1
00:00:00,000 --> 00:00:07,640
I've always had a problem comparing Apple Silicon to PC parts because gaming wasn't

2
00:00:07,640 --> 00:00:15,240
ready. I've always wanted to know how Apple Silicon compares to off-the-shelf PC GPUs and CPUs.

3
00:00:15,240 --> 00:00:19,360
And that's not something we've been able to test because Macs and PCs now use completely

4
00:00:19,360 --> 00:00:25,480
different architectures, and there were no games that ran natively on both.

5
00:00:25,480 --> 00:00:33,640
But that's changed. Check this out. I'm playing Baldur's Gate 3, one of the hottest video games right now, on a Mac.

6
00:00:33,640 --> 00:00:40,360
And it's running natively on Apple Silicon. There's now a growing list of games that support both Apple's graphics hardware and

7
00:00:40,360 --> 00:00:44,200
metal graphics APIs without any sort of translation layer.

8
00:00:44,200 --> 00:00:48,760
So with the help of LMG Labs, we're going to finally compare Apple chips with PC graphics

9
00:00:48,760 --> 00:00:54,680
cards and see what GPUs from NVIDIA and AMD compete with Apple's M series.

10
00:00:54,680 --> 00:01:00,160
Though not in this video. That's part two because in part one, we need to build the test benches that the GPUs will

11
00:01:00,160 --> 00:01:05,400
be on. And those test benches have CPUs, so we're going to figure out what CPUs are the closest

12
00:01:05,400 --> 00:01:13,840
to Apple Silicon. And I need to ask Labs if they'll help me.

13
00:01:13,840 --> 00:01:16,840
What did I think? Oh my god. Well, it's going to be interesting.

14
00:01:16,840 --> 00:01:22,800
It's going to be a learning experience. We haven't tested Macs really all that much, so this is going to be the first time Labs

15
00:01:22,800 --> 00:01:27,040
really tackles Macs. Nicholas, hey.

16
00:01:27,040 --> 00:01:30,520
This is Nicholas Harris. He's LTT Labs software developer.

17
00:01:30,520 --> 00:01:36,880
Part of his job has been developing and automating the tests for PCs and other parts here.

18
00:01:36,880 --> 00:01:42,960
The first step is figuring out the tests. And so Nicholas worked with test technician John Deuren to figure out how to measure these

19
00:01:42,960 --> 00:01:51,320
two completely different computers. Our primary goal was to find tests that can natively work on either system to avoid the

20
00:01:51,320 --> 00:01:55,840
Rosetta layer, because that's another variable that we want to isolate for.

21
00:01:55,840 --> 00:02:00,280
But we're also isolating for the CPU that also limits us, because we do have tests that

22
00:02:00,280 --> 00:02:05,440
test the whole system, but we're not looking to test the memory and the SSD and the graphics

23
00:02:05,440 --> 00:02:12,000
card yet. But our current test suite for Markbench is very Windows focused, so there wasn't really

24
00:02:12,000 --> 00:02:18,520
anything we could just reuse from that. There is a test framework out there called Pheronix, which has been there out there for

25
00:02:18,520 --> 00:02:26,200
a long time. So we tried to find some that did compression, stuff that did maybe some encoding, things

26
00:02:26,200 --> 00:02:30,000
that were just pure computational. Okay, so I have to confess something.

27
00:02:30,000 --> 00:02:35,160
I had this idea, fantasy really, that Labs would test a bunch of Macs and then compare

28
00:02:35,160 --> 00:02:39,960
it to a matrix of CPU data that would show us what desktop and laptop CPUs perfectly

29
00:02:39,960 --> 00:02:46,480
match their M-series counterparts. But we have to use new tests, so we're starting from scratch.

30
00:02:46,480 --> 00:02:49,720
While we would have loved to investigate every CPU, it is wholly unreasonable to get

31
00:02:49,720 --> 00:02:56,360
Labs to test them all. For instance, here in Logistics, there are about 150 different CPUs available to test,

32
00:02:56,360 --> 00:03:00,120
and that doesn't even include the laptop CPUs and all the shapes and sizes they're

33
00:03:00,120 --> 00:03:06,440
cooled in, either. So we have to make tradeoffs, and this is about graphics cards, and so that's why we're

34
00:03:06,440 --> 00:03:14,200
only including desktop components. If you're after gaming performance, the question then becomes like, okay, cool, we have three

35
00:03:14,200 --> 00:03:19,840
contenders, right? Apple, Intel, and AMD. Intel changes their socket all the time.

36
00:03:19,840 --> 00:03:24,760
But AMD, you can go back three generations on AM4 on the same platform.

37
00:03:24,760 --> 00:03:28,760
And so that's why we're going to be sticking with AM4 chips in our conclusion, though we

38
00:03:28,760 --> 00:03:32,280
did run the tests on a few Intel chips earlier in this project.

39
00:03:32,280 --> 00:03:36,240
Alright, so let's go through the tests and the results.

40
00:03:36,240 --> 00:03:40,160
The first we did is, of course, Cinebench, it's widely used in the tech media space

41
00:03:40,160 --> 00:03:43,960
and they just came out with an update for it, though we did R23.

42
00:03:43,960 --> 00:03:47,400
It includes both a multi-core and single-core score.

43
00:03:47,400 --> 00:03:53,200
We chose Cinebench because it kind of chooses itself as it's the prolific go-to processor

44
00:03:53,200 --> 00:03:59,400
benchmark, and it's really good that it supports Apple natively as well as Windows.

45
00:03:59,400 --> 00:04:03,200
Looking at the single-core results, you can see how the newest chips rise to the top,

46
00:04:03,200 --> 00:04:08,200
but once all the cores get involved, you can see just how the 24 and the M2 Ultra push

47
00:04:08,200 --> 00:04:14,080
it to the top. Along with single-core tests, Lab did a Flak encode test where they encoded a bunch of

48
00:04:14,080 --> 00:04:17,600
copies of a 9-inch nail song from Wave to Flak.

49
00:04:17,600 --> 00:04:21,920
We actually struggled to find single-core tests because most tests are all about loading

50
00:04:21,920 --> 00:04:26,380
the CPU and trying to, you know, how fast you compute this thing.

51
00:04:26,380 --> 00:04:31,640
In this test, and another you'll see, Apple Silicon is so far ahead of the other chips,

52
00:04:31,640 --> 00:04:37,920
and they're all grouped together. That's why we're going to be weighing this test less when we figure out our matches.

53
00:04:37,920 --> 00:04:43,160
The last single-core tests are the XZ and LZ4 compression tests, with both compressing

54
00:04:43,160 --> 00:04:49,640
an Ubuntu image. We actually tried like four different compression algorithms, or compression tests, but not

55
00:04:49,640 --> 00:04:55,380
all of them worked. Sometimes they worked on one, but not the other, even though they're advertised for cross-platform.

56
00:04:55,380 --> 00:05:00,960
So we did find that XZ and LZ4, we were able to compile for both natively.

57
00:05:00,960 --> 00:05:06,080
LZ4 single-core shakes out slightly differently from Cinebench, with most of the Ryzen 5,000s

58
00:05:06,080 --> 00:05:13,040
closer to the M3 generation. But it appears that with the XZ compression test, Ryzen has a bit more strength than it

59
00:05:13,040 --> 00:05:18,360
does in Cinebench. What was it like to do all the testing?

60
00:05:18,360 --> 00:05:25,280
Illuminating. Testing's pretty straightforward. Once you identify the tests and you come up with your test suite, the execution is just

61
00:05:25,280 --> 00:05:31,200
while you do your testing. And now as results come in, it's like putting your puzzle together, right?

62
00:05:31,200 --> 00:05:34,640
As the pieces slowly fit in more, you get more of the picture.

63
00:05:34,760 --> 00:05:38,480
However, the difference is you're completing a puzzle that doesn't, you don't know what

64
00:05:38,480 --> 00:05:43,920
the end picture is. So it's interesting that way to kind of see the story reveal itself to you.

65
00:05:43,920 --> 00:05:48,360
Alright, so how about the puzzle pieces that tested multiple CPU cores?

66
00:05:48,360 --> 00:05:53,120
Blender is a popular 3D modeling program, and in it, Labs rendered the barbershop scene.

67
00:05:53,120 --> 00:05:56,240
We might need to find a new scene as it can render fairly quickly.

68
00:05:56,240 --> 00:06:00,240
It's a popular scene used to benchmark rendering performance in Blender.

69
00:06:00,240 --> 00:06:04,960
In Blender, the mid-range AMDs provide a transition between the M2 and M3s.

70
00:06:04,960 --> 00:06:08,720
The 5600X and G are surprisingly weaker here.

71
00:06:08,720 --> 00:06:11,920
Another render type test that I'd never heard of is C-Ray.

72
00:06:11,920 --> 00:06:15,720
It's a simple ray tracer that outputs this 90s looking image.

73
00:06:15,720 --> 00:06:23,080
C-Ray gives us another render type of test, but mostly we chose it because it works on

74
00:06:23,080 --> 00:06:27,040
both. That's definitely got for some of these.

75
00:06:27,760 --> 00:06:32,320
It's a very simple, efficient load to multi-core.

76
00:06:32,320 --> 00:06:37,480
AMD is relatively weaker in this test with the 3600 sitting between the M1s and their

77
00:06:37,480 --> 00:06:44,560
different cooling. The next test Labs did is LibRAW, which tests how well CPUs handle raw photographs.

78
00:06:44,560 --> 00:06:49,680
LibRAW is also nice that it has a built-in post-processing benchmark, which we run 30

79
00:06:49,680 --> 00:06:54,480
times on the test image that comes with Fronix, and then it spits out like a megapixels per

80
00:06:54,480 --> 00:06:59,080
second. LibRAW is the other test we're going to have to weigh less because of Apple Silicon's apparent

81
00:06:59,080 --> 00:07:04,760
supremacy. It does feel like the Macs are especially tailored to calculate audio and visual codecs.

82
00:07:04,760 --> 00:07:08,440
And lastly, if you're into numbers, there's PrimeSiv.

83
00:07:08,440 --> 00:07:12,920
We chose it because Y Cruncher doesn't work natively on Mac, because we do favor Y Cruncher.

84
00:07:12,920 --> 00:07:19,160
It's a very popular benchmarking one, but we found PrimeSiv, PrimeSiv, PrimeSiv, PrimeCV,

85
00:07:19,160 --> 00:07:23,200
PrimeSciEV, it calculates Prime numbers up to a certain length.

86
00:07:23,200 --> 00:07:29,440
So we consider that it's our stand-in for Y Cruncher and that it's something computational,

87
00:07:29,440 --> 00:07:32,880
generating number over a long period.

88
00:07:32,880 --> 00:07:38,040
It's multi-core as well. We learned a lot by doing this, and that is that this is hard.

89
00:07:38,040 --> 00:07:42,720
For one, picking AM4 means that we've got an array of chips that don't quite fit with

90
00:07:42,720 --> 00:07:47,720
single-core performance, as that's where Apple Silicon shines, and these are old.

91
00:07:47,720 --> 00:07:52,080
But then, with AM5, there aren't any low-end chips to compare with the lower-end Mac chips

92
00:07:52,080 --> 00:07:57,280
either. Single-core is more important for gaming, so we waited higher, but we waited Flak and

93
00:07:57,280 --> 00:08:01,800
Libra less because they favor Apple Silicon egregiously in a way that's not related to

94
00:08:01,800 --> 00:08:09,080
gaming. Alright, our picks. These choices for CPUs are still even a best guess, because we don't know what's the bottleneck

95
00:08:09,080 --> 00:08:13,320
CPU-wise once the GPUs are installed in running games.

96
00:08:13,320 --> 00:08:18,480
So these are not exact matches. I really was in Fantasyland thinking this was possible.

97
00:08:18,480 --> 00:08:25,080
But we've learned a lot. We're going to use the AMD 5800X3D as a control and to match the M2 Ultra, because

98
00:08:25,080 --> 00:08:30,280
its 3D cache really helps in gaming, and the M2 Ultra screamed well ahead in every test

99
00:08:30,280 --> 00:08:36,840
we threw at it. The M2 Pro and Max, as well as the M3 Pro chips, will be matched against the 5800X.

100
00:08:36,840 --> 00:08:42,960
The 5700X were pitting alongside the basic M3, and the basic M2 and M1 chips are matched

101
00:08:42,960 --> 00:08:47,200
against a 5600G. I'm feeling as well as I could.

102
00:08:47,920 --> 00:09:02,240
And that, I guess, is the biggest lesson on this journey.

103
00:09:02,240 --> 00:09:08,680
And it's that we're always learning. But now that we have our CPUs figured out, the next step will be to test the GPUs.

104
00:09:08,680 --> 00:09:12,960
Have NVIDIA and AMD met their match? We're going to have to see where they line up.

105
00:09:12,960 --> 00:09:21,080
But things are looking good in their own little way. Personally, I wasn't expecting the Mac, the Apples, to be as strong as they were.

106
00:09:21,080 --> 00:09:25,360
I knew they were super efficient, so this is my first really experience.

107
00:09:25,360 --> 00:09:31,800
Now during this project, I started daily driving the 15-inch MacBook Air, and I loved it.

108
00:09:31,800 --> 00:09:37,640
The fact that I could close it, neglect it for three days, and it still had power.

109
00:09:37,640 --> 00:09:40,800
I mean, I ended up buying one. What?

110
00:09:40,800 --> 00:09:44,240
You bought a Mac from this project? I mean, I did immediately stick or bomb it.

111
00:09:44,240 --> 00:09:48,840
How dare you? Thanks for testing this Mac Address, Labs.

112
00:09:48,840 --> 00:09:53,520
If you want to check out another video we did, check out the iPad tier list video.

113
00:09:53,520 --> 00:09:57,680
And I'm curious in the comments below, who of you are like Nicholas and bought a Mac

114
00:09:57,680 --> 00:09:58,400
for gaming?