Get email delivery of the Cadence blog featured here
The last day of TechCon had two keynotes rich in deeper technical content, from Greg Yeric on Process Technology Limbo, and Rob Aitken, on How to Build and Connect a Trillion Things.
Greg is focused on what new technologies are going to be important. We had planar scaling, we went to FinFET, what next? For example, last year he talked about contact over active gate and Intel introduced it this year. This is not pitch scaling, but it results in higher density.
He pointed out that there are more important things in the mobile market. The most important phone features (to the consumer) are design build and quality, screen, camera quality, headphone jack, battery life, processor power, price. We are working on the penultimate one in the list, processor power.
Lots are going in on More than Moore. 2.5D is now mainstream, 3D is coming. 3D with wafer bonding too. He had a photomicrograph of a Sony 3-layer chip (the same chip as I wrote about earlier this year in System in Package). This is a 3-layer chip, with 90nm CMOS image sensor, 30nm DRAM, and 40nm logic. Note that those nanometer numbers are not the process technology, they are the thickness of the three die (or, rather, the almost unbelievable thinness). Getting everything so close allows a camera to deliver 103MP with 120 pictures per second, or to deliver movies at 960fps. Using this technology for real 3D, where the logic is partitioned among more than one die, or built monolithically, is going to come, but it breaks all the tools.
The next transistor, after FinFET, looks like being GAA, gate-all-around, but with flattened wires known as nanosheets. If you want to read more about this, including photomicrographs, see my post from last year's Samsung's foundry forum Samsung Foundry Forum: Beyond FinFET and FD-SOI.
Another trick that looks like it may come is folding the N and P transistors over each other, known as Complementary FET or CFET. Like the contact over gate, this is a one-time gain. But at this point in Moore's Law, one-time gains are very important.
Switching to memory, like everyone else, Greg is most scared about DRAM. There is no super memory yet, making the industry hungry for a high-density fast memory. It is important since so much of the power and the clock cycles go to memory.
There is an ancient Chinese proverb, "unless we change direction, we are likely to end up where we are headed." One big challenge is power, where if Moore's Law continues, then electronics will consume half the energy on the planet. See Why Moore's Law is Going to Destroy the Planet.
The biggest thing to change is the materials. DARPA has a program on this, the Electronic Resurgence Initiative. There is a lot of fundamental materials research, especially in 2D materials, like graphene. Greg showed the titles of a lot of papers on 2D materials—those are the titles on the screen behind him in the above picture—and then he pointed out that these were just press releases from a single month earlier this year. Or, in electronics, see the IEEE Spectrum article The Most Complex 2D Microchip Yet from April.
Greg wrapped up by pointing out just how much in limbo we really are. All of the applications, architecture, circuits, design technology and more are in limbo. None of this stuff is yet ready for prime time, just at the press release stage. If we put together what we know, we can maybe get to 2nm—but that is just 6-8 years from now. Whereas, for example, MRAM was invented 30 years ago and has taken until now to get into production. The challenge is how to extend Moore's Law for as long as possible, and then bring some of these new technologies to industrial readiness.
Rob Aitken focused on the idea of a trillion IoT devices. Is it possible? Would it take more silicon that we can manufacture? Does the whole population need to become IC designers? Are there enough batteries?
A trillion is a lot. A newsreader reads 1000 words in a 5-minute newscast. A trillion word newscast would last 6,000 years. There is also the chef's cooking problem: doubling a cookie recipe works, multiplying it by a million does not.
So the first thing to ask is whether there are a trillion things anyway, even before we worry about making them smart. The answer is yes, but you have to include consumables. For example, there are 400 billion beverage cans per year produced.
How about batteries? A trillion batteries (little coin-cells) would need 109,000 metric tons of lithium. Today we produce about 32,000 metric tons of lithium, so 109,000 tons is a lot. That's not even the total since that assumes that nobody else wants any, like Tesla, or for laptops. So it looks like "things" will largely need to do energy harvesting.
Manufacturing? at 2mm2 per chip, there are 35,000 per 300mm wafer, so this is 28M wafers. That's about 3X TSMC's capacity or 30% of current worldwide production. So not completely impossible but still a lot.
These chips need to communicate to the next node along the chain. There are challenges since there is only so much bandwidth. When you get to the level of the cell network, there are also regulatory and license limits. There are also energy limits, and dollar limits, everywhere too.
How about designing them? There are 5,000 RF designers, so that's about 200M per engineer. 20,000 design teams so 50M per tea. But there are 20M people who can write code, so just 50,000 things per coder. So clearly a lot of the diversity has to come from software design. In the hardware space there will be a set of platforms configured by software, or perhaps we manage to make hardware design a lot more like software design (although there are still fixed costs with manufacturing, no matter how cheap design becomes).
What IoT devices need to do is well-understood. Gather the data, make some useful information out of it, and maybe do something. What it takes to do this is really a four-step process.
Rob wrapped up pointing out that these devices need to be invisible to users. They have to "just work" or the trillion scale internet won't happen. Oh, and security, which also needs to just work, invisibly.
Sign up for Sunday Brunch, the weekly Breakfast Bytes email.