CPU wars: Intel vs. AMD vs. ARM

Hi there, everyone, and welcome to a new episode of Syscast, the show where we talk Linux, open source, and pretty much anything else that interests me. It’s been a while, so it’s good to be back, sort of. And what better way to kick off a new season than having a chat with one of my best friends, Jan Somers.

How are you, buddy?

Hi, I’m doing great. It’s nice to do something else. It’s supposed to be something that could be a new hobby.

I don’t know. Let’s see how this goes.

Well, the reason I wanted to get you on is I have been struggling a bit lately with everything that is hardware related. I fell into a rabbit hole trying to come up with a proper set of servers that I could hire or purchase for Odeer, the monitoring service. I was looking for servers that had a lot of CPUs.

I didn’t need the memory or the disk space, just a lot of CPUs and CPU cores. And then you fall into the rabbit hole of should you go Intel or AMD? And hey, suddenly ARM or ARM has become interesting as well.

And that’s a deep rabbit hole because there are so many acronyms and technologies. Well, shit just got real, as they say. And I totally got lost.

So what better thing to do than ask you about all of the things that I do not understand. So I hope you’re ready because I have a lot of questions.

Okay. See, one of the first things that you notice is Intel sort of dominates the server market. Every expensive server that you can buy nowadays probably has an Intel CPU in it.

Why is that? Why is Intel instead of AMD or instead of ARM, why are they suddenly the de facto default that we should look into? Yeah.

Well, Intel is actually the founding father of the instruction set that’s used in the modern computer that’s running your Windows and currently your Apple laptop and stuff like that. But due to some weird plot twists in history, they needed a second source for supplying their chips to IBM in the days when the computers were… still becoming popular and rising to popularity. So they actually licensed to other parties like AMD, Cyrex and other companies that unfortunately cease to exist.

Only AMD is currently still in the market. for some high-powered chips. But that put Intel at an advantage because they only licensed the tech, but not the prints of the entire chips themselves. So it’s always been a neck-and-neck race between Intel and AMD.

But they, well, Intel, when I refer today, have a very good marketing department. So if you can market your chips very good and they’re high-performant and they outperform your competitor, Well, that gets you a market advantage. Look at Microsoft.

They just wiggled their way around the market as well. And that’s pretty much the story of war between AMD and Intel the last couple of decades. There have been a few twists where AMD overtook Intel for a couple of years, but then Intel beat back.

And that’s about the time period we’re in right now where Intel was the king and dominant and the one and only king on the hill in size of performance in the data center, laptop, desktop. They were everywhere. But then AMD researched from six, seven years of being mediocre or just so-so at best at desktop and laptop CPUs.

And then they fought back with a completely new architecture named Zen. And they have been at Intel’s throat as of now. And currently they have overtaken them and by quite the margin.

And as of CES, it’s the Consumer Electronics Show, which just happened a few days ago in Las Vegas. They have launched new laptop chips and they have now beaten Intel on each and every front. The data center, more cores, less price, better performance.

Desktop, same thing, more cores, more performance. And laptops as well, because they’re just using the new design, a new smaller process. And that’s pretty exciting to follow.

It is indeed. So if I can just go back a step, you mentioned that Intel sort of created AMD by accident then because they just couldn’t keep up with production? Or how does that happen?

How do you create your own competitor?

Well, they were forced to create their own competitor or their worst enemy, to say so. Because when they wanted to sell their chips to vendors like IBM for the upcoming computers, they were not satisfied that Intel could suffice in the production of enough chips to supply their demand. Because back then, manufacturing chips was a bit flaky.

The yields weren’t all so good. Because every… every production plant was slightly different. They weren’t a carbon copy and everything was a bit off on some production lines, which yielded in bad or no chips.

So IBM, for example, they demanded to have a second source, another company who could provide them with exactly the same chips. Just so if one of the fabs or factories of Intel blew up or didn’t produce any chips, that they were guaranteed that they could deliver their personal computers to customers. So they were forced to do that.

So they went into… With a bit of history, they ended up licensing their technology to… AMD, which gave them the keys to the kingdom and the ability to create, well, x86 compatible CPUs.

So you can just install your software on either Intel or AMD platform. So that’s how that happened.

Okay, I’m already hearing a couple of acronyms that we really need to dive into soon. You mentioned that Intel, so AMD has some kind of license from Intel that allows them to produce a certain chipset. Does that mean that if AMD would suddenly become the new default on servers, basically Intel also benefits because of all the licensing involved?

And so they can’t really lose this game?

Well, that’s quite how you can put it. It’s, as they say, a gentleman’s agreement that they lend their new technologies and research and additions to the x86 instruction set that they license it to each other without any royalties. A really good example of that would be one of the first times that AMD overtook Intel.

Don’t ask me what year it is, but it’s with the Athlon 64 CPUs, the first 64-bit CPUs. Intel was working on a similar project with their Itanium line, which ended up to be bust because they were too complex and they were really marketed as… servers and high available compute. But that actually flopped because of the high cost and the complexity.

But AMD actually extended the current x86 instruction set and added 64-bit instructions and capability to it without sacrificing backwards compatibility with 32-bit because a lot of software was back then… or exclusively 32-bit because it’s a chicken-egg thing. So because of that, it’s called the AMD64 instruction set, which is appended to the x86. And because of the gentleman’s agreement, Intel could also look into the specs and they share those IP.

So they can’t really end up losing the game unless they just go bankrupt. So technology-wise, they will always innovate and try to outsmart each other in other ways. But in pure, the instruction sets, they have to be compatible.

Otherwise, you’ll just get a divine in the market.

Smart folks. So they basically allowed AMD to exist. AMD then got better and improved chipset and they gave it back to Intel.

And that’s the quid pro quo kind of deal where they give some, they take some and everybody’s happy.

Yeah, pretty much it. If you use some software like CPU ID or other… software like that, you can actually check out the instruction sets that are supported by your CPU, like SSE, SSE 1, 2, 3, 4, etc. Those are instruction sets created by Intel, but you can also find them in various AMD chips from the same era and up until now.

So if they create one extra instruction set to do hardware acceleration on XYZ, like video encoding and compression and other crazy math things that I don’t know half a thing about. The other party will have it as well. It’s just bilateral.

Okay, cool. So you mentioned one of the first acronyms that… I’ve seen before because it’s all over the Linux packages every time you install it.

You mentioned x86. What is that and what are the alternatives?

x86, it’s just something that Intel came up. It’s just instructions to fetch from memory, decode the instruction, execute it and write it back to memory or a way to add up certain… It’s a bit like a framework like in PHP or something like that.

It’s not really defined as much what it is. It’s just how it’s supposed to be built and what’s the idea behind it.

Okay. So at the very beginning, you mentioned that AMD at Consumer Electronics Show, they announced new CPUs for desktop laptop servers. You mentioned that they became smaller.

I think you’re referring to the nanometer process at that point? Yes.

Yeah, the fabrication process, the node, the amount of nanometers, that’s all the same. The current notification of nanometers is more of a marketing term. I’ll get into that somewhat later.

But it defines how small each of the features of each individual transistor are on the chip itself. So the smaller every part of the transistor can be, the more you can cram into the same or smaller size for the same power budget. Or if you have the same footprint of your chip in total, you can cram more transistors in there.

So more logic, more… or place for more instructions or more complex instructions, more cache. You can do a lot with that. So it can go both ways.

Either you go with a more energy efficient process, so your transistors are optimized for energy efficiency or for high performance compute. So it depends on the structure size and how you… manufacture them that it either goes to a efficient process or a high power feature size

As a rule of thumb, I would say yes. If it’s a smaller feature size chip, so if the… so-called nanometers are smaller, or as they advertise, it’s baked on a smaller process, then it would generally have more transistors and more logic in it. And it would perform better than the other part because it just has more hardware to do the same thing or do it better than that.

It’s not always the case, but in most general cases, yes.

Yeah, that’s exactly the thing. Okay. Like I mentioned, AMD has overtook Intel for a few times.

And the exciting thing for us hardware enthusiasts actually happened with the Dozen architecture of AMD. And they’re currently on their second revision of that architecture. And later this year, they’re going to do Zen 3, which is the third revision of that, which is going to be a big improvement on its own.

The great thing about that is that AMD got an opening to be this dominant in the market because of Intel’s somewhat larger ambitions than was feasible at that time. The feature size or the process that they bake their chips on, they were previously on the 14 nanometer process. But Intel wanted to go from 14 to 10 nanometers, which doesn’t seem like a very big jump.

Usually they do a jump from 14 to 12 or 11 or something like that. But going to 10 is very ambitious. And they really had a lot of trouble getting it up to speed, running, having greater yields or the same yields as their current mature process, which resulted in delays after delays after delays.

A lot of… Apple fans, I think, especially for the laptops, have been bashing Apple for the fact that their laptops haven’t gotten faster and faster. That’s mostly due to Intel not being able to keep up with their roadmap because as chip designs go and laptop designs go and roadmaps for any big enterprise company…

They all plan like five or more years ahead. And if those roadmaps and plannings don’t go ahead because one company or one chain in the link doesn’t come up to its part, then you have big troubles. So Intel is really in a pickle because they’re scrambling to bring out more products with more cores since the competition AMD is bringing out so many chips with so much more cores for a very small price compared to what Intel is asking.

And it’s really putting them in a bind.

So because Intel tried to skip a generation, I guess, by going 14 to 10, that failed and that gave AMD the breathing room to take it up a notch and then try to dominate that market. So why? Okay, I’m now going to show you how very little I know of hardware.

Intel is, from what I can tell, it has really good CPUs. That’s probably why all servers have them. If AMD now is suddenly a lot better and they apparently use the same chipset, can I just yank out my Intel CPUs and put an AMD CPU in the same motherboard?

Or is there more to it than that?

If you were in the 90s, you might have could have done that because they shared the same socket. So the actual physical connection between the pins on your chip and the main board. But since I know…

95, 96, they actually began to make their own designs, their own sockets, so their own shapes and pin layouts. And they drifted apart from that. So they have their own way of configuring their chips and doing their pin layouts.

And they have their own chipsets and their own motherboard. So unfortunately for the consumer, you can’t just drop in and replace an Intel chip with an AMD chip. You’ll unfortunately have to change your main board as well.

So if you want to change, that’s going to be pricey because these main boards cost about 100 to 300 or 400 euros depending on the amount of bling and RGB lights you want on these days.

So I was thinking about the potential future here. If AMD really is doing this great and Intel is a bit lacking in that case, it isn’t that easy to completely switch either a server farm or the… I’m imagining if you’re a cloud provider and you’re selling hardware, it might not be that easy to suddenly start selling the AMD CPUs because it requires, as you say, a new motherboard, which might have implications for the memory slots that you have, the disks you can attach, the rate controllers you can attach to that motherboard.

So the entire chain gets thrown upside down if you just want to swap out the CPU.

Basically, that’s it, yeah. But AMD has been having a few generations of the new server chips with Zen architecture already out. So they were trying to warm up the market for their re-entry to the server market because Intel was absolutely dominant there.

But yeah, unfortunately, for a lot of cloud vendors and IT integrators, which use Intel in their data centers currently, they’ll have to go through a lot of testing, QA, validations in order to see, is this new platform up to spec? Are the drivers for all our virtualization stacks working? Are they okay?

Are they stable? Because every new hardware platform has their own kinks. Every chipset has their own disadvantages.

And you always have to have software to drive the link between your software and the hardware between it. But everything’s been looking up for them. It’s pretty promising.

But I think a lot of companies have been testing the waters with the previous generation and seeing what it does, but they’re very enthusiastic about the new generation, the current Rome Epic server chips. Intel has their Xeons, and AMD has their Epics. It’s just like the ring of it more, but hey, that’s me.

But they really offer… a lot of bang for their buck because the epic cpus they have like 64 physical cores 128 treads and a buttload of cash in order to support these chips with enough data to keep them busy pretty much 100 percent of the time which makes it possible for people who are now using two Xeon CPUs in one hypervisor, for example, to get enough cores for a maximum distribution between memory, CPU, cost, etc., that they can now do it with a single socket EPYC CPU, which drives the cost down considerably, which really opened up the market to whole new ideas. So you can either go way beyond what Intel now has with two socket CPUs, or you can have the same thing with one CPU on the EPYC side, or you can have two EPYC 64 core CPUs, which is a buttload of threads, cache, et cetera. Depending on what you want, build a supercomputer, go to Mars.

You can go to town in the AMD shopping department if you want.

Yeah. Intel actually marketed what SMT is, simultaneous multi-thread, I presume. That’s the acronym for it.

But the thing isn’t new. Intel just marketed it as hyper-threading. It’s just a fancy name for SMT, which is an industry-wide thing.

AMD just calls it SMT as well. SMT is just the hyper-threading non-branded version, which AMD uses as well. So if you say multi-threads or X amount of threads, that’s the hyper-threading variety of what Intel would call it.

Okay, now we’re talking about cores. So a CPU has X amounts of cores. Usually the rule of thumb is if you have more cores, the clock speed of each core goes down.

I don’t know why that is, but that seems to be the rule of thumb. But I somehow, I can remember if I go back a decade or so in the hardware realm, there have been Intel CPUs that were clocked at 2.4 gigahertz, which were amazingly fast. And then you had AMD CPUs, which were clocked at 4 gigahertz, so almost twice as fast. which seem to be just as fast.

What’s the deal here with core counts versus clock speeds? Does it even matter nowadays?

It’s always a delicate balance between a lot of things because one thing you didn’t mention, which is very important in scoping this out, is the pipeline that they’re using. The pipeline is the different stages that your instruction set can go through. And the difference between the 2.4 Intel CPU and the 4GHz AMD CPU at that time was actually the pipeline design which AMD used.

Intel went for a very elaborate and long, extensive pipeline, which they still do. use and improved upon. And AMD opted into the, let’s go for a lot of cores and a lot of megahertz or gigahertz, but with a very short and simple pipeline. That ended up hurting their performance because the market was going another way and it just didn’t work out.

On paper, it was a very good design, but the shorter pipeline, which was more simplistic, just cost more steps to do the same instructions. at a faster pace. So that’s how they could keep up somewhat, but they fell behind very quickly, comparing to a slower Intel part, which had a more elaborate and complex pipeline. So that’s a bit of a way to draw comparisons between those two and how those are all interlinked.

Because clock speed isn’t everything. It’s also your design. It’s your branch predictor to see how and when and what instruction is going to go where and how to optimize using all the parts of your CPU to keep them busy at the max.

How big are your caches, the level one, two and three caches, et cetera. It’s a very complex thing to balance between all these things in order to make a great design and also innovating upon new technologies and not just refining what you have.

Well, it was a bit of a gamble because AMD also came back from using the shorter pipelines and went to the more complex and elaborate pipelining system with… lots of branch predicting and improving their cash hierarchy, et cetera. So they gambled, but they gambled wrong. Like I said before, those chip designs, they are designed and thought of four or five years in advance.

And, well, you have to guess what the future is going to. So what are people going to need? What is… the data center going to do?

How is virtualization going to end up? What are the needs? Is everything going to VDI?

Is everything going to microservices, which they thought was going to be the thing? So small containers with small threads, with small executions, which had to be fast and snappy, or is it one big monolithic VM or a computer? it’s very hard to look into the future and see, aha, this is what we need. They just gambled and gambled wrong.

And now they’re just doing different designs, which took them a lot of time in order to draw up, execute, test, and bring out. And that’s why the Intel has been king of the hill for so long and had a sort of monopoly asking monstrous prices for CPUs. But they’re coming back from that because AMD is hurting them a lot on the price front and performance front.

So they have to cut back on prices, which is good for the consumer. That’s true.

So the pipelines that you mentioned, I’m trying to… understand i think what a pipeline means in this case so if there’s an instruction an instruction can be i think add one plus one there are probably a couple of ways to do that you can have a dedicated instruction set to do a addition of two numbers or you can go a very weird way where first you subtract some numbers and then you multiply it to get the same result. So is the pipeline something that I can look at as if there’s a dedicated instruction to do some kind of operation, then it’s okay to have a short pipeline because you only have one instruction. But if it’s a complex operation that you want to do in your CPU… the longer pipeline wins because, see, I’m totally lost.

What the hell is a pipeline in this case? And how do those instructions come into it?

A pipeline, as far as I can explain it, I hope I don’t make too many mistakes, but the general gist of it is that your CPU has a lot of different specialized calculation parts or cores or pieces of hardware. And a pipeline, as far as I understand it, is that you can string a lot of these parts together in a consecutive order in order to execute one instruction. So if you have two or three simple units stringed together, which are multipurpose, and you execute a thing, a complex instruction might pass one or two or three times through that pipeline to refine or execute the entire command.

But if you have a very complex pipeline and you have a lot of small, very highly purposed instruction cores, you can actually end up running the commands or the instruction in the pipeline once and be done with it. So you can end up with a lot of efficiency gains doing it in a long pipeline, but it also draws more power. It uses a lot more… silicon and transistor space on your chip so it’s a delicate balance on how large are you going to do your pipeline and that’s why just to jump onto another thing ASICs are so efficient at what they do because they have a pipeline only and only for that what they have to do mine bitcoins or calculate something or encode video if they’re in a video recorder etc so That’s basically what a pipeline does.

It’s stringing together a lot of specialized parts in order to try to execute things all at once, only once. And if it’s not possible, they might have to go through the pipeline a few times.

Yeah, that’s basically it. The only thing is that the instructions are exactly the same. It’s just how they execute it behind the curtain because they speak the same language, but their household or the way they do it internally just differs slightly.

And that indeed could cost them a few clock ticks in order to do the same thing, which is what they call IPC, so instructions per clock. which is now the current measurement for efficiency and performance per watt power they consume. So that’s exactly that.

Okay, so in dev terms, you’re talking to the same API, basically your instruction sets, but that API does the processing in an entirely different way, making one CPU sometimes more efficient than the other. Okay, that makes sense. Now, we’ve been talking about Intel and AMD.

Why is there no third CPU provider with a different set of acronyms that is x86 compatible? Why do we only seem to have two major CPU providers in the world? Why is there no third or fourth or whatever?

In the past, there were some other providers which unfortunately went bankrupt or diverted their resources into other markets. I mentioned Cyrex before. Those were very popular CPUs back in the day of the Pentium, etc.

They were drop-in replacements for… and were considerably cheaper, but weren’t always clock for clock more efficient or more powerful, but they were a lot cheaper. So for the hobbyists and the builders, that was okay. But as I mentioned, most of these vendors or brands went out of business due to fierce competition between the… superiorly Intel and the vastly bigger competitor AMD because AMD wasn’t that small at that time or they diverted their resources into other niche businesses so that’s why we don’t have currently any other competitors because Intel does not need to do the second source agreement with other partners anymore because there are two parties on the market and they’re not going to repeat the same They’re going to repeat history again by licensing x86 to other parties.

So that’s pretty much it. Because otherwise, if they have another party on the market that can compete with them, their market share would decline because they have three parties competing or four parties competing. So they’re pretty comfy at where they are.

So they don’t have any need of licensing x86.

No, I don’t see it happening anytime soon or ever for that matter because there’s some other fronts like RISC-V and ARM, if you want to name it, that are also playing on the same field or trying to get in the same field and vice versa with Intel and AMD trying to get in the mobile and the low power market. So it’s… It’s a mix of different things, but as far as the current consumer market go for laptops and for desktops, since AMD is back at the game and have some really good products as well as Intel, they’re still good.

They’re just behind in the game. Don’t get me wrong. Prices will go into a more favorable way because Intel had a monopoly for so long because AMD had no counter answer to that.

So as long as there are two parties which compete head-to-head, that could be quite interesting on technical level and for pricing.

Now, you already mentioned ARM. So let’s have a look at that if you can. We have Intel versus AMD, which are both running the x86 instruction set.

Now, the reason I got into this entire rabbit hole was because I wanted to find a server with a lot of cores. And once you start looking for a lot of cores, you also find ARM machines or ARM servers. What’s the difference between ARM and the Intel and AMDs of the world?

ARM is a bit of an oddball in this game because ARM is actually just a company licensing their intellectual property for their instruction set. And they just license it out to anyone who wants to buy it. So they develop the designs as well for their cores and other integration options.

And they also… do consultancy to help you design your own chips. And if you are a big enough company, you can hire your own set of designers and create your own mobile SOCs in order to power your smartphone. So it’s a very versatile thing because you have a blueprint and can improve upon it.

You just have to pay the license fee. So they don’t manufacture their own physical chips. They just design pipelines, instructions, et cetera. but they are of a completely different order since there are, well, more than that, but currently x86 versus ARM, that is just to say, it’s just a difference between the way their instruction sets are built out of.

So going back to x86, that’s what they call a CISC chip. It’s complex instruction set computing. And ARM is from the beginning built upon the RISC instruction set, which is a reduced instruction set computing.

The difference between the two, viewing it from the point of arm, is that RISC is built in order to have a very few or as few as possible and as simple as possible instructions. And they also have the fixed length. So each instruction is always X amount of bytes or bits long.

This is totally different from what x86 does because their instruction sets vary in length, which makes it very powerful because they can just make up whatever they want a new instruction set. They don’t have to look at the size of the instruction set in bits because they can just throw it in there. or in their branch predictor and just divide it into micro operations, which end up to be RISC instructions. So they convert CISC instructions to RISC instructions and then push it through the pipeline on their end.

But the major advantage of what ARM has started is because of their fixed instruction size. they can, in theory, have a lot easier to develop and less complex chip designs. Because you don’t have to incorporate very complex designs in order to… read, decode, and distribute these varying lengths of instructions, which makes it a lot easier to make these chips. So they can implement more features like power efficiency modes.

And in general, if your pipeline is much smaller, more simple uh i think on a silicon level you can also distribute a lot more over over your your chip size in order to distribute heat dissipation power etc um it has a lot of advantages so am i right to assume then that arm has a potential good future for general purpose devices so i’m thinking

x86 has just evolved into whatever the market needs. It will always outshine in multi-purposeness because they can do pretty much anything you can throw at it. Where other chips like ARM, generally are only good at what they are built for.

They’re going more general purpose and they’re growing towards each other, what both can do. But it’s pretty much a thing of the right tool for the right job because a lot of smartphones have arm socks in there because they are energy efficient. They’re designed only what the phone needs to do and you can slap a lot of hardware onto it like graphics cores in order to play your games like Angry Birds and whatnot.

So it depends on what you need. So as far as I can see, there’s still a divide in the markets where x86 offers a lot of flexibility and raw performance for people who are content creators and high performance compute. Because you could do pretty much anything with it.

And ARM is very specialistic, but it’s also growing to the more general compute at the expense of power and complexity, of course.

They can always add instruction sets to the ARM license pool or the capabilities of ARM. But as far as I can see, I think they’re going to earn and have earned their place in the data center due to their power efficiency because Amazon is also making their chips. They have their second revision coming up and have instances running on it.

But as far as I’m concerned, I think the pure economics of what ARM can offer, they have relatively good or too great performance depending on the task you’re going to run for a pretty much lower power budget than their counterparts in x86. So if you see what the major costs are in data center is the power to just run your chips and your servers. And on the other hand, it’s also the cooling of these chips.

So if they are more efficient and they generate less heat and they just consume less power, it also means that you have to invest less power in order to cool these chips. So that’s pretty much a very large cost reduction for Amazon. So I think that’s why they’re also experimenting with these chips.

That and the flexibility that gives them in order to design a chip that only has the parts they need because the x86 chips… They have pretty much every bib and bop you need in server hardware or desktop hardware. But with ARM, you can go with, I want a pipeline that only has the X, Y, Z, and I want to add a little bit of this and add a little bit of that in order to make the thing you want it to do.

So it’s, like I said, a bit of both a thing about economics and…

the best tool for the job. That’s cool. So you might trade a bit of performance for a substantial gain in power efficiency.

And if your workload happens to be multi-threaded and would benefit from having multiple cores, you can essentially run two servers instead of one and get less power usage, but perhaps more combined clock speeds available from your ARM CPUs.

Yeah, it depends what you need. Like you said, with ODEAR, you just need a lot of cores. I don’t think they need to be fast.

They just need to do their thing. And if you can do it with ARM instances, which are maybe less highly clocked, but a lot cheaper to run. That might be pure economics-wise a better fit for your business because you don’t need those high-powered beasts in order to run your app.

So if you’re going for a microservices or a very small server that just needs to run XYZ, it’s very small tasks, your home monitoring system, you name it. It doesn’t have to cost as much as the others do.

So you mentioned ARM CPU, but essentially that thing doesn’t exist. So ARM is just a licensing thing and other manufacturers make the CPUs. So if you say there’s an Intel CPU in a server, you know that it has been made by Intel, the company.

But if it’s an ARM CPU… You’re actually just referring to the instruction set that a third-party silicon provider happens to have made a chip out of.

Exactly. A few of the most known… Producers of ARM chips or with the ARM instruction set are Apple with their A-series chips.

They do heavily customized versions of those based upon the basic designs of what ARM provides them. And since recently, they also designed their own graphics cores. you also have Qualcomm, um, creating their own, uh, SOCs, uh, and Samsung did it until recently, I think, but I think they abandoned it because it’s, it’s really high cost business to be in, to design your own chips, even if you license it and have most of the designs yourself, because you still have to compete with other businesses doing the same thing. And also, um, putting their SOCs available to other vendors to make phones out of.

But those vendors don’t always use the ARM license in order to build mobile phone socks. I think Qualcomm is the creator of the Snapdragon 8CX. It’s a really high-powered SoC built for laptops.

And that’s one thing that Microsoft has demoed recently in order to run their Windows tablet slash mobile laptop platform on. So ARM is finding their way into the mobile segment, but it’s a far way off from the capabilities of what x86 can do, but it’s getting pretty damn close to what it can do. On the other hand, Microsoft is also experimenting with the new mobile chips from Intel as AMD.

So… It’s going to be a mixed market depending on what people want. Is it a thin and light with large battery but yet acceptable performance?

You can go to ARM. If you want a high-powered tablet with pretty decent battery life, etc., you can go to an x86 chip. So it depends, but many varieties of ARM chips are made for smartphones as for the desktop, as for the data center when we look at Amazon, for example.

It’s not getting any easier comparing CPUs online, is it? I’m already struggling with Intel versus AMD, but then there’s ARM, which exists from multiple vendors that each can implement their own version of that ARM CPU. So it’s going to be the only way, and this was to be expected, of course, the only way to know is a particular server fit for your purpose is to basically…

Buy them all, benchmark them, and then throw the ones away that don’t match what you need. That sounds like a lot of work.

That’s where the tech press come in, and that’s also a new source that I follow very vigilantly. Also on Twitter, because most of the reporters have their own Twitter account and do some on-the-site reporting, like on CES and many other developers or, let’s say… keynote presentations from major vendors like Intel, AMD, Nvidia, etc. What are they up to?

And then they’re reporting on it. So you can really get into the minutiae of things. And it’s a great community online as well.

If you have questions or want to get into it, it’s pretty open. And there’s a lot of interaction because if people have questions, you can actually add those people like a few days ago. Hey, can you ask this to company YZ about their new product on CES?

And that’s how really interesting debates and conversations happen. since the internet happened many many years ago it’s really easy if you know where to look and have some time to read upon a few things to get a sense of what do I need and where do I get it or just ask your friendly neighborhood connoisseur or something like that but we don’t all have a yan at our disposal to ask questions so some have to trust the mainstream press

So how did you ever fall into this rabbit hole? How come you’re my hardware guy? What is it that excites you about all the hardware that’s going on right now?

There has to be something that gets you passionate about all these things.

Well, I think it started when I was quite little and I took apart my dad’s alarm clock just to see what it… did and how it worked because it’s it’s one of those digital ones with the red segment led displays displaying the time and make the nasty chirping noises when you had to wake up in the morning um and i was just utterly fascinated about how did this work because you you have this analog clock thing you can see the handle moving you can see gears and stuff like that but the digital uh versus analog thing really captivated me i don’t know why but it’s just like I wanted to know what’s inside. So as anyone would do, I presume, I took it apart. And the challenge is always to put it back together and in working order and have no spare bits in your hands afterwards.

It’s a bit of a tricky venture because with that particular alarm clock, it took me a few tries over a few days. So my dad wasn’t very pleased in order to get it functioning correctly. So that’s how it started.

After that, with the rise of the internet and I got my own computer, I found some forums about CPUs and overclocking them and just got really… enamored with the fact that you can have an off-the-shelf product that is set at a certain frequency or can do this, that you can push it even higher. And okay, it wasn’t safe. It was a very interesting thing, even though I didn’t understand a thing about it.

Just reading and just mucking about with computers over the years. gave me a bit of an understanding how to do this. And as I mentioned, the tech press and Twitter, specifically Twitter, there’s some really good tech press and some… Data miners that dig up a lot of information about new chips that are about to come and the fact that you can design these intricate logic-based things at such a minuscule scale and it’s continuous race about getting those feature sizes smaller, getting those clocks up and just trying to… get ahead of the competition in order to just inch out in front of them, just maximizing their performance.

It’s not quite tangible, but it’s more of a feeling. I’m really, really happy to see when new hardware is out. So it’s something that always excites me for not always a reason that I… can define, but it just captivated me.

I think that’s interesting because I consider myself to be a reasonably well-informed IT guy. But the more I think about it, the more I realize that my knowledge and my skillset is almost entirely in the software realm. And there’s an entire hardware realm outside of this that is… if not more complex than the software side that I know absolutely nothing about.

And you seem to be drawn especially hard to the hardware side. I think that’s a cool niche to have knowledge in as well. It also doesn’t seem to stop evolving.

It never stops. You have the arms race between Intel and AMD. Then ARM comes in.

There are open source RISC CPUs being built. Makes me wonder though, where are we in like a year or five years or… A decade out.

Where do you see the CPU market in this case? And where do you see us evolve?

That’s a very tricky question because any assumption I would make is inherently wrong, but I can dream though. I think that the current way of chip design, which AMD pioneered a few years ago with their Zen architecture and now the Zen 2 architecture, is that they split up their logic and their compute into two different dyes. Up until a few years ago, every chip was made as a monolithic block of silicon, which made it prone to an error, which you could bin the CPU if it has a flaw.

So it would render it utterly useless, which makes it very expensive. So you can only get X amount of CPUs out of a big circular die, which they make the chips out of. But if you make the… compute parts very small, like on the new process, the seven nanometers, then you can have a lot of seven nanometer parts.

And if you have a very good yield, that means that the defect density on these 300 millimeters of wafer that they use is very small. That means that the chance that such one of these small chips… is defective and can’t be sold or is a lower tier part, it’s very low. So that’s very lucrative.

On the other hand, it’s less expensive to do your logic part or the part that does the interaction with your different buses to your cards and the memory and your USB, et cetera. If you make it on the current process, which is refined and cheap, and has very good yields, then you can make those two on a separate assembly line and merge them together at the last part. So I think that is a part of technology that Intel is also going to pursue because they also announced it.

They’re going to call it Voveros. But they’re doing it slightly different. They’re going to stack the different layers of the silicon on top of each other.

So they’re planning to do the I.O. part and then the compute part on top of that and then maybe a stack or two of memory on it. So system on a chip is going to become a whole different meaning. So I think that is going to be the future going ahead, stacking everything on top of each other because the time it needs the single to run from A to B is also going to be important.

Less latency is more performance. And just makes up for very cool chips. And just from a physics point of view, how do you cool a stack of very hot running silicon that’s stacked on top of each other?

How do you do that? So… Except that, we’re just going to go down the nanometer rabbit hole any further, even further.

We’re currently at the, quotes, seven nanometer process. We’re going to go to, I think, six. step between that uh five three uh one and a half and then beyond that it’s it’s it’s uh not certain um we have the new machines um that that are being built um that are going to make it possible, that run ultraviolet light as a very, very, very short wavelength in order to make those patterns on the wafers happen, which use up a few kilowatts, if not more, of power just in order to run that laser beam in order to make… the ultraviolet light happen. It’s going to be some crazy physics involved, which is a different and interesting part of it.

But those two things, just scaling it down, down, down, down, down, and trying to stack things up on top of each other or just using different chiplets, as they say, in order to combine awesome things, which also makes it possible to have a chiplet for purely graphics and then for graphics uh level 4 cache or just your memory and your i o and your compute so you can just make a chip a la carte if you want to make a super computer for example which you which you can just make whatever you want tailor-purposed so that’s cool so cpus are going to become more modular Yeah, pretty much modular because it’s just no longer doable to make monolithic mega dies because if you shrink down your transistors even further and further and further and further, if there is a defect happening, the defect will cover more and more transistors, which… Those chips are built for somewhat redundancy. If a cluster of transistors is hit, there’s somewhat of a backup routing path or something else, or parts of the chips they can turn off in order to sell it as a lower tier part.

But that’s going to be an issue. And that’s why they’re probably also from an economic standpoint are going to do that. But I do remember that I said those nanometers, that they weren’t all that.

It’s more of a marketing term somewhat earlier in the podcast. And before I forget to explain that, I’ll just jump into that. The term nanometers was used up until a year or eight ago, I think, for the actual feature size of a transistor.

So gate pitch, et cetera. But don’t go in that because that’s a whole other podcast on its own if it needs to be. But that was the actual physical manifestation of a transistor on its own.

That was when they used the planar transistor. That’s like a crossbar happening where the electrons could flow when one of the bars is turned on and off. But if you shrink and shrink and shrink and shrink such features on a chip, you get a problem which is comparable to single to noise.

It’s like in the audio world. If you have too much noise, you can’t make up if it’s a one or a zero. because there’s not enough electrons flowing through or they’re escaping or they’re leaking out on another way because there’s no longer a big enough or defined enough path. In order to combat that problem, they have developed what they call FinFET transistors.

So the path of the transistor where the electrons flow through, they have replaced that with a couple of fins which are standing in parallel next to each other going through the flow gate. That way you have more… surface where the electrons can flow through, which eliminates that process. But when they use the planar transistors, they could say, okay, my chip is X amount of nanometers big.

But then they used the FinFETs, which were a completely different design and also were a bit different in size and footprint. And that’s where, as they say, the marketing nanometers came into play. And everybody who had a fabrication facility could say, we have X amount of nanometers.

That’s good to point out that the current AMD chips are made at 7 nanometers and Intel were struggling with their 10 nanometers. Actually, in C, those two processes are producing pretty much the same size of transistor and are not that much different as they seem out just from the numbers point of view. So the 7 nanometers is pretty much the same as the 10 nanometer chips.

They’re just differently designed. They vary in their geometric 3D form from each other and the way they produce it with different steps. But just pure marketing term.

So don’t let marketing fool you that 7 nanometer chips are better than 10 nanometer chips. They’re pretty much the same. It’s just more of a PR rating as they call it.

Damn liars. And they had me fooled because I almost started buying CPUs with less nanometers thinking that they were better. So who is lying then in this case?

If the 7 are actually pretty similar to the 10 nanometer processes, which one is most correct? Is it more 7 or more 10 nanometers?

Well, I would say that the 10 nanometers is more accurate, even though it’s wrong on its own, because nothing on that rating of nanometers really coincides with the… the meaning of what it should be so in order to state that so but it’s it’s um uh it depends on how you view it it’s like uh a lopsided box um if you view it from one side it’s uh smaller uh more narrow than the other uh parts so basically everybody lies that that’s the point Yeah, that’s pretty much it. It just depends on your view because if you measure it from that angle, it’s that amount of nanometers if you offset it. And if you view it from the other angle, it’s X amount of nanometers.

It’s pretty much of a…

marketing term so lies as long as your chips end up being good that that’s the only thing that matters the benchmarks will uh tell you everything yeah that’s probably a more accurate result look at how fast things work how much uh thermal and power throughput you’re achieving basically look at the end result instead of the marketing term that’s on the box um okay I think that’s a good place to wrap up. I’ll conclude that every CPU is a lie and that I have absolutely no idea what I should be buying, except that I’ll need to benchmark a lot of servers to see which is best for my purpose. Okay, man, thanks a lot for coming on the show and for sharing your expertise.

I may have added more questions to my paper than I had resolved, but I am a lot smarter now if I look at my knowledge of CPUs. So I am very grateful for that. I’m assuming that we will have a lot of links to share in the show notes as well.

So if people want to read up on things, if they want to understand what the die or a yield or something is, you can find some links in the show notes that can keep you busy for a couple of hours. But Jan, if people would like to find you online and ask all my questions again to you, where can people find you?

Mostly on Twitter, I think. That’s the best place you can reach me. If I’m not mistaken, my Twitter handle would be at J underscore Somers or Somers.

So you could also put in the show notes if you want to.

No, it was really nice to talk about these things because I do that on and off in small bits with colleagues, with you guys. It’s just fun to talk about things that I’m passionate about. Although I’m not an expert, I just like to make people understand how these things work and see how I can improve upon their knowledge.

It’s just fun for me to do. It’s nice.

Yeah, same here. Just a lot of fun to talk about geeky stuff like this. If you, the listener, have made it to the end, first of all, congratulations and welcome.

I’d love to hear feedback. This is the first episode in a pretty long time. I took a really long break and I hope to get into this more regularly.

But I could use some feedback. Was this an episode that you liked? If not, what didn’t you like?

If there were bits that you did like, please let me know as well. Don’t just give me all the negative feedback. That’s not really motivating.

Let me know what works and what doesn’t work. I’m trying to find a good method of podcasting, a good format, length, content, you name it. But in order to nail that down, I need feedback.

I’ll give all my contact details in the show notes as well. So you can reach me on Twitter or… via email perhaps I should do some kind of anonymous contact box so you can just troll me but I need feedback so please do let me know and Jan as well what you thought of this show and what you would like to see more of with that said I’ll talk to you next time goodbye