Episode 1045: Nvidia GPUs Destroy CPUs in the AI Race, Math Error Exposing the Tech Bubble

Martin Shkreli (0:00)

Do I use Codex or Quad? Yes. I saw both IonQ press releases, they're both fake, and got a very small amount of Jagex. We use Cypher, it's fake and full of gate. I mean, I'm an expert in everything faking, so that helps a lot, but IonQ is just kind of a fake company, like, and they kind of are, you know, they're sort of in semis, sort of. They're making, you know, so-called quantum computer. You have to learn a lot about computing, you have to learn about what algorithms are all about. And it's really not that hard, you know, I think it scares people, because, like, the book on algorithms, the main book on algorithms is called CLRS.

And it's like, you know, we have a copy over there, actually. And it's like a thousand pages, and it's really dense, and it's not fun. They try to make it easy, but it's not easy. But the reality is, like, the basic idea is, if you had 52, you know, cards, ordered cards, and you shuffled them, how fast can you reorder them? And what algorithms or what strategy is the fastest? And you might say, well, I'm gonna have the deck, and I want to take one card. And if it's higher than the middle, I'm going to send it to the back of the deck.

And if it's not, I'm going to move it to slide at one position. So you can only move the cards one position at a time. That's a very slow strategy. There's other types of strategies where you go from one slot to the next. So if the first slot isn't the first card, which maybe is called Ace of Spades, then you move that card to the right, and you move the card to the left. That replaced it. And then you're going to go to number two and you do the same thing. Eventually, you will sort all the cards this way. That's also pretty slow. And so the running time of these algorithms is called complexity.

And some computers, most computers are the same. There's no special thing about any one computer versus the other. They all perform at the same algorithmic speed or complexity. The difference is that it used to be the clock speed, where you'd have a gigahertz, then a two gigahertz, then a three gigahertz. And eventually, so hertz is a cycle per second. So when I was a kid, 100 megahertz was fast. It's 100 million cycles a second. We'll get into what a cycle is in a second. But the more cycles per second, the better.

And eventually, the laws of physics kind of stop you from going much faster than five gigahertz. It's pretty hard. Otherwise, you're going to run into thermal issues and other just things that just make it impossible. Otherwise, we'd love to go a terahertz or faster, but you can't. So what these companies decided to do is to make more cores. So we went from all computers before like 2000 had one core, and then they decided to go to two cores, and four, and six, and eight, and 12 Yeah, 16 is pretty typical. Now, a GPU has 60,000 cores.

That's why it's so fast. It's not because it has a high clock speed. In fact, the clock speed is lower.

It's around 2 gigahertz. But between 2 or 3 gigahertz doesn't really matter when you have 100 times the cores or 1,000 times the cores. So that's what makes GPUs so much faster. But you have to be able to split the work into the different cores. There's some work you can't split, and there's some work that's very good to split. And that's called parallelizable. If you can parallelize the work, GPUs are really good for that kind of work. A lot of work is parallelizable. So that's the good news. But in any event, a quantum computer sounds pretty fancy. But at the end of the day, it has to win on clock speed, right? Because that's how you beat it. Does it have a clock speed that's faster than a CPU or GPU? No. In fact, they're tremendously slower.

Okay, does it have more cores? No. There's a single core, as a matter of fact. So how is this thing exciting? Why would anyone give up on a quantum computer? And the answer comes in the complexity, the running time of the algorithms. Now, most algorithms that we're using today revolve around matrices. But there's one quantum algorithm that's much, much faster on a quantum computer, and it involves factoring integers. And unfortunately, factoring integers isn't really a thing people do often. So if I tell you 35, you know it's 7 and 5 But if I tell you a really long number, it's actually very hard to figure it out, and that's what factoring integers is all about. But a quantum computer can factor integers almost instantly, whereas any other computer would take a certain amount of time, often the square root of the size of the integer, whereas quantum can do it instantly. As you get a big enough integer, like the integers that are on blockchain, there's no computer fast enough to factor it. You would be searching for factors forever, even if you had every computer in the world. But with a quantum computer, it could find the answer right away, which is pretty amazing. That's the only thing you could do faster. Everything else, like sorting a deck of cards or multiplying two matrices or anything like that is much, much, much, much slower on a quantum computer. So you have to find the algorithm that makes sense for a quantum computer. And unfortunately, there's only really one that's good. And, you know, it's kind of a trick that nobody needs. We get it. So it's a learning about computing is tricky because you end up becoming a professor instead of a trader.