EP028: ARM In Post Production Part 1

00:00:00:00 - 00:00:14:19
Robbie
Hey there, and welcome back to another installment of The Offset Podcast. And today we're talking about ARM processors and why they should be important to you and in the post industry in general.

00:00:14:21 - 00:00:33:13
Joey
This podcast is sponsored by Flanders Scientific leaders in color accurate display solutions for professional video. Whether you're a colorist, an editor, a DIT, or a broadcast engineer, Flanders Scientific has a professional display solution to meet your needs. Learn more at Flanders scientific.com.

00:00:33:15 - 00:00:40:01
Robbie
Hey everybody, I am Robbie Carman and with me, as always, is my partner in crime, Joey D’Anna. Joey, how are you doing, buddy?

00:00:40:03 - 00:00:41:06
Joey
Good. How are you?

00:00:41:08 - 00:01:01:18
Robbie
I'm doing swell. Now, Joey, I know that this is an episode that you have been waiting to record for a long time. Simply because you are a computer nerd. And I have to give it a little inside baseball behind the scenes. A little, you know, sort of story here for our audience. And that is, many of you might know that Joey has a large vehicle collection.

00:01:01:18 - 00:01:27:02
Robbie
He has, I don't know, 15 cars, 20 motorcycles. You know, probably boats and planes that we don't know about. But besides that large collection of vehicles, Joey also has a large collection of antique computers. He's probably the only person that I know that has a legitimate super computer array in his basement. He has various Atari's and next computers and SGI boxes.

00:01:27:04 - 00:01:52:01
Robbie
Going back to, you know, the early to mid 80s is a little bit of a connoisseur and collector of these things. So when it comes to computing, computing history, how computers do their thing and work, Joe is always my go to, reference for these kind of things. And, you know, just this week, Apple, announced some new, new computers, some new, new Mac studio.

00:01:52:01 - 00:02:14:19
Robbie
A couple months ago, they announced on Mac mini, and they, as you all probably know, have for the past several years have been championing their own processors, having having made a switch over from, Intel processors, for about a decade or so that they were using those, and now they're creating what they generically refer to as Apple Silicon, but it's not something that is brand new.

00:02:14:19 - 00:02:30:22
Robbie
Actually, this is based on, what's called ARM technology or ARM processors. And so we want to talk a little bit, a little bit about that today because we got a lot of questions, popping up, in email and different forms that people, who people who know are saying, hey, what new Mac should I buy?

00:02:30:22 - 00:02:56:13
Robbie
What what's the deal with unified memory? What's the deal with GPU cores, all that kind of stuff. And honestly, Joey, I think it's a it's a subject we've been skirting for a while because, you know, Apple, you know, has been Apple. And you know, they use a lot of really fancy marketing. And it's, you know, it's it's tough sometimes to wade through that, but also I think that we generally have the attitude of we're not sick about computers, right?

00:02:56:13 - 00:03:15:14
Robbie
It doesn't matter if it's a windows box, a mac, a Linux system like do whatever you're comfortable with, do what your, you know, it gets the job done best. But in this particular case, we thought that, it probably be a good time to talk a little bit about ARM just because this new announcements from Apple, a lot of people are in that cycle.

00:03:15:14 - 00:03:36:00
Robbie
Maybe they invested in the M1 series, you know, 3 or 4 years ago. Now it's time for an upgrade. So let's cover this. But before we do, I want to make one asterisks. Caveat we are not going to, in this episode, go into a deep, deep, deep, deep, deep dive of chip manufacturing. What I say is lithography, but you say it.

00:03:36:00 - 00:03:36:15
Robbie
How do you say it?

00:03:36:20 - 00:03:37:16
Joey
Lithography.

00:03:37:18 - 00:03:38:23
Robbie
Lithography. Okay.

00:03:38:23 - 00:03:42:02
Joey
However, plate photography. But with list.

00:03:42:04 - 00:04:04:07
Robbie
There we go. The idea of how the fabrication of chips are made. There's plenty of information that on that on the web. If you want to take a deep dive of how the billions and billions of transistors and interconnects, we'll touch on some of that where necessary. But I just want to just preface this by saying, if you are a, you know, a CPU designer, this is not your episode, okay?

00:04:04:07 - 00:04:24:10
Robbie
We're talking to, fellow colorists, editors, people who are looking to use this technology, for creative, creative reasons, methods rather and reasons. If you if we say anything that you think could use more clarification, just let us know. Along those lines, of course, you can always, follow us on social media. We're on Instagram and Facebook to search for the offset podcast.

00:04:24:10 - 00:04:50:03
Robbie
You can always go over to offset podcast.com, and check out show notes. Of course, we're on YouTube as well. And wherever you find the show just to its favorite, like subscribe, download, tell your friends, the more people we get, eyeballs on, the better it is. So, Joey, let's, let's dive in. Man, now that we got all that out of the way, I guess place I want to start on, give us a little history of, of, what is ahm, why is it, you know, that kind of thing?

00:04:50:08 - 00:05:18:21
Joey
Ahm, overarching as a CPU architecture and what's called an instruction set, it's basically what the CPU instructions do represent, how they're laid out in an actual, you know, piece of hardware. It's basically the standard of what the CPU or central processing unit of the computer adheres to. So that doesn't mean that all ARM processors are the same, just like all x86 or x64 compatible processors aren't the same.

00:05:18:21 - 00:05:54:00
Joey
You've got processors compatible from Intel from AMD. Back in the day, there were a couple other companies. These days it's really just Intel and AMD, but multiple manufacturers make their own CPUs that are compatible with the ARM architecture. That means when you compile software for ARM, it will work on that CPU. It's different than what we're used to with Intel, x86 64, which is kind of the that was pretty much the standard for workstation level computing for a very long time because it got really, really good.

00:05:54:00 - 00:06:01:22
Joey
That's why Apple adopted it. You remember long before the Intel Macs, there were the power PC Macs.

00:06:02:03 - 00:06:12:19
Robbie
Which was it? That's great. It's kind of like the same analogy, right? Where back then, you know, even though Apple wasn't making or, you know, they were relying on a different company other than AMD and Intel. In this case, it was.

00:06:12:19 - 00:06:14:20
Joey
IBM or IBM.

00:06:14:22 - 00:06:15:02
Robbie
Then.

00:06:15:03 - 00:06:23:23
Joey
Who was maybe people called it the Apple G3 or G4 or G5. It was a Apple specified IBM unit.

00:06:24:01 - 00:06:25:23
Robbie
Got it now.

00:06:26:01 - 00:06:52:03
Joey
Same thing moving forward with with ARM. Apple is specifying a lot of stuff and then having their OEM manufacturers make them. Now right ARM is what's called a RISK architecture. Reduced instruction set computing where essentially the amount of things the processor does, like the amount of options you have for instructions to give it, whether it's like, oh, move this memory, add this, subtract that, multiply that.

00:06:52:08 - 00:07:24:10
Joey
They're broken down into a much smaller group of fundamental primitive instructions. So whereas certain architectures like Intel might have instructions for really detailed mathematical operations that take a long time to run, the idea of risk is all the instructions run really, really, really, really, really fast because they're all really small parts of the program. And because of that, we can optimize the processor to do lots of cycles really quickly, and we can make them more efficient.

00:07:24:12 - 00:07:49:06
Joey
RISK and non-RISK have kind of butted heads back and forth over the past decades as to who was in practice, really the best performer for the best workloads. And a lot of this really goes into software engineering and how you take advantage of the processor. But these days, ARM is a risk architecture that Apple has decided to move their entire product line to.

00:07:49:08 - 00:08:13:06
Joey
You know, and it is it's been around for a very long time. It's a very mature, architecture. It's very powerful, and it has some really cool features and things about it that really work well for desktop computers, phones, smaller computers, and in our case, for post-production and image manipulation.

00:08:13:08 - 00:08:30:08
Robbie
Well, before you go on, I want to there's a lot to unpack there. I want to make a couple, statements, questions about this. So, so ARM, as you said, has been around for a while, but where where has it been? Right. Because, you know, I think from, you know, from my perspective as a consumer. Right. Of course, I know about Intel.

00:08:30:08 - 00:08:48:05
Robbie
Of course I know about, AMD and that's kind of like those were the two, you know, two games in town, right. Like. Yeah. And before that it was pretty obvious because, oh, Apple was pushing PowerPC with, you know, G, 3G, 4G, five, that kind of stuff. But it's not like ARM hasn't been around, as you said. But where has it been?

00:08:48:05 - 00:09:05:21
Joey
It's actually mostly been in embedded applications, small little efficient microcontroller or above microcontroller level things. In your smart devices, in your car and any kind of thing that needs processors.

00:09:05:23 - 00:09:06:13
Robbie
Okay.

00:09:06:15 - 00:09:35:01
Joey
And the cool thing about ARM is it's flexibility. We'll talk about this a lot in terms of flexibility. Right ARM can scale from they just came out with one that's like less than a millimeter size. It's like the new record holder for the smallest CPU to as we'll talk about gigantic 128 core 256 core data center monster's. ARM is designed to be able to scale in ways that other architectures might not have been able to.

00:09:35:02 - 00:09:56:22
Robbie
Okay, so that makes sense. And so now. So basically there is a foundry, you know, TSMC or somebody else making these ARM processors, these companies like Apple and others are now saying, well, Apple being the size that they are, as you pointed out, it's not like they've, you know, they they say they designed the chip. That's a little bit of a misnomer if I'm guessing what you're saying.

00:09:56:22 - 00:10:07:13
Robbie
They've they've had a lot of input on the features and the architecture that they would like to see, but it's not like they're the ones actually making the chips.

00:10:07:15 - 00:10:33:04
Joey
Yes and no. Apple has gone to what I consider to be kind of a higher level of OEM with this. Yeah, yeah, most people have, you know, they're not just saying, hey, I need a ARM of this clock speed with this much memory on board and this much cache and this many cores. Right? They are going in and designing their own interconnects, which we'll talk about later.

00:10:33:04 - 00:11:06:16
Joey
And most importantly, they're designing their own GPUs. Whereas the CPU part of this system on a chip package is, you know, there's parts of it that are kind of used between multiple products. The GPUs on the new Apple systems are that's all 100% designed by Apple. And I have to say, everybody that's used one of these new ARM Macs can attest to, they kind of came out of nowhere and became like heavy hitters in the GPU world in terms of performance.

00:11:06:18 - 00:11:28:22
Robbie
All right. So that makes sense. So they have, a core design that they're outside of the regular vendor. Like if you and I just wanted to get some ARM processors for our new, you know, widget, whatever, we would probably buy something off the shelf. But Apple being the size of they are, the product they're moving has a lot more say to go, no, these this is the design considerations we want you to make make this for us.

00:11:28:22 - 00:11:29:14
Robbie
Like this.

00:11:29:16 - 00:11:50:03
Joey
They're designing what's called the package, the system on a chip or an SoC. So that's got the CPU multi-core CPU. It usually in the case of these systems has memory built in as well. And the GPU and anything like external interconnects, whether or not it has access to PCI slots, whether it has serial ports, all of those things.

00:11:50:03 - 00:11:58:14
Joey
Apple is designing the entire package. Whereas, okay, if we were going to make a product that had an ARM in it, we would buy something that you could buy a development kit for.

00:11:58:16 - 00:12:21:11
Robbie
Yeah, right. Or go to a CD or something, buy buy some processors. Right, I got it. Okay. That that that all makes sense. So where I guess one of the things that I was, you know, the rumors about, Apple specifically going to ARM started a long time ago. You know, there was oh, we're unhappy with Intel. Intel's, you know, too much thermal problems, too much, you know, power problems.

00:12:21:16 - 00:12:52:07
Robbie
So I think that writing was on the wall for a long time. But when they when they made that change, you know, for me, I think I think the first time that I even really hurt ARM was when Apple I think released to developers, they released a mac mini style developer CPU. Right. Because as you said earlier, part of the challenge with this changeover was, hey, we have a whole different way of, you know, a whole new instruction set and a whole new way that, you know, underlying program has to be handled by this new architecture.

00:12:52:11 - 00:13:03:17
Robbie
So they had they got it in developer's hands for a while to, to get it out. And then they had, you know, for a long time, I still I think it's still around, actually. But, you know, Rosetta was an application that they had to kind of translate.

00:13:03:17 - 00:13:05:10
Joey
It's still in the OS. Yep. Yeah.

00:13:05:15 - 00:13:30:05
Robbie
Older trans older instruction sets, you know, Intel or x86 to this new architecture. But now they, they came out with M1 3 or 4 years ago, something like that. And now we're up to M3 and M4. So it's obviously generation. And to riff on what you just said a second ago about the combination of Apple and the, the, the actual foundry, you know, that has been an really interesting thing to watch.

00:13:30:05 - 00:13:52:17
Robbie
Kind of the development, obviously core counts gone up, the type of core, whether it's what they call efficiency cores versus performance cores, how they do on the GPU. So it's not unlike the old, you know, moniker of Chip generations. They refer to this sometimes as nodes or die generations or something like that, where it's like, hey, we're going up and getting capabilities of each one.

00:13:52:20 - 00:14:02:07
Robbie
So where we're now on M4 is the latest and greatest. M3 is also being supported. There's still some M2 in the system as well. All right. That oh that all makes sense.

00:14:02:07 - 00:14:26:14
Joey
And really we should be calling them like M 25 or whatever because. Right. You know these SoC packages have gone way back before because all of your iPhones before Apple even moved ARM for their other stuff, were all ARM based SoCs. This that's kind of what this evolved from with Apple. They you know, they started making the phones out of ARM SoCs and then they realized, wait a minute, these phones have a lot of compute in them.

00:14:26:16 - 00:14:38:13
Joey
And if we just make this a little bit bigger, well then we can kind of control the whole ecosystem. And we could have one unified family of architecture. And that's really attractive to Apple.

00:14:38:15 - 00:14:53:13
Robbie
That's a great segway into what I want to ask next. Right? Is like, why? Why would a company the size of Apple make this move? And I just want to be clear about one thing, because I you sort of alluded to it, but it's not in the consumer space. It's not obviously just Apple who's doing this now. Right?

00:14:53:13 - 00:15:15:03
Robbie
There's Qualcomm, Microsoft, all of the big players are going, wow, you're right. This is a big improvement on a lot of levels from what we're doing. And they're all offering products now that shape or fashion support ARM with windows or hopefully we'll see some more Linux out there. But I think that's a good segue into why this move was made.

00:15:15:03 - 00:15:25:18
Robbie
And I think the vertical integration part of it that you just, you know, I guess both horizontal and vertical, that you mentioned about Apple with the phones is a big part of this, right?

00:15:25:22 - 00:15:47:13
Joey
Y but, you know, you want to talk about some of the advantages, you know, why would a company move their products to, you you hit the nail on the head when you said efficiency. Yeah. Right. Any computer puts out heat, takes electricity. So for any amount of computing power you have, it takes a certain amount of electricity.

00:15:47:18 - 00:16:20:04
Joey
And it makes heat out of some of those. Some of that electricity. And it makes compute out of the rest, ARM massively better power efficiency and thermal efficiency than Intel. That is the motivation. That's why it was great for embedded applications. That's why it was great for phones. It's good for battery life. That's why, you know, when it comes to these smaller systems like the Mac Studio on the Mac mini, they could put incredible power in something that doesn't need a gigantic brick of a power supply or massive heat sinks or massive fans.

00:16:20:09 - 00:16:39:09
Joey
And that concept of it's just more efficient scales all the way up from your Apple Watch to the biggest data center in the world. Because for every Hertz of compute that you do, if it's more efficient, it doesn't matter how big you get, you get advantages from that efficiency.

00:16:39:11 - 00:16:56:01
Robbie
Yeah, I see that. And I think that's something that they've always been interested in is power efficiency, battery life long term. And I think that comes out of that phone. But I had to look this up because, you know, it's one of those things where it's like, you Mac Studios become a really popular computer, like power savings. Is it real?

00:16:56:03 - 00:17:03:17
Robbie
Well, let me look at I look this up, and I just want to tell you that the Mac Studio is a max power of 480W. Full, full bore.

00:17:03:19 - 00:17:04:23
Joey
No, I'm just.

00:17:05:01 - 00:17:30:08
Robbie
I'm just thinking to myself. I'm looking over this shoulder to a server I have sitting here in a rack. That's like a mask that I built and that has, I think, a 1600 watt or an 1800 watt power supply in it. Right. And like, that's kind of become like the normal for a workstation. Performance is like, you know, somewhere between 1500 and 2000W when you factor in CPU of 280 and the GPU pulling 500 or whatever, it all adds up.

00:17:30:10 - 00:17:31:22
Joey
Can I put it in perspective?

00:17:32:00 - 00:17:32:13
Robbie
Yeah, yeah.

00:17:32:17 - 00:17:45:02
Joey
You mentioned I do collect old computers, and part of that is I have a vintage, 1996, I think it was made full rack Silicon Graphics, Onyx two.

00:17:45:04 - 00:17:46:02
Robbie
It's a big computer. Yeah.

00:17:46:02 - 00:18:04:15
Joey
Basically supercomputer. It's got a computer module and a graphics module. Each one has half of the rack. When you plug that thing in and you turn it on, you don't even turn on the computer and you don't even turn on the graphics. The only thing it comes on is one of the fan trays. That's 800W. I've measured it.

00:18:04:16 - 00:18:06:10
Robbie
It's crazy. It's crazy. So.

00:18:06:12 - 00:18:17:04
Joey
You know, it's not just bringing your power bill down. It's the fact that they can put this in a smaller package and put more in the same amount of space for the same amount of power.

00:18:17:06 - 00:18:34:16
Robbie
Yeah. And I think I think the overarching point, though I want to make is that, yes, there's the efficiency, the power savings, the thermal performance, I think is a big one, especially if you're, you know, if you're working with a machine at your desk versus, you know, in a machine room, you want it to be quiet. You don't want those fans spinning up and annoying you.

00:18:34:16 - 00:18:52:01
Robbie
And like that. All that adds up to smaller, quieter computers that pound for pound punch above their weight class. But you know, I think a lot of people go, oh, well, it's the end all be all power efficiency. It doesn't get better. I mean, like, power still wins, right? And we've seen that. We'll discuss this in a little bit.

00:18:52:03 - 00:19:19:09
Robbie
You know, if you have the ability to throw 2000W at something and you don't care about the thermal performance because it's in a rack in a data center. Yeah, you're okay. You're not going to be like, no, I don't think anybody's making the argument that, loaded Mac Studio is ever going to run circles around, you know, whatever the latest Nvidia GPU with, you know, 96 core Threadripper, you know, that has pulling 3000W or something, you know?

00:19:19:13 - 00:19:41:15
Joey
Well, that kind of brings me to my next kind of advantage of ARM And this isn't a ARM specific thing, but ARM and especially in the context of the ARM Macs have made it a big thing, that differentiates it from the Intel generation setups is this thing that we talked about a little bit. We're going to talk about more called unified memory.

00:19:41:17 - 00:19:54:16
Joey
Unified memory is an advantage that I cannot state enough is basically a cheat code for our specific type of computing need. It's okay.

00:19:54:16 - 00:19:55:07
Robbie
So before load.

00:19:55:08 - 00:19:57:03
Joey
For all computing needs.

00:19:57:05 - 00:20:19:10
Robbie
Right. So before you go on, I want to make sure I have this part right. Because I think this has caused a lot of confusion because I see, people still referring to this memory as Ram, which is I don't think I don't know if that's technically right or wrong. It's still Ram, but it implies to me when people say that it implies a slightly different architecture than what you're describing.

00:20:19:10 - 00:20:38:18
Robbie
Right? So tell me if I got this right in a traditional computer motherboard processor set up, you got your processor sitting in a socket that interfaces with the motherboard, and there is a, there's a, there's a bus, a memory, a pathway, a highway going out to that memory. And the CPU talks that memory and it's bidirectional, goes back and forth and data is moved.

00:20:38:18 - 00:21:09:03
Robbie
Instructions are moved from the memory of the Ram back to the processor and vice versa. The advantage of having this all on what you referred to earlier as SoC or a system on a chip, is that that pathway is right there with the CPU next to it, with no way to have to navigate. No, you know, bus or whatever to have to go out to it's integrated, meaning that it can talk loads faster to get things done, considerably faster than having to go out on that highways.

00:21:09:03 - 00:21:10:11
Robbie
I mean, that's an oversimplification.

00:21:10:11 - 00:21:33:02
Joey
That's but that's that's part one. Part one is that, yes. When you put the memory physically next to the CPU, it actually does make a big difference in terms of performance and capability. But whereas when you buy an external GPU for your Intel machine, it has it's own memory. So the GPU needs to do operations. It's looking at its memory when the CPU needs to do operations, it's looking at its memory.

00:21:33:02 - 00:21:41:10
Joey
And what the CPU is going to have to do is it's going to take stuff from its memory, give it to the GPU, tell the GPU what to do, then takes off the GPU memory, give it back to the system memory.

00:21:41:14 - 00:21:43:13
Robbie
Well, there's a lot of back, a lot of back and forth, a lot.

00:21:43:13 - 00:22:02:16
Joey
Of back and forth memory bandwidth to and from a GPU in a traditional system is a massively important thing to be cognizant of. That's why PCI went from, you know, different generations, every generation. It gets way, way, way faster. And that's always a big important thing is if you're using an external GPU, you need to be able to talk to it.

00:22:02:18 - 00:22:04:00
Robbie
As fast as you possibly.

00:22:04:00 - 00:22:11:01
Joey
Can. Awesome. Because think about this. What's our workload? What is our compute need for a GPU?

00:22:11:03 - 00:22:16:23
Robbie
It's an old French. Yeah, right. But it has it has to happen. We expect it to happen basically in real time. Right. So we.

00:22:16:23 - 00:22:34:21
Joey
Want actly. So the instead of saying, hey, here's a 3D complicated 3D scene, render it right where in that case you give the GPU a little bit of information and then you're waiting on the GPU to turn, turn, turn, turn and render the image. Right? No, no no no. We're saying, hey GPU, make this one stop brighter, right?

00:22:35:02 - 00:23:02:20
Joey
That's a really fast operation to do. But then you say, hey, make this one stop brighter at 24 frames a second uncompressed floating point 4K resolution, and then do it maybe 3 or 4 times over because I'm doing a temporal effect with noise reduction. Right. You're moving huge amounts of data in and out of the GPU in real time all the time, and then asking it to do very comparatively simple math to it.

00:23:03:02 - 00:23:33:18
Joey
So where does unified memory come into play here? Well, on these new ARM systems now, not all ARM systems are like this, but on the Snapdragons and Apple ARM systems, they have what's called unified memory. The memory itself on this die that also shares the CPU and the GPU is shared by the CPU and the GPU. So if you have 128 gigs of memory, you have 128 gigs of memory combined CPU and GPU, and the system can divvy it up how it how it sees fit.

00:23:33:20 - 00:23:55:03
Joey
Two big advantages there one. When was the last time you bought 128 gig GPU right there. That would be pretty expensive. Well guess what, now you get it in your Mac studio so you get more GPU memory, but more importantly, okay, once the CPU reads some frames off, your storage puts them into memory, the GPU just needs to know what address they're at.

00:23:55:09 - 00:23:56:03
Joey
They're already there.

00:23:56:05 - 00:23:58:07
Robbie
It's it's instantaneous. Pretty much. Yeah, I guess.

00:23:58:07 - 00:24:18:22
Joey
Then if the GPU says cool, here's your one stop over version, I'm going to put that into memory for you. Well, the CPU could just say cool. What's the address, bro? Because, I've got that same memory. We're essentially taking the send to the GPU memory and then get from the GPU memory steps completely out of the equation.

00:24:19:03 - 00:24:42:20
Joey
So for a workload like color grading or any kind of real time video or image manipulation, unified memory, look, I'm not going to say it's the end all, be all. And you couldn't build a discrete GPU system that has enough bandwidth to beat out these new unified memory systems. But unified memory is a big, big, big performer for what we do.

00:24:42:22 - 00:25:05:01
Robbie
Yeah. Now it makes it makes a lot. It makes a lot of sense. So the other thing that strikes me about these systems, and I don't know how to say this in a succinct way, but it's like the you alluded to earlier, the scalability up from the smallest, you know, one millimeter size thing up to the big, you know, data center behemoths.

00:25:05:02 - 00:25:25:20
Robbie
And what struck me about specifically what Apple is doing, and I think, Qualcomm's doing this a little bit with Snapdragon as well, is, you know, kind of the, the, the levels of gradation of these chips, right? Like from, you know, an Apple parlance, it's the regular, you know, just no, no, no added name, just M4. And then it's, you know, pro, then it's Macs and then it's ultra.

00:25:25:20 - 00:25:46:08
Robbie
Right. And let's talk about that for a second, because I think that, you know, on the surface it's kind of like the old days. Apple I don't know if anybody remembers. But in the old days, back in the PowerPC days, the Apple Store actually used to be good, better, best kind of like options. And, you know, nobody wants to go back to that nomenclature because they don't want to tell people that you're buying a, you know, an old computer.

00:25:46:08 - 00:26:15:11
Robbie
And I tell you, you're buying an awesome one. But the difference between those generations is, is, is a little bit of a multifactor thing. It's obviously, number of cores that you're getting, you know, both GPU and CPU cores. Yeah. What what type of, types of forces are whether they're performance or efficiency cores. Right. But there's also like some differences in terms of, like each version of that SoC, the bigger you go can do things like support more unified memory, for example.

00:26:15:11 - 00:26:25:19
Robbie
Right. So like get more more and more memory can be used for the GPU and CPU. Is there anything else that I'm missing between like those like versions like Macs Pro Ultra, right.

00:26:25:19 - 00:26:54:22
Joey
I mean basically the way all of these companies are scaling these things up is by adding more small pieces, right? So when you go from like the M1 to the M2 to the M3, the individual cores are getting faster and more performant right now. But when you go between the sizes of them where you have like, okay, here's well, I forget what they call like the max has a bunch of cores and then like the ultra has two of the maxes together.

00:26:55:00 - 00:27:22:15
Joey
That's kind of where the scaling really starts coming into play. Right? Because these what they call the package, which is the, the, the chip that has the CPU cores, the GPU cores and in Apple's case, the memory as well. Right. That can only get so big, right? They're never going to be able to make that big enough to fit a monster, you know, workstation level computer.

00:27:22:17 - 00:27:45:10
Joey
So what they do is they take multiple of them and put them together in a multiprocessor configuration with a very, very high speed interconnect right next to them on the same board, like we talked about with memory. The closer it is, the less physical connections that has to be made. The more efficient it can be. And that's why you get into these huge configurations like the like.

00:27:45:16 - 00:28:03:20
Joey
Okay. So on the new Mac studio, you have the M4, which is the latest generation processor, but they don't have the ultra or basically dual chip version of that because they haven't made an interconnect fast enough to take two of those in fours and zip tie them together.

00:28:03:20 - 00:28:20:02
Robbie
Yeah. So in Apple parlance, they call that the Ultra Fusion interconnect or something like that. Right? That's just a it's a fancy marketing term for that physical connection between two separate SoCs that they've linked together to act as one.

00:28:20:04 - 00:28:43:16
Joey
Exactly as those SoCs get faster and faster and faster, obviously you will need a faster, more data capable interconnect. That's why in all of these generations, one thing a lot of people ask about is, okay, why don't we have the latest and greatest, highest number basically in the biggest configuration. And it's important to remember, you know, essentially all of these Apple associates start on the phone first.

00:28:43:18 - 00:29:02:14
Joey
They start on the phone on the iPad first, and they scale up from there. That's how it's been since day one. The original dev kit was basically an iPad without a display, and then they start making them bigger and bigger and bigger, and then they get to a point where it's so big they can't really make it any bigger.

00:29:02:16 - 00:29:25:13
Joey
So then they start adding two of them. And that's where you get these monster Mac Studio configurations that really compare with modern day Intel workstations, because they just have so much GPU and they're so powerful. But what you're looking at there is basically two of the highest capacity chips that they can make and still put an interconnect on, put together.

00:29:25:15 - 00:29:42:01
Robbie
Yeah, yeah. So one thing I think is interesting about this architecture and it's, it's it's been a pretty market change, I think, from the way that I used to think about CPUs in particular. So like when I in the past when I would go CPU shopping, one of the first things I would look at is the clock speed, right?

00:29:42:01 - 00:30:04:02
Robbie
I go, oh, well, you know, a three gigahertz chip's going to be faster, better than a, you know, a 2.2GHz chip. And, you know, you'd look at other numbers like cache size and stuff. And noticeably missing from a lot of this ARM marketing is that discussion of clock speed. And I think the more that I dug on this, it's partly because that clock speed is is highly variable with these chips, right?

00:30:04:02 - 00:30:25:20
Robbie
Depending on what it's doing. Right. Like M4, I think can have a burst of up to, 4.3 or 4.5, you know, gigahertz, something like that. But it's interesting from a marketing point of view that, that, that that has been a shift like, rather than getting people to fixate on actual clock speeds, it's more about core performance, the thermal, like they've changed the narrative.

00:30:25:20 - 00:30:27:08
Robbie
It seems like a little bit on and.

00:30:27:08 - 00:30:54:11
Joey
I gotta say, I think that's the right thing to do. I don't think Clock Speed has ever told the whole story, especially on these modern CPUs, because you look at we keep using the Apple example because it's the most, most prominent, you know, on that chip. It has built on H264, HDR 65 and ProRes encode and decode. So it wouldn't matter if the CPU was 1GHz or 1,000,000,000GHz for decoding ProRes, or.

00:30:54:13 - 00:30:55:21
Robbie
It's got a dedicated chip for it. Right.

00:30:55:22 - 00:31:12:05
Joey
Because it's already built onto a hardware decoder right there. All it's doing is waiting for the hardware decoder to give it the information. So the overall design of the system is now taking kind of more of a front seat versus what I consider to be somewhat antiquated metrics, which.

00:31:12:05 - 00:31:36:22
Robbie
Yeah, I understand. So I think I think, you know, now that we've gotten our head around the basic design of it, I think one outstanding piece that still seems like to be a battle. And I think to be fair, the comparisons, the benchmarks do have some validity to it is that, you know, in our line of work and editorial and post and color and whatever the GPUs obviously become arguably the most important part of the kit.

00:31:36:22 - 00:31:56:20
Robbie
Right? Because that's the that's the thing that's doing them. Heaviest lifting on image processing, has, you know, perhaps the biggest impact on real time this, you know, of performance that we've come to expect. And so I think it's a little bit of a weird thing for people to go. What do you mean? There's not a, you know, I don't have to choose a GPU for this, right?

00:31:56:22 - 00:32:22:02
Robbie
I'm just getting told this is the GPU. Right? And so, you know, there's a lot of people who want to compare, oh, this is, you know, the latest and greatest from Nvidia or AMD. This is how it's comparing to Apple. And it seems to me that there's still a gap. Those dedicated discrete GPUs, you know, most recently you know from Nvidia like the 5950, 80 those ones, you know, those are still beefier GPUs.

00:32:22:02 - 00:32:42:12
Robbie
I mean, they're still doing heavier lifting and systems that are set up properly, but they do have some of the issues that you talked about. You know, bus speeds a factor memory, dedicated memory on them. You know, there's only so much they can you have what's your feeling about? And you alluded to this earlier where you said, hey, surprised everybody that Apple's been, you know, competitive at least with GPUs.

00:32:42:14 - 00:33:00:19
Robbie
Do you think we're going to get to a point where the difference between a discrete GPU in, in an embedded Asus GPU? Because I think a lot of people have a little PTSD about embedded because they think about like, oh, you mean like that Intel, you know, GPU, you know, crap that we had for years and suffered through that never performed right.

00:33:00:21 - 00:33:12:19
Robbie
But but where does that lay out now? Because it's like it's not quite as good as a dedicated GPU yet, but it's still holding its own. What do you think the future, if you had to probably prognosticate a little bit, stands for that. That SoC comes.

00:33:12:19 - 00:33:39:11
Joey
Down to the workload. In a lot of cases. Like I said, these these integrated GPUs are really good for, one particular task. They're really not as good for things like 3D rendering or machine learning training, stuff like that. In those worlds, the big discrete Nvidia guys will absolutely wipe the floor with any SoC GPU, but even in color grading you get a super heavy noise reduction, super heavy temporal stuff.

00:33:39:13 - 00:34:02:01
Joey
Yeah, those new 5090s will outrun the the best Apple has to offer. That is where we kind of get into this. Interesting question is there's no reason why you can't have an ARM CPU with a discrete GPU. It's just nobody's really put that into a desktop setup yet.

00:34:02:03 - 00:34:18:20
Robbie
All right. Hold that thought for a second, because we've talked about a lot and we're already running a little long. So how about this. How about we come back in part two. And in that part two we'll talk about what we think the future of ARM is. And what Apple's doing, prognostications of where the industry is going with cloud based ARM and all sorts of fun stuff.

00:34:18:20 - 00:34:24:23
Robbie
So, stay tuned for a part two. Until then, thanks for watching. I'm Robbie Carman

00:34:25:01 - 00:34:32:00
Joey
And I'm Joey D’Anna