[Music] Hey guys, welcome to the show! Today I have with me the 128 GB version of the

MacBook Pro M4 Max, also the 64 GB version, and I figured I'd just do a comparison

on why you might need 128 GB. Do you know why? Can you tell me why? Write in

the comments right now why you need 128 GB. Future-proofing? But next year this one is going to be faster, so, you know, do

you need to future-proof? Do you need to proof it, or do you want an excuse to buy next year's one? So maybe

it will be good out of the box for that. But anyway, I'll tell you the answers and I'll

show you some demonstrations after I tell you what happened. So, if you want to run Llama, the 405 billion parameter

version, it's not going to work on the 64 GB version. However, you can get it running on 128. It looks like you there you go.

And 5 years later we'll get a response. It runs very slowly, very, very slowly, but it is possible. You just have to use the

quantized version, the, you know, the compressed version, and it runs. And that is pretty impressive. If you want to do

some sort of private AI on your local system and you don't want to reach out to the ChatGPT privacy people in the

world stealing all your data, allegedly. I know what you're doing. Legal disclaimer: no, I don't. If you want to convert text into

video, it does run on a 64 GB version in standard situations. However, it doesn't run out of the box on a 128, why? Because

most people do not have 128 GB. So there is currently a bug in which it asserts, which means it crashes into the Metal

shaders. That bug's been around for the past 2 years and it's on people with 128.

GB systems haven't been fixed because there aren't enough people complaining about it. So maybe if you get 128 and

complain about it, you might be able to fix it. So, Metal Performance Shaders and its array situation, it's

probably an easy fix if someone gets around to fixing it. It's just an assert; there is a workaround. I did work around

it, and I ran them both on these two systems, and the 128 actually ran more than twice—it finished the

job in 10 minutes, whereas the 64 GB, it did do it, but it did it in 21 minutes.

And that's because it had to constantly thrash the model in and out of memory between the SSD and back into memory,

and that really was a big bottleneck for it. So that's what really slowed it down compared to my PC with only 16 GB RAM

and a 10 GB graphics card, the RTX 3080. That one ran it faster; it was about 50%

faster running it than my M4 Max. But that being said, this code isn't really that optimized for Max's situation. It

was doing; it was running on a GPU, but via the CPU rendering pipeline. So there are improvements to be made. So the fact

that you actually got it up and running is impressive, and the fact that it ran Llama 3.1 405B is impressive to

me, even though it's very, very slow. Other things you might want to do, um, if you want to do some training

that requires about 50 GB of RAM for some models, if you want to just add some more images for the database for it to

figure out, if you want to ask it about different stuff—they both work on both systems. And like I said, it's going to be

slightly faster if you are memory constrained on your system. Other situations like unroll engine or, um,

Blender render test performance, all that kind of stuff; no difference whatsoever because those applications are memory-constrained. What you want to do is you

want to go into Activity Monitor and you want to see your memory usage over here. That green bar, if you see it in the

red, that's when you might get performance—well, you probably will get performance improvements by going for a

better spec system. Whereas if it's mostly in the green, which most of the time it is, you're probably not going to get

any improvements. And if you're the kind of guy that maybe has a lot of multitasking applications running at the

same time—if you've got virtual machines running, hogging that piece of memory; you've got Chrome hogging that piece of memory;

Final Cut Pro—basically, if you don't want to close your applications, then maybe having a 128 GB system might be kind

of useful. But for now, let's jump into the test and show you the kind of impact you get and the difference it makes. I've got

a suite of different applications to show you if there's going to be a difference in your use case: virtual

machines, all that kind of stuff. But I'm going to start the show with machine learning, AI—Apple intelligence,

intelligence, that kind of stuff like that. So I've got this—this one here is a text-to-video generator. It's going to

use models in the machine learning world to get a bit of text and turn it into a piece of video. And look what

happens when I try to run the conversion. I'm going to convert—go ahead, start on 64. 64 is go ahead and starting to run,

and on 128 it's running as well. And if you look at the terminal right here, they're both proceeding along. I'm not

going to get into the detailed monitor yet because something's going to happen on 128 when it starts to load that last model: boom!

Right there, Python has quit unexpectedly, yet the 64 GB version is running successfully. So there's some issue in

the 128 GB version. Same piece of code, just the hardware is slightly different. It says right here that it's crashing in the

Metal Performance Shaders because the array has a 2x32 issue. And if you type that code into Google,

you'll find that this bug has been around for two years now, since 2022. It hasn't been fixed in two years. There is a workaround,

so we're actually going to get some performance metrics going. There is a workaround to get it going, but yeah, it

hasn't been fixed in two years. So if you are getting a 128 GB [device], one of the issues you might run

into is that some applications might have bugs for this reason, and the bugs might not be fixed because most people out

there don't have 128 GB, so they're not going to fix it. You have to be the person pioneering out there, fixing it. Knock on

Apple's door! Apple, please fix it! Yeah, so let's get this running over here. I'm going to just stop there. So memory

wise, you can see on the left here, on the 128 GB version, it's already using 96 GB of

memory on this application, and on the 64 GB version, it's using 50.9 GB. So the 64

GB RAM version is on step number two out of 20 with rendering, out of 20 frames, whereas the 128 GB version is already on step

number three, and the iterations per second here, it's now dropped to 31 seconds per iteration, so 30, whereas on

the 128 GB version it's 22 seconds. So that's about a 50% slowdown. And if you

keep it running, what you're going to find is this test is going to finish, and it's going to take the 64 GB RAM version

21 minutes to finish, and the 128 GB RAM version is only going to take 10, and that is Activity Monitor is our friend.

Here, I'll just make it slightly bigger on the 64 GB RAM version because there isn't enough memory to fit all of the

required data in the memory. It's having to thrash out the memory between the actual RAM and the SSD, so that thrashing

that has to go around loading and unloading data is causing the GPU to not work at 100%. So you can see on the left here,

128, it's just rocking the GPU just non-stop—boom, boom, boom, boom, boom—whereas on the 64 GB RAM, because of all

that thrashing from the memory, it's having to—it's not going at 100% speed, and that causes that bottleneck

and that causes the slowdown. But on the positive, what you find is the 64 GB RAM version actually works; it's slower for

this intense use case—it's slower, yeah, half the speed—but you know, if you want to test it out, it still works. And

something you can do is you don't need, if you're working on these systems, unless maybe your Mac's for life

and you don't want to go anywhere else, you don't need to render everything at full resolution. So, for example, if I

reduce the resolution of the output image—if I make it 512 by 256—so on the right here, on the 64 GB one, it takes 3.5

seconds per iteration; on the left here, it's also 3.5 seconds per iteration. So when you reduce the extra required RAM,

it is the same speed. So it's only if you're at a bottleneck. Now, these situations may compound if you're doing

multitasking. So, for example, if you've got a virtual machine running in the background that hogs memory, if you want

to run a bit of Chrome that's going to take another 10 GB of memory—all these browsers, all these—super

Running lots of applications at the same time, you're going to get a bit of a hit with the 128GB, but the only question

you need to ask yourself is: is this M4 system the system to actually go for? Because, for example, if I just log into my Windows PC

my Windows PC with the Nvidia graphics card, you see an activity monitor here. We've only got a 10GB graphics card,

Nvidia, it's 2 years old, a 3080, and we've only got 16GB of memory. Yet this

computer—not this one right here, the one I'm remoting into—can actually perform this conversion about twice as fast as

the M4 Max. So maybe, maybe spending an extra lump on the memory for, you know, machine learning kind of stuff,

maybe you're better off just getting a Windows Nvidia solution for this stuff. But if you are Mac for life, then I understand.

And there's a lot of headroom to get a lot of performance out of these systems. For example, while this is being run on

the GPU on the PC, on the actual M1, it's running on a GPU, but via a CPU

rendering solution. It is approved by Apple—Apple is the one that actually released it and got it going—but

there is a lot of headroom to get the algorithms running faster potentially. And the great thing about Mac is that it

uses unified memory. So whereas on the PC you'd be limited to getting 24GB on an Nvidia card unless you're kind of

like Elon Musk level rich and you can afford the proper, the top-of-the-line. You've got to wait like two years for an H100 anyway. If

you're a consumer, you're probably limited to 24GB on the graphics card. So maybe for some cases, some programmers

out there, maybe you guys out there could make things run a lot faster out there in the world. So potentially there is a

a lot of movement to get the stuff working better on Mac, and if you do, please let me know, cuz I'd love to test systems and

work with you guys. That'll be amazing. But I'm hoping that you guys are getting a sense of how much the performance can

improve between the 64 and 128 when you switch it up. So I'm going to go back

into Activity Monitor up here. We can see everything's finished, and it took 95 seconds on a test on both systems, and

that's because we reduced the resolution of the video that we made at the end. And if we just look at the output

result, you can see that they're all exactly the same. So the results are exactly the same right there; the

difference you're going to be making here is if you pump up the resolution to, let's say, 1080p. So we've got what we

got right now is when we're trying to run this maximum texture version, just minimally maximum, slightly maximum; we're

using 6.59 GB of RAM, but we're also using 39 GB of swap. So it's

allocated a piece of the SSD, and it's just constantly writing in and out for that memory, and it's 38 GB. And if we

128GB vs 64GB M4 Max MacBook Pro | Ultimate AI RAM Performance REVIEW

Xcreate

128GB와 64GB MacBook Pro M4 Max 비교 및 RAM 성능 검토 소개

128GB와 64GB 버전 MacBook Pro M4 Max 비교 리뷰 시작

128GB 필요성에 대한 시청자 의견 요청

미래 대비 또는 단순 구매 욕구에 대한 질문 제기

LLaMA 4050억 파라미터 버전 실행 결과 예고

64GB에서는 LLaMA 실행 불가, 128GB에서는 양자화 버전으로 가능하다는 설명

다양한 AI 작업 환경에서의 RAM 용량 비교

개인 AI 및 데이터 프라이버시 관련 언급

텍스트-비디오 변환 작업에서 64GB와 128GB의 차이점 설명

128GB 버전의 Metal Performance Shaders 버그 문제 언급

128GB 버전이 64GB 버전보다 2배 이상 빠른 작업 완료 시간 비교

50GB RAM 필요 모델 트레이닝, 이미지 데이터베이스 추가 등의 작업 비교

메모리 사용량과 성능 향상의 상관관계 분석

Activity Monitor를 이용한 메모리 사용량 확인 방법 설명

메모리 사용량이 높을 때 성능 향상 가능성 언급

메모리 사용량이 낮을 때 성능 향상의 어려움 언급

멀티태스킹 환경에서 128GB 시스템의 유용성 설명

실제 애플리케이션 테스트 및 성능 비교

다양한 애플리케이션을 이용한 성능 비교 테스트 계획 설명

머신러닝 및 AI 애플리케이션을 이용한 테스트 시작

텍스트-비디오 생성기를 이용한 테스트 진행 및 결과 분석

메모리 사용량 비교: 128GB는 96GB, 64GB는 50.9GB 사용

해상도를 낮추면 64GB와 128GB의 속도 차이가 없다는 점 언급

멀티태스킹 환경에서의 RAM 용량 제한 문제점 언급

M4 시스템 대비 Windows PC의 성능 비교

맥 사용자를 위한 128GB RAM의 장점과 단점 비교

LLaMA 3.1 4050억 파라미터 모델 실행 테스트

LLaMA 3.1 4050억 파라미터 모델 실행 가능성 테스트

64GB 버전에서 메모리 부족으로 인한 크래시 발생

128GB 버전에서도 스왑 메모리 사용량 증가 및 크래시 발생

양자화 버전을 이용한 4050억 파라미터 모델 실행 결과

128GB 버전에서만 가능한 작업에 대한 결론

일반적인 사용 환경에서는 RAM 용량 차이가 크지 않다는 점 강조

3년간 64GB RAM 사용 경험을 통한 장점 설명

장기간 사용을 위한 RAM 용량 확보의 중요성 강조

향후 CPU 및 GPU 성능 향상에 대한 기대감 언급

시청자의 의견 및 사용 사례 공유 요청