it is.  Having recently  announced  that   Broadwell’s 14nm   deployment  would be somewhat delayed by yield issues, it’s possible that the  manufacturer felt  the need  to address rumors of serious problems in the fab line — or maybe it just wanted to gloat.

To hear Intel tell it, 14nm yields aren’t the only thing ramping well. The process node is on target to deliver the kind of marked improvements in cost, performance, and density that have become dicey at foundries like TSMC or GlobalFoundries.

The left graph is a yield graph showing 22nm and 14nm graphed on the same time scale. As you can see, 22nm takes a dip in June — that’s the launch of Haswell — but quickly recovers. The dip at the very beginning is likely related to the Haswell volume ramp, which would have pulled total yields down compared to Ivy Bridge alone. We can see where Intel ran into trouble on 14nm in June, July, and August. That’s possibly the result of ramping production at new facilities for the first time. So if Fab A was building 14nm in May and June, Intel spins Fab B up in July and August, yields take a whack before coming back up to projected rates. Alternately, the company may have hit a deployment milestone on 14nm that hurt yields at multiple fabs at the same time.

On the right graph, up the side, we have “Switching Energy Change” — that’s a measure of how much power the CPU is using. Across the bottom, “Delay Change” — how much power the chip uses. This graph tells us the following, in aggregate:

• Intel’s 14nm process consumes less power than the 22nm process, at every point on the curve.
• At the far right, the 14nm chips take a far smaller frequency penalty, but still draw significantly less power. 22nm Haswell gave up about 65% of its frequency to hit a 25% switching speed.

If that doesn’t seem to square with Intel’s published TDPs and clock speeds, keep in mind we’re seeing one part of the curve against an arbitrary 100%. But a quick glance at Intel’s Ark does show us the relationship between TDP and clock. The 15W dual-core i7-4550U has a 1.5GHz base clock and a 3GHz Turbo Mode. The 37W i7-4600M is a dual-core chip with a 2.9GHz base clock and a 3.6GHz Turbo Boost. What this chart implies, at the bottom, is that Broadwell will see clock gains at the same TDP rates.

What about the left-hand side of the equation? The implication here is that Broadwell will also scale better than Haswell. If so, that bucks the trend we saw with 22nm Ivy Bridge and Haswell itself, where high-end overclocking has become increasingly dicey. If the chip Intel is using for comparison is a mobile processor, it’s entirely possible that both curves continue bending sharply to the left, until there’s essentially no improvement by the time we arrive at desktop TDPs.