1
00:00:00,160 --> 00:00:06,960
You guys just saw our income breakdown, so you know that unless you buy an awful

2
00:00:04,319 --> 00:00:11,920
lot of Framework hoodies, I am not making back what I spent on this brand

3
00:00:09,200 --> 00:00:17,359
new M3 Ultra Max Studio. So, the question is, how can we possibly justify

4
00:00:14,960 --> 00:00:22,960
buying it? And the answer may surprise you. You see, depending on what you're

5
00:00:19,680 --> 00:00:26,720
using it for, the M3 Ultra Studio is

6
00:00:22,960 --> 00:00:28,240
actually, dare I say it, a good deal.

7
00:00:26,720 --> 00:00:33,600
Not for everyone, of course. If all you're doing is some light photo editing, you're going to be better off

8
00:00:30,640 --> 00:00:37,600
doing that on your phone. But if you work with AI, for instance, or if you

9
00:00:36,000 --> 00:00:42,719
have heavy demands when it comes to content creation, then even with our

10
00:00:40,559 --> 00:00:48,399
configuration, which has the topsp spec GPU and cost us about

11
00:00:46,039 --> 00:00:54,399
$10,000, but this little puppy could end up making you a lot more than you spent.

12
00:00:52,079 --> 00:00:58,719
Let's take a look after, of course, we uh look at the accessories. There's a

13
00:00:56,480 --> 00:00:58,719
power

14
00:01:00,760 --> 00:01:09,360
cable and paper. And that's it. So, after that

15
00:01:07,760 --> 00:01:15,360
very short side quest, let's take a closer look at the machine itself. Starting at the front, we've got an SDXC

16
00:01:12,799 --> 00:01:19,360
reader as well as two Thunderbolt ports. And these are one of the big

17
00:01:17,200 --> 00:01:23,360
differentiators for the new Mac Studio. Whether we're talking about the topsp

18
00:01:20,960 --> 00:01:30,000
spec M3 Ultra version that we've got here or the lesser

19
00:01:26,439 --> 00:01:32,960
M4 Max version that has faster cores,

20
00:01:30,000 --> 00:01:36,560
but not nearly as many of them. And the reason that these are such a big deal is

21
00:01:34,479 --> 00:01:41,439
because now they are Thunderbolt 5, which even under normal circumstances

22
00:01:39,200 --> 00:01:47,439
offers up to double the performance of last gen Thunderbolt 4 because it

23
00:01:43,759 --> 00:01:49,840
transitions from PCIe Gen 3 to PCIe Gen

24
00:01:47,439 --> 00:01:54,200
4. That means more bandwidth for things like connected SSDs or card readers or

25
00:01:52,880 --> 00:02:00,399
external GPUs, at least on non-MAC machines. But

26
00:01:58,000 --> 00:02:05,680
it also has another trick up its sleeve. Instead of operating at 80 Gbit 80 Gbit

27
00:02:03,680 --> 00:02:12,400
transmission and receive, it can actually stack it and do up to 120 GB

28
00:02:09,119 --> 00:02:14,800
transmission and 40 GB receive in boost

29
00:02:12,400 --> 00:02:20,239
mode. This allows the new Mac Studio with its Thunderbolt 5 ports to drive

30
00:02:17,520 --> 00:02:25,760
extremely highresolution, high refresh rate displays, and a lot of them. Apple

31
00:02:23,360 --> 00:02:29,360
boasts it'll do up to eight connected displays in the ultra configuration.

32
00:02:28,000 --> 00:02:34,800
Now, let's take a look at the back where we find, that's right, another four of

33
00:02:33,120 --> 00:02:38,239
those Thunderbolt 5 ports, allowing you to connect pretty much whatever you

34
00:02:36,400 --> 00:02:41,760
could possibly want. remembering, of course, that Thunderbolt can daisy

35
00:02:39,680 --> 00:02:46,879
chain, so you can have up to six devices off of each of those six ports. We also

36
00:02:44,560 --> 00:02:51,920
get a 10 gig Ethernet connection, power input, and Mickey Mouse style, two USBA

37
00:02:49,840 --> 00:02:55,599
ports that operate at up to 5 GB per second, an HDMI port, as well as a

38
00:02:54,000 --> 00:03:00,959
headphone jack, and of course, conveniently located rear power button.

39
00:02:58,640 --> 00:03:04,640
We also get a nice beefy looking exhaust to go along with our nice beefy intake

40
00:03:03,120 --> 00:03:08,879
on the bottom. I haven't heard it in person, though I will very shortly. The

41
00:03:06,239 --> 00:03:13,760
Labs tells me is pretty darn quiet. A Kensington lock for the bottom. And

42
00:03:11,360 --> 00:03:18,720
that's about it. It's just a big block of aluminum. A big block of aluminum

43
00:03:16,239 --> 00:03:23,040
full of some very impressive specs. Let's talk speeds and feeds while I get

44
00:03:20,560 --> 00:03:28,480
this thing set up. The M3 Ultra SoC in here is equipped with up to 32 CPU

45
00:03:25,680 --> 00:03:33,360
cores. That is eight more than Apple's last gen flagship. But perhaps more

46
00:03:31,040 --> 00:03:37,440
importantly than that, all eight of the new cores are performance rather than

47
00:03:35,920 --> 00:03:42,959
efficiency cores, meaning this thing should be a multi-threaded beast. In

48
00:03:40,239 --> 00:03:47,200
addition to that, we get a new 32 core neural engine that is apparently much

49
00:03:44,879 --> 00:03:52,879
faster than last gen. And on top of that, we get up to an 80 core GPU with

50
00:03:49,920 --> 00:03:57,280
up to 800 gigabytes per second of combined memory bandwidth that is shared

51
00:03:55,040 --> 00:04:02,680
across all of those processors. This was a really key Apple innovation in the

52
00:03:59,040 --> 00:04:05,920
last few years, their unified memory. Of

53
00:04:02,680 --> 00:04:09,040
course, fast memory is great, but when

54
00:04:05,920 --> 00:04:11,680
it comes to AI, more memory is in some

55
00:04:09,040 --> 00:04:15,519
cases more important, and that is where this thing is really appealing to AI

56
00:04:13,680 --> 00:04:22,320
developers. It can be equipped with up to 512 GB of LPDDR5 memory. That means

57
00:04:20,400 --> 00:04:29,040
that compared to last gen, which topped out at 192, we can run much larger AI

58
00:04:26,800 --> 00:04:33,520
models. And that ignores the fact that the hardware is faster, allowing us to

59
00:04:31,280 --> 00:04:37,440
run them in a more performant manner. Of course, it's not all for AI. Apple

60
00:04:35,520 --> 00:04:42,560
loaded the M3 Ultra up with a new media engine that they claim supports up to 24

61
00:04:40,080 --> 00:04:47,440
streams of ProRes playback in addition to supporting hardware encode and decode

62
00:04:45,120 --> 00:04:53,919
of many of the most popular formats, including AV1 decoding. Note that I said

63
00:04:50,400 --> 00:04:56,639
AV1 decoding, though. One major omission

64
00:04:53,919 --> 00:05:02,000
is that Apple's SOC's still do not support AV1 encoding. And this is true

65
00:04:59,040 --> 00:05:06,560
of both the M3 Ultra configuration as well as the M4 Max configuration. Either

66
00:05:05,199 --> 00:05:11,919
way, these machines are going to be Photoshop and Premier Powerhouses.

67
00:05:08,880 --> 00:05:15,759
Because even limited to CPUAV1 encoding,

68
00:05:11,919 --> 00:05:17,440
the M3 Ultra beat a Threadripper 7980X,

69
00:05:15,759 --> 00:05:22,720
a chip that consumes more power than this entire box and has twice as many

70
00:05:19,639 --> 00:05:24,080
cores. Pretty impressive. It's just

71
00:05:22,720 --> 00:05:28,240
worth noting that it doesn't support hardware encode, which is a shame when

72
00:05:25,759 --> 00:05:32,479
you look at the incredible performance of Apple's hardware encoder for H.264.

73
00:05:31,120 --> 00:05:37,039
Now, another thing that's worth noting is that Apple doesn't have a clear

74
00:05:34,960 --> 00:05:42,560
flagship with the Mac Studio anymore. I mean, they obviously have one that costs

75
00:05:38,560 --> 00:05:44,960
more, the M3 Ultra version, but the M4

76
00:05:42,560 --> 00:05:49,280
Max version may end up outperforming it in some cases. Now, we don't have an M4

77
00:05:46,800 --> 00:05:53,919
Max yet, but we do have an M4 Pro machine. And we found that in cases

78
00:05:51,280 --> 00:05:59,520
where singlethreaded CPU performance is more important than having a big beefy

79
00:05:56,080 --> 00:06:01,840
GPU or lots of CPU cores, it could end

80
00:05:59,520 --> 00:06:06,080
up being the more performant option. Of course, you'll also be limited to a

81
00:06:03,199 --> 00:06:10,560
maximum of 64 GB of memory. Now, before we go any further into performance,

82
00:06:07,840 --> 00:06:14,759
let's thank our sponsor, Dbrand. Uh, Dbrand told me they want you to go

83
00:06:11,759 --> 00:06:18,759
outside and touch grass.

84
00:06:14,759 --> 00:06:21,680
Okay, that shouldn't be that hard. I see

85
00:06:18,759 --> 00:06:26,720
some over there. I think there's some grass in there. There you go.

86
00:06:24,479 --> 00:06:31,199
Oh, sorry, boss man. I misread. They actually want you to touch grass on your

87
00:06:28,479 --> 00:06:36,720
Mac Studio. Oh, this. They actually told me they

88
00:06:34,400 --> 00:06:40,800
were working on this a while back. I didn't know they were ready to ship yet.

89
00:06:38,479 --> 00:06:47,600
That is really impressive. Apparently, every square centimeter contains a,000

90
00:06:44,960 --> 00:06:51,759
grass strands. Man, you can really feel the density of it. It's like having a

91
00:06:50,000 --> 00:06:56,479
putting green on the top of your Mac Studio. This is not an April Fool's

92
00:06:54,319 --> 00:06:59,680
joke. This is a real product. Well, like the Touch Grass skin is a real product.

93
00:06:58,400 --> 00:07:06,319
They don't make skins for the Mac Studio, so get wrecked. I guess Dbrand has

94
00:07:04,400 --> 00:07:09,599
fantastic guides on their site for how to apply their skins, but I don't give a

95
00:07:08,319 --> 00:07:13,759
if those guys want me to follow their guide, they're going to have to pay me more and be a little

96
00:07:17,319 --> 00:07:23,520
nicer. There you have it, folks. This is as real of grass that you can stick to

97
00:07:21,599 --> 00:07:28,039
your phone. It's ultra limited edition. Once it's all sold out, it's sold out.

98
00:07:25,440 --> 00:07:30,960
You can get it now at shortlininus.com. All right, let's talk

99
00:07:29,599 --> 00:07:34,800
a little bit more about the Mac Studio, shall we? I want to fire up some kind of

100
00:07:33,039 --> 00:07:38,880
a load and see how loud this thing gets for myself. Let's go with Blender. We're

101
00:07:36,720 --> 00:07:42,000
going to do the Barberhop render. I'm going to see if I can hear this

102
00:07:44,360 --> 00:07:52,479
thing. I mean, surely it's got to ramp up. This is CPU mode, too. This is with

103
00:07:49,759 --> 00:07:56,240
all the CPU cores going full tilt. I can barely hear it. And I'm pretty much the

104
00:07:54,560 --> 00:07:59,840
biggest diva there is when it comes to computer noise. Like to the point where

105
00:07:57,680 --> 00:08:04,639
I put my computer in a separate room, so I don't have to hear it. If I could game

106
00:08:02,720 --> 00:08:12,080
on a Mac Studio, I probably wouldn't even bother. To be clear, I can game on

107
00:08:08,560 --> 00:08:13,759
a Mac Studio. Uh the performance is fine

108
00:08:12,080 --> 00:08:17,919
whether we're looking at games that rely on Rosetta or whether we're looking at

109
00:08:16,000 --> 00:08:22,879
games that run natively on Apple Silicon.

110
00:08:19,759 --> 00:08:25,840
But what it isn't is a tremendous value

111
00:08:22,879 --> 00:08:30,720
and the compatibility is still a major struggle for the macOS platform

112
00:08:27,680 --> 00:08:33,120
compared to Windows or even Linux these

113
00:08:30,720 --> 00:08:37,039
days. With that said, I still feel that GPU benchmarks are relevant because they

114
00:08:35,360 --> 00:08:41,360
give us some idea of what Apple has improved from generation to generation.

115
00:08:39,279 --> 00:08:45,680
And at least when it comes to gaming, it looks like not much. Most of this

116
00:08:44,240 --> 00:08:51,560
performance difference compared to the M2 Ultra could be explained by going

117
00:08:47,680 --> 00:08:54,000
from 76 to 80 GPU cores in the maximum

118
00:08:51,560 --> 00:08:58,640
configuration. But what's curious is that if I wasn't running this in CPU

119
00:08:56,399 --> 00:09:06,360
mode, if I was running Blender in GPU mode, it is much faster in GPU Blender

120
00:09:03,320 --> 00:09:09,760
rendering. And the same holds true in

121
00:09:06,360 --> 00:09:12,399
AI. Actually, in AI, performance is up

122
00:09:09,760 --> 00:09:18,320
across the board. So depending on the application, it'll use either the neural

123
00:09:15,120 --> 00:09:20,800
processor, the GPU, or the CPU,

124
00:09:18,320 --> 00:09:26,399
typically in that order in order to get the best performance out of the M3 Ultra

125
00:09:23,040 --> 00:09:29,920
SOC. And we found that across the board,

126
00:09:26,399 --> 00:09:32,560
the M3 Ultra is much faster than the M2

127
00:09:29,920 --> 00:09:36,640
Ultra in AI. And that's in addition to being able to run much larger models.

128
00:09:34,160 --> 00:09:41,600
So, this was the main justification that Nick, and I'm naming him, Nick from the

129
00:09:39,040 --> 00:09:48,880
lab gave me for us buying this machine because we actually use our 192 GB M2

130
00:09:45,360 --> 00:09:50,560
Ultra Studio daily to run local LLMs

131
00:09:48,880 --> 00:09:56,320
that help with programming for our coders. He says, and he promises me that

132
00:09:54,080 --> 00:10:02,640
this machine will justify itself because we'll be able to run better LLMs. Is

133
00:09:58,560 --> 00:10:03,920
that right, Nick? Yep. Okay. Oh, wow.

134
00:10:02,640 --> 00:10:08,480
I'm just looking a little closer at those gaming results. This thinging

135
00:10:05,920 --> 00:10:13,360
sucks for gaming. This is like on par with a 4060 Ti. The way it kind of keeps

136
00:10:11,200 --> 00:10:19,680
up in Blender, though, is so bizarre to me. Like, yeah, it's slower than a 7980X

137
00:10:16,560 --> 00:10:23,040
with an RTX 5090, but like not by that

138
00:10:19,680 --> 00:10:25,120
much, you know? Like a 4060 Ti would get

139
00:10:23,040 --> 00:10:30,880
crushed by that configuration, and this doesn't. What a strange strange machine.

140
00:10:28,560 --> 00:10:38,399
And boy, will you ever pay for it. The M4 Max configuration starts at $2,000.

141
00:10:34,640 --> 00:10:41,440
And the M3 Ultra starts at $4,000, going

142
00:10:38,399 --> 00:10:44,320
up as high as about $14,000 if you max

143
00:10:41,440 --> 00:10:49,120
out both the RAM like we did, and the storage, which we did not. A terabyte is

144
00:10:47,839 --> 00:10:53,600
fine when you're going to be accessing most of your storage over the network.

145
00:10:51,680 --> 00:10:58,240
Oh, with that said, could we maybe try to do a storage upgrade for an LT DIY? I

146
00:10:57,200 --> 00:11:04,160
don't know if there's a way to do that on this one yet. We'll have to figure that out because of course like other

147
00:11:01,839 --> 00:11:08,720
Apple silicon machines, Apple does not officially sanction any kind of storage

148
00:11:06,800 --> 00:11:12,800
or memory upgrade after the fact. You got to buy it and that's what you get.

149
00:11:11,040 --> 00:11:19,040
And actually, for that matter, I don't think there's any official way to upgrade like the the Wi-Fi chip in it

150
00:11:15,839 --> 00:11:20,640
either, is there? Which is already kind

151
00:11:19,040 --> 00:11:26,720
of an issue because it only ships with Wi-Fi 6E. So, you'd have to get like a

152
00:11:23,360 --> 00:11:30,040
Thunderbolt to PCIe to Wi-Fi 7 adapter

153
00:11:26,720 --> 00:11:32,279
card solution if you wanted faster

154
00:11:30,040 --> 00:11:37,000
Wi-Fi. Oh, Apple. Subscribe to ShortCircuit.