In the first IDF keynote of 2013, Intel CEO Brian Krzanich introduced a new sub-Atom CPU called Quark. While the keynote was woefully short on details and tech, SemiAccurate managed to dig some of that up anyway.
The promise was a sub-Atom device to compete with, well nothing directly but it does address a massive market that by unit volume could potentially dwarf phones. There were no specs, dates, device details, process tech, ISA, or even die size mentioned. When we hunted down a chip it looked like this.
Baytrail on the left, Quark on the right, $.45 (base18) piece below
If nothing was directly stated about Quark, what was intoned? Officially the device is 1/5th the size of an Atom, 1/10th the power, and it is synthesizable was the extent of things. 14nm was strongly hinted at, x86 wasn’t mentioned, and a bunch of nebulous use cases and a reference design board was briefly shown off. That said, what are the real Quark specs?
Imagine if you took a venerable Pentium CPU like the P54C which was originally made on a .6µm process and shrunk it down to a modern 32nm process. That is a hot topic of debate amongst geeks when the Star Trek physics battles peter out for the evening but no one is daft enough to actually do that, right? Well until Quark the answer would be yes, but Quark is said to be based off a P54C on a 32nm process.
Because of that it is small enough, the core, not the chip, is 1/5th the size of an Atom core and the core very likely only does pull 1/10th the juice of an Atom. The entire chip however takes more but how much more and what that is relative to an Atom’s uncore is not known but the gap should decrease quite a bit when the full chips are compared. On the up side though that entire debate is now over so Federation cosplayers can go back to Warp Factor 9+ science minutia.
Quark is also said to be synthesizable which usually means a design that is portable across different foundries. When asked about making Quarks at different places, Brian Krzanich made it quite clear that the core was a black box with open interfaces, not a portable design that can be modified. While no process tech was stated, SemiAccurate did confirm that it was and will be made on a 32nm process initially. Partners can add their IP to the design but they can’t peek in to the core, modify it, or move it outside of Intel to manufacture.
What does a buyer end up with? A 32nm P54C Pentium class CPU with an updated uncore and possibly a bit more added IP. How much do they pay for a Quark? Sources say about $5 and up but not below, in the cheap microcontroller market that is a tough ask for a 32-bit device. Adding external IP to a device like this is even a more tenuous proposition, mask costs for a 28nm device are closer to $10 million than $5 million, how many $5 not million devices would you have to make at what margins to pay for the masks alone? What odds do you see of hitting one out of the park like that? Would you risk it with your company?
That brings up the last point also related to process size. The $5 and up controller market is indeed huge but it is also cost sensitive, usually pin bound, and not too performance sensitive. Some are power sensitive but not as much as the <$1 market, a $5 chip is too powerful to run off a battery for multiple years. This means power is not a make or break proposition, nor is it a lever that people will pay notably more for.
This is why most people in this industry make their $5 controllers on far older processes, not a -1 or -2 node but a -5 node can be just fine. 130Nm? 90Nm? Likely fine. Why pay $5000 for a wafer if you don’t have to, much smaller node will do just as good a job for a fraction of the cost. If you are paying $5000 for a wafer you need to get more than 1000 $5 devices off of it if you are going to make a positive return. That means about 50mm^2 or less per device, 250mm^2 on a $1000 -4 node wafer. Before you say that you can fit more transistors with higher performance on the $5000 wafer, if you are pin bound and don’t care about performance on that level, it is still a net loss.
So Quark is an interesting fab filler, a 32nm Pentium P54C with some added IP, a 32-bit microcontroller with cutting edge mid-90s tech. Because of the market Quark is going in to that is just fine, more than enough. For process tech 32nm is severe overkill and I question the cost. Intel said that most of their 32nm process tools could be converted for use at 22nm so there isn’t paid for equipment kicking around to build it on. For these reasons, Quark may be an OK fit for the intended market, may possibly be a viable business for Intel, but it makes no financial sense for anyone else to touch. It may fill fabs but what empty 32nm fabs does Intel have? Technically Quark may be an OK $5 32-bit controller with a non-industry standard ISA, financially though it makes no sense.S|A
Latest posts by Charlie Demerjian (see all)
- More on Intel’s 10nm process problems - Sep 17, 2018
- Intel puts out another 14nm 2020 server platform - Sep 11, 2018
- Why Can’t Intel Supply Enough 14nm Xeons? - Sep 10, 2018
- Intel can’t supply 14nm Xeons, HPE directly recommends AMD Epyc - Sep 7, 2018
- AMD reintroduces the Athlon name with two CPUs - Sep 6, 2018