I’ve had my Macbook Pro M5 Max now for a couple of days and thought I would share some initial thoughts with regards to the positives and negatives. Would be interested to know how it compares with others and hopefully at least one person find it useful if they’re weiging up an upgrade now or waiting for an as of yet unannounced alternative.
Spec-wise I have the top end M5 Max 18 core cpu and 40 core GPU, 128GB RAM and 4TB ssd. My primary usage is for coding and data science. I do sequence modelling on it (e.g. human languages, DNA sequences). I also have a couple of side projects in the computer vision and 3d modelling space.
General thoughts - performance
First off, this thing is fast. I’m yet to notice a problem with thermal throttling under extreme load (some reviewers found the thermal treatment was dodgy in some early deliveries affecting performance on extensive benchmarks - although I’ve not actually run the same benchmarks yet so cannot rule it our). Compilation is fast, and while Xcode is still a steaming pile of 💩 it is at least snappy and large projects build in seconds (Xcode itself feels like it’s in perpetual beta, and each update breaks as much as it fixes - but this is a software issue).
I don’t really game, but I loaded up Civilisation VI and it performs amazingly at the high res of the screen. The game is exceptionally poorly optimised and I kind of think it’s a shame machines are so powerful and developers are just able to get away with writing slow code, but from a user perspective at least we don’t need to put up with slow software.
General thoughts - hardware
Coming from a Macbook Air, and having previously used a touch bar intel Macbook Pro, the immediate thing to jump out is the quality of the speakers. It’s pleasant to listen to music on, bass is deep and mids clear even when the volume is pushed. When travelling I do sometimes use my laptop to watch videos and there will be a huge improvement here (almost worth the extra heft). The fan doesn’t come on often, but there can be some clicking when it engages and if I’m not playing music it is audible. When I work alone in the room I have music playing through the speakers and in a room with others I have headphones on. You won’t find it too distracting (as compared to a Lenovo I have for the odd windows task which sounds like a vacuum cleaner even when idle).
So far I couldn’t be happier. I need to travel to client offices a bit more to see how I get on with the larger mass of the laptop. I was close to getting the 14 but was really worried about thermal throttling, this is the unfortunate tradeoff and the big benefit is those lovely speakers.
AI - the good
The new hardware shaders (from the m4?) and neural accelerators in the M5 are a huge step up from those on my M2. Local LLM performance can be really fast, certainly sufficient for a lot of the tasks I’m working on at the moment (for example, creating knowledge graphs from a collection of academic papers). The memory bandwidth is really good and the large amount of memory allows several big models to be loaded into memory concurrently.
Image classification, depth mapping, pose detection, segmentation are all blazingly fast. Opens up some really useful new possibilities.
I’ve a few personal projects that utilise Apple intelligence and I notice a hige improvement running on the new machine compared to, say, my old M2 Air and my M4 ipad (the latter has improved shader performance so for very small models is much faster than my M2 Air).
The SSD speed make a huge difference, especially when loading in large models.
AI - the bad
The hardware shaders don’t support fp8. For image/video generation this is a bit of a pain as a lot of new models are distributed in fp8 only (some enterprising individuals have written scripts to convert them back to fp16 so that they will run, however the loss when quantised isn’t reversed, you just end up with a slower/larger model). Fingers crossed Apple adds fp8 support in the M6 - this will be enough to get me to upgrade again within a year. If you want a good example of how frustrating this can be, download latest comfy ui desktop and try running some of the supplied templates. I’ve run a few pytorch and MLX based models in fp16 and they are much faster than before, but this is still an area where Apple is lagging behind nvidia. For some AI workloads unfortunately Nvidia GPU still really a hard requirement (I just spin up an AWS VM for these cases, but if you regularly need it then don’t spend the money on a Max).
How about other people? Have you bought one, made a decision to wait?
Spec-wise I have the top end M5 Max 18 core cpu and 40 core GPU, 128GB RAM and 4TB ssd. My primary usage is for coding and data science. I do sequence modelling on it (e.g. human languages, DNA sequences). I also have a couple of side projects in the computer vision and 3d modelling space.
General thoughts - performance
First off, this thing is fast. I’m yet to notice a problem with thermal throttling under extreme load (some reviewers found the thermal treatment was dodgy in some early deliveries affecting performance on extensive benchmarks - although I’ve not actually run the same benchmarks yet so cannot rule it our). Compilation is fast, and while Xcode is still a steaming pile of 💩 it is at least snappy and large projects build in seconds (Xcode itself feels like it’s in perpetual beta, and each update breaks as much as it fixes - but this is a software issue).
I don’t really game, but I loaded up Civilisation VI and it performs amazingly at the high res of the screen. The game is exceptionally poorly optimised and I kind of think it’s a shame machines are so powerful and developers are just able to get away with writing slow code, but from a user perspective at least we don’t need to put up with slow software.
General thoughts - hardware
Coming from a Macbook Air, and having previously used a touch bar intel Macbook Pro, the immediate thing to jump out is the quality of the speakers. It’s pleasant to listen to music on, bass is deep and mids clear even when the volume is pushed. When travelling I do sometimes use my laptop to watch videos and there will be a huge improvement here (almost worth the extra heft). The fan doesn’t come on often, but there can be some clicking when it engages and if I’m not playing music it is audible. When I work alone in the room I have music playing through the speakers and in a room with others I have headphones on. You won’t find it too distracting (as compared to a Lenovo I have for the odd windows task which sounds like a vacuum cleaner even when idle).
So far I couldn’t be happier. I need to travel to client offices a bit more to see how I get on with the larger mass of the laptop. I was close to getting the 14 but was really worried about thermal throttling, this is the unfortunate tradeoff and the big benefit is those lovely speakers.
AI - the good
The new hardware shaders (from the m4?) and neural accelerators in the M5 are a huge step up from those on my M2. Local LLM performance can be really fast, certainly sufficient for a lot of the tasks I’m working on at the moment (for example, creating knowledge graphs from a collection of academic papers). The memory bandwidth is really good and the large amount of memory allows several big models to be loaded into memory concurrently.
Image classification, depth mapping, pose detection, segmentation are all blazingly fast. Opens up some really useful new possibilities.
I’ve a few personal projects that utilise Apple intelligence and I notice a hige improvement running on the new machine compared to, say, my old M2 Air and my M4 ipad (the latter has improved shader performance so for very small models is much faster than my M2 Air).
The SSD speed make a huge difference, especially when loading in large models.
AI - the bad
The hardware shaders don’t support fp8. For image/video generation this is a bit of a pain as a lot of new models are distributed in fp8 only (some enterprising individuals have written scripts to convert them back to fp16 so that they will run, however the loss when quantised isn’t reversed, you just end up with a slower/larger model). Fingers crossed Apple adds fp8 support in the M6 - this will be enough to get me to upgrade again within a year. If you want a good example of how frustrating this can be, download latest comfy ui desktop and try running some of the supplied templates. I’ve run a few pytorch and MLX based models in fp16 and they are much faster than before, but this is still an area where Apple is lagging behind nvidia. For some AI workloads unfortunately Nvidia GPU still really a hard requirement (I just spin up an AWS VM for these cases, but if you regularly need it then don’t spend the money on a Max).
How about other people? Have you bought one, made a decision to wait?
