WEBVTT 00:00:00.000 --> 00:00:05.180 Cores, cores, cores. Whether you're an Intel fan or an AMD diehard, 00:00:05.180 --> 00:00:08.280 both of the big PC chip makers are trying to sell you 00:00:08.280 --> 00:00:11.360 on the idea that you need as many cores as possible. 00:00:11.360 --> 00:00:14.920 But wait a second, why can't we just have CPUs 00:00:14.920 --> 00:00:18.480 that have one lone super powerful core? 00:00:18.480 --> 00:00:22.040 Wouldn't that be a lot more straightforward? I mean, think about it. 00:00:22.040 --> 00:00:27.880 Those old school parallel ports that we used to use for printers carried lots of data streams at once, 00:00:27.960 --> 00:00:34.200 but they were actually very slow. Meanwhile, USB ports only handle one data stream 00:00:34.200 --> 00:00:39.600 at a time, but they're very fast. Couldn't we just apply that same idea to CPUs? 00:00:39.600 --> 00:00:44.560 To answer, we spoke with Eric Jihon, Intel's CPU SOC architecture lead, 00:00:44.560 --> 00:00:49.720 and we'd like to thank him along with Thomas Hannaford and Ben Benson for helping us out with this episode. 00:00:49.720 --> 00:00:53.320 As it turns out, a single core is certainly capable 00:00:53.320 --> 00:00:57.320 of juggling multiple programs at one time. This is done by scheduling, 00:00:57.320 --> 00:01:01.760 where each clock cycle of the CPU, and there are billions of those per second, 00:01:01.760 --> 00:01:06.840 is assigned to execute instructions from a certain program or multiple programs. 00:01:06.840 --> 00:01:12.760 Scheduling can be done by the operating system, or more optimally, it can be handled by the CPU's hardware. 00:01:12.760 --> 00:01:16.120 For example, simultaneous multi-threading or SMT 00:01:16.120 --> 00:01:21.000 has been with us since the early 2000s and actually makes your operating system see 00:01:21.000 --> 00:01:24.760 one physical core as two logical cores. 00:01:24.800 --> 00:01:28.160 With SMT then, the operating system will always make sure 00:01:28.160 --> 00:01:31.440 that each physical core has two instruction threads 00:01:31.440 --> 00:01:36.960 to work on so that if a single thread stalls, because the CPU has to wait for additional data, 00:01:36.960 --> 00:01:40.600 the CPU core can still work on the other thread. 00:01:40.600 --> 00:01:45.520 But even with SMT, scheduling still has inefficiencies. 00:01:45.520 --> 00:01:49.440 Each core can still handle only so many instructions per clock cycle, 00:01:49.440 --> 00:01:53.240 and if you're trying to run a ton of programs or processes, 00:01:53.240 --> 00:02:00.080 the whole scheduling paradigm can become overwhelmed, meaning that adding a second or third or fourth core 00:02:00.080 --> 00:02:05.760 can help remove the reliance on the scheduler and allow the CPU to actually do more work. 00:02:05.760 --> 00:02:09.280 But the fact that scheduling is an imperfect solution 00:02:09.280 --> 00:02:12.720 isn't the only reason that we can't just have one big core. 00:02:12.720 --> 00:02:16.240 You also have to think about how much CPUs cost. 00:02:16.240 --> 00:02:19.240 Higher end chips are already expensive enough, 00:02:19.240 --> 00:02:23.360 but if you tried to consolidate a modern, say, eight core chip 00:02:23.360 --> 00:02:27.640 into one big core, the price tag would be even higher. 00:02:27.640 --> 00:02:34.000 You see, in modern CPU fabrication, if you have one core that's bad due to a manufacturing defect, 00:02:34.000 --> 00:02:40.080 you can just disable it at the factory and then sell that chip as a model with fewer cores. 00:02:40.080 --> 00:02:43.440 But if you had one massive core and it was defective, 00:02:43.440 --> 00:02:49.600 you'd have to throw the whole thing away, meaning more waste and higher costs for the consumer. 00:02:49.600 --> 00:02:52.960 There's also the consideration of how you link up a CPU 00:02:52.960 --> 00:02:56.640 to the rest of your PC. A multi-core CPU is set up 00:02:56.640 --> 00:03:00.040 so that each core has its own connection or fabric. 00:03:00.040 --> 00:03:04.680 That gives it access to memory. As you make these fabrics wider and faster, 00:03:04.680 --> 00:03:09.240 it becomes more and more difficult to make a fabric that has the same signal integrity 00:03:09.240 --> 00:03:12.280 as the smaller ones that were designed for multi-core chips. 00:03:12.280 --> 00:03:16.120 But beyond all of these concerns regarding the CPU itself, 00:03:16.120 --> 00:03:20.440 the way we use computers these days really tilts the equation in favor 00:03:20.440 --> 00:03:26.720 of having multi-core processors. Although back in the days of Windows 95 and even XP, 00:03:26.720 --> 00:03:31.800 you could get away with a single core CPU. Modern versions of Windows are often running 00:03:31.800 --> 00:03:35.920 hundreds of processes and thousands of threads at once. 00:03:35.920 --> 00:03:39.080 And because PCs are such general-purpose machines 00:03:39.080 --> 00:03:44.560 that need to be able to adapt to a wide variety of workloads, it's better to have multiple cores 00:03:44.560 --> 00:03:48.880 rather than just hoping that your applications will be optimized enough to take advantage 00:03:48.880 --> 00:03:53.240 of one big single core, especially as it's becoming more and more common 00:03:53.240 --> 00:03:58.720 for programs and even games to actually be able to take advantage of more than one core. 00:03:58.720 --> 00:04:01.960 But there's one alternative that we didn't really talk about. 00:04:01.960 --> 00:04:05.240 If you want one huge core, you could just forget about CPUs 00:04:05.240 --> 00:04:09.680 and go cut open an avocado. That would solve your problem. 00:04:09.680 --> 00:04:14.480 And you wouldn't be hungry anymore. Thanks for watching. Leave a like or a dislike depending how you felt. 00:04:14.480 --> 00:04:19.080 Check out our other videos and leave a comment if you have a suggestion for a future fast as possible.