WEBVTT 00:00:00.000 --> 00:00:05.360 One of the specs you'll hear thrown around when you're shopping for a CPU is the process node 00:00:05.360 --> 00:00:10.800 measured in nanometers and how a smaller one is better. Just check out the tech headlines and 00:00:10.800 --> 00:00:16.000 you'll see plenty of stories about how chip makers are racing to cram more and more tiny 00:00:16.000 --> 00:00:21.680 transistors onto their processors. And why not? More transistors means better performance and 00:00:21.680 --> 00:00:26.800 efficiency because the electrons don't have to travel as far through each transistor so they 00:00:26.800 --> 00:00:32.480 can switch on and off and therefore process information more quickly. But do process nodes 00:00:32.480 --> 00:00:38.240 really tell the whole story? To get some answers, we reached out to Jason Gorse and Bruce Feinberg 00:00:38.240 --> 00:00:43.040 from Intel and we'd like to thank them for their contributions. The process node was originally a 00:00:43.040 --> 00:00:48.320 measure of how long the gate in the transistor was. This is the part that actually controls the 00:00:48.320 --> 00:00:53.360 flow of electrons from the source to the drain. This was considered an accurate enough proxy 00:00:53.360 --> 00:01:00.480 for transistor size up until about 1997 when the 350 nanometer process was popular. The reason 00:01:00.480 --> 00:01:05.600 this is important is because when you double the number of transistors on a chip, it's fair to 00:01:05.600 --> 00:01:11.680 expect roughly double the performance at a given die size. And for a long time, these 00:01:11.680 --> 00:01:17.600 doublings, if you will, took place at such predictable intervals that Moore's law came to be, 00:01:17.600 --> 00:01:22.160 stating that the number of transistors on a chip would double about every two years. 00:01:22.160 --> 00:01:27.520 This gave the chip makers an easy cadence to follow for naming each process node because they could 00:01:27.520 --> 00:01:34.320 expect each one to be smaller by a factor of about 0.7. Why 0.7, you might ask? Well, the 00:01:34.320 --> 00:01:41.440 transistors are roughly square in shape and if you multiply 0.7 by 0.7, you get 0.49 or roughly 00:01:41.440 --> 00:01:47.920 one half. So for example, when the industry went from the 1000 nanometer process node to the 700 00:01:47.920 --> 00:01:52.880 nanometer process node, this marked a rough doubling of the number of transistors that they 00:01:52.880 --> 00:01:58.800 could fit in a given area, even though the name of the process only reduced by a factor of 0.7. 00:01:58.800 --> 00:02:05.680 Thing is, in 1997, while manufacturers were able to start shrinking the gate length by more than a 00:02:05.680 --> 00:02:11.920 factor of 0.7, other parts of the transistor weren't shrinking as quickly anymore. So gate 00:02:11.920 --> 00:02:18.080 length was no longer a good proxy for the overall transistor density in the entire chip, 00:02:18.080 --> 00:02:22.720 and therefore the performance. Rather than changing the naming scheme outright though, 00:02:22.720 --> 00:02:28.560 we started to see a process node defined by the size of a group of transistors called a cell. 00:02:29.200 --> 00:02:33.600 This was done to give people an estimate of the equivalent level of processing power 00:02:33.600 --> 00:02:39.040 accounting for components that weren't shrinking as quickly. So the first node we saw under this 00:02:39.040 --> 00:02:46.080 new naming system was the 250 nanometer process. Performance was about double the previous node, 00:02:46.080 --> 00:02:51.280 as you would expect from the name, but the gate length was actually around 190 nanometers, 00:02:51.280 --> 00:02:56.240 which is much smaller. It's just that there was other stuff that prevented the transistors from 00:02:56.240 --> 00:03:02.000 being packed more tightly than that. This scheme involving cell area lasted until around 2012 00:03:02.000 --> 00:03:08.000 and the 22 nanometer process when a whole new type of transistor was introduced, FinFET. 00:03:08.560 --> 00:03:14.320 Chipmakers found that at these sizes, the gates were so small that you could have electrons leaking 00:03:14.320 --> 00:03:19.680 through them due to quantum tunneling. This could cause undesirable behavior, so engineers needed 00:03:19.680 --> 00:03:25.040 a way to make their chips more powerful without shrinking the gates even further. The solution 00:03:25.040 --> 00:03:31.040 was to take the channel the electrons go through and raise it up like a shark fin, hence the name 00:03:31.920 --> 00:03:36.640 increasing the surface area of the channel and allowing lots more electrons to pass through. 00:03:37.200 --> 00:03:42.000 Of course, this also meant that transistors were now three-dimensional instead of planar, 00:03:42.000 --> 00:03:47.520 making it much harder to accurately measure their size. Now the industry has still continued to use 00:03:47.520 --> 00:03:54.720 that 0.7 factor to describe a generation of improvement, like going from 14 to 10 to 7 00:03:54.720 --> 00:04:00.240 nanometer processes. But the truth of the matter is that these numbers don't actually measure the 00:04:00.240 --> 00:04:06.240 real size of the transistor anymore, and they can even vary wildly between different manufacturers. 00:04:06.240 --> 00:04:10.880 Intel, for example, attempts to measure a process node by taking the weighted average 00:04:10.880 --> 00:04:16.560 of the two most common standard cell sizes. Really, a more important consideration though 00:04:16.560 --> 00:04:22.480 is transistor density. That's how many can be packed into the same space without decreasing 00:04:22.480 --> 00:04:27.840 the size of the actual transistor features very much, if at all. In addition to density, 00:04:27.840 --> 00:04:32.000 chip makers are using other techniques like improved materials to boost performance. 00:04:32.640 --> 00:04:36.720 This can include everything from squeezing the crystal structure of the channel to make the 00:04:36.720 --> 00:04:42.240 electrons go through faster, to low resistance traces between transistors, to gate materials 00:04:42.240 --> 00:04:48.160 with a high dielectric constant for better control of electron flow. Of course, this process can 00:04:48.160 --> 00:04:53.440 require some trial and error. Intel's well-publicized difficulties with their 10 nanometer process 00:04:53.440 --> 00:04:59.680 were due in large part to them trying to overscale. In other words, pack more than double the number 00:04:59.680 --> 00:05:04.800 of transistors into the same space, which required them to try lots of new technologies inside the 00:05:04.800 --> 00:05:10.720 chip all at one time, which caused delays and manufacturing problems. But as our technology 00:05:10.720 --> 00:05:16.320 continues to improve, chip makers look poised to keep Moore's law, even if it's a little slower, 00:05:16.320 --> 00:05:21.760 alive to some extent, as well as keep silicon, the base material for our processors, for a long 00:05:21.760 --> 00:05:26.000 time to come before we have to really start considering more exotic solutions, like carbon 00:05:26.000 --> 00:05:31.680 nanotubes. In the meantime, I hope you enjoyed this deeper dive into processor sizes. Just remember 00:05:31.680 --> 00:05:37.200 that the process node isn't the be-all and end-all when you're shopping for a CPU anyway. It's 00:05:37.200 --> 00:05:42.560 always more important to pay attention to the real-world performance that you'll see in games 00:05:42.560 --> 00:05:48.400 and applications that you actually use. Thanks for watching, guys. Like, dislike, check out our 00:05:48.400 --> 00:05:52.240 other videos, leave a comment with a video suggestion if there's something you want to see, 00:05:52.240 --> 00:05:57.440 and don't forget to subscribe and follow. You could even ring the bell if you really like us. 00:05:58.480 --> 00:06:01.760 Or if you don't like us, you just want to, like, hate watch us. That's fine too.