Intel’s 14nm Core M detailed: Broadwell delivers huge performance, dramatic power reduction

When Intel first unveiled its 14nm process, the company made aggressive claims about its performance and power consumption. The delays and fab cancellations that followed took a great deal of wind out of its sails, leaving us to wonder if the 14nm delays would push back the entire product stack. Now Intel is striking back against its naysayers — today, the company has released an enormous amount of information on its upcoming Broadwell Y product, the laptop- and tablet-oriented Core M. Let’s dive into Core M, and also take a look at what it might mean for the desktop Broadwell chips that will surely follow.

Everything but the kitchen sink

According to Intel, it’s rearchitected nearly everything about the Core M in order to hit its aggressive power and performance targets. The memory controller, L1/L2 caches, I/O, PCI Express interface, and even the chip’s packaging have all been overhauled.
The Core M (codenamed Broadwell Y) will debut on Intel’s low-power 14nm process variant, which helps account for the chip’s dramatically improved power consumption — Intel argues it can deliver a greater-than 2x TDP improvement, 50% smaller packaging, a 25% physically smaller board, and a 60% reduction in SoC idle.
Broadwell basics

Intel is also claiming that Broadwell Y’s SRAM scaled to 58% its 22nm size when built on 14nm. That’s particularly impressive because it cuts against what other foundries are expected to deliver. Compare these claims below against what’s expected for the mainstream foundries like Samsung and GlobalFoundries (Intel first). Intel SRAM scaling

It’s a bit small, but you can see that as clock speed increases, cache size improvements decrease sharply. At 500MHz, 16nm SRAM (green line) is 1.6x more dense than 28nm — but by 1.5GHz it’s only about 10% faster. If Intel pulled off an SRAM density this large, some of its claims about matching the foundry partners may stand up to scrutiny.
Everything about Broadwell Y is more efficient, from its FIVR (fully integrated voltage regulator) to its power management schemes. The company has also implemented a new throttling mechanism that’s supposed to deliver performance more efficiently.
Say Broadwell needs to deliver the equivalent of 200MHz “worth” of performance over four seconds. Because there’s a minimum voltage threshold required to run a CPU core, the power-saving impact of lower clock speeds eventually flatlines — running a chip at 100MHz vs. 200MHz may not result in any effective power savings at all.
Broadwell Y power cycling

Instead of running at 50MHz for four seconds, Broadwell Y will run at 100MHz for two seconds, then idle for the other two seconds. The dotted line represents the average clock speed over time — by bursting up and dropping off, the chip conserves more energy than it would save by sitting at a flat low clockspeed.

Performance (CPU and GPU)

On the CPU side, don’t look for much — Intel claims Broadwell is up to 5% faster than Haswell thanks to some minor tweaks to the TLBs (translation lookaside buffers), better address prediction, and a larger scheduler. Real life impact is likely to be in the 2-3% range. That would be normal — Intel’s die shrinks tend to offer only minimal performance improvements.
Graphics, however, is a bit different.

Add caption

Intel is claiming a 20% increase in compute units and a 50% increase in shader throughput with better overall thermal headroom. This implies that final performance uplift could reasonably be between 20-30% — but it’s not clear from this graph which solution Intel is referencing. The company’s current low-watt parts use the Intel HD Graphics 4200 solution which a maximum frequency of 850MHz and 20 EUs.
Since the GPU in Intel’s current top-end chip (the Core i7-4770K) currently clocks up to 1.25GHz, it’s possible that Broadwell’s ultraportable GPU performance jump will simply put it on par with Intel’s high-end desktop solution. That would still be a significant and noteworthy increase — but it won’t exactly reinvent the wheel as far as GPU performance is concerned. Whether or not it can keep up with a native solution like Tegra’s K1 is something we’ll have to wait for benchmarks to determine.

Add caption

The Llama Mountain reference design tablet, for the Broadwell Y (Core M) SoC

Intel’s Great Leap Forward

There’s a lot to get excited about in these slides — if Broadwell Y delivers like Intel is promising I don’t think its an exaggeration to say it could redefine what thin-and-light means to a lot of people in the x86 space. If Broadwell Y delivers on time and with the performance and power consumption characteristics Intel is claiming here, it’ll punch into all-new form factors and could light a fire in the thin-and-light market.
Whether or not it takes on ARM in the tablet space, well, that’s a different question altogether. This wouldn’t be the first time that Intel unleashed a vastly improved product that failed to grab much market attention. Qualcomm has taken over the LTE modem business on the strength of integrated radio in tablets and smartphones alike, and Intel isn’t expected to have a 28nm radio ready in volume until into 2015. The overall competitive position between ARM and x86 is complicated by OS, software, and form factor considerations — the two architectures have yet to go head-to-head in the way most analysts expected back in 2011.
None of this is to say that Broadwell Y won’t be a huge jump forward for Intel’s traditional x86 business, but after multiple years of watching the company’s Android attempts fizzle, it’s just not clear what it will mean for Intel’s larger business. On the desktop side of things, the 5% IPC improvement will certainly be better than a wet slap in the face. The more important question is whether Broadwell will improve on Haswell’s max clock speeds — and for that piece of info, we still have quite a few months to wait, I’m afraid.