Hacker News Reader

Comment by maccard | Hacker News Reader

> I am a programmer, so my needs are not that high

This isn't true for all programmers. I'm a programmer, and until 3 years ago a good chunk of my day was spent waiting for compiles to finish.

> Or perhaps it is that after 12 years, the laptops are now on bar with desktops. I think the biggest improvements for me in last 20 years have been the SSDs and then the NVMe.

Not even remotely close. I have an AMD threadripper [0] in a desktop PC under my desk that I use for development. This is a mainstream, readily available x64 processor with 32 cores (and hyperthreading) that can sustain just 3.8GHz for at least 30 minutes (I've not pushed it any farther than that because I've needed to). Based on PassMark benchmarks [1], the threadripper is substantially quicker than the M1 Max.

[0] https://www.amd.com/en/products/cpu/amd-ryzen-threadripper-3...

[1] https://www.cpubenchmark.net/compare/AMD-Ryzen-Threadripper-...

View on HN

Thanks for taking the stand here; generally new laptops are fast and especially the M10 max seems like a nice rig that beats some desktops for some programmer workloads. Maybe even many desktops for most programmer workloads, IDK. But besides not all programmer workloads being equal, desktops also compromise high end platforms like your sTRX4, and those are a little bit different.

Also, I'd like to add memory constraints as an example of workload inequality: The M10 max only seems to support 64GB. Good for many, but would be a huge draw back at my last dev job. The problem could inherently use huge amounts of memory, and exact memory usage was not always predictable. Especially for those colleagues which did a lot of work for customers, changing one parameter could mean coming back to a OOM after the program run for a night (or longer). Thanks to 128GB this became much rarer, but the for the type of program at some point the only viable option to safe memory is to accept a worse result.

> Also, I'd like to add memory constraints as an example of workload inequality: The M10 max only seems to support 64GB

That's a really good point. In my last job, our machines had 256GB of memory to fully utilise the 3990x cores.

> desktops also compromise high end platforms like your sTRX4, and those are a little bit different.

They're a little bit different, but they're still mainstream, readily available desktop envioronments. You can go on Scan here in the UK and order one of those systems just like you can order any other intel or AMD system.

My 2010 Mac Pro has 48 GB of RAM, this 2021 M1 Max has 32 GB. So not much has changed on that respect either. Except on 2010 I imagine it was struggle to get even 16 GB on a laptop.

I dunno man.. You're comparing a CPU which costs as much as the complete machine (or two M1 airs). It's not portable has a TDP of 280W. Still it's 30% slower single threaded. For multithreaded it's 2.7x as fast. So yeah, maybe for rendering it's faster, but it won't be as snappy as the M1.

Then again, if you wanna use it for rendering, just buy 4 M1 macminis for the price of one CPU. Or 6 minis if you add all the other stuff you need.

The claim was that laptops are on par with desktops, not that they're good replacement for many workflows. This is just flat out not true.

> It's not portable has a TDP of 280W

That doesn't bother me in the slightest for a workstation.

> So yeah, maybe for rendering it's faster, but it won't be as snappy as the M1.

Nobody who cares about performance is doing rendering on a CPU, it's all GPU accelerated. It's not about "snappiness", it's about suitability for work. I upgraded from an 8700k to this machine and saw a 5x speedup (my compiles dropped from 30m to about 6 minutes).

> Nobody who cares about performance is doing rendering on a CPU, it's all GPU accelerated.

Has OSL already been ported to GPUs? You're implying that OSL users such as film studios "don't care about performance" but I suspect this simply isn't true.

Sorry, I can't edit my original post to update but you're right.

A large amount of commercial renderers and commercial rendering software is GPU accelerated which was what I was trying to refer to (and would like to update my comment to but can't).

That said, given the render times for indiovidual frames a laptop is going to end up thermally throttling which will wipe out the single threaded gains, and being limited to very small amounts of ram is likely going to make using a current M1 Mac impossible for anything like OSL.

> not that they're good replacement for many workflows. This is just flat out not true.

it is a good replacement for many. It might not be the fastest of all, but that doesn't mean it's not a good replacement

It's faster single threaded, and will be faster until you fully saturate more than around 14 threads.

Couldn't find a linux compilation benchmark, but here a comparison with the M1 on compilation performance:

https://openbenchmarking.org/test/pts/build-eigen

Compilation test shows that it's even faster.

It looks like this is basically a serial benchmark, so yeah of course the higher raw clock speed will win out. The run is also less than a minute, so most laptops will happily burst for that long, however what happens if you throw a 5/10/15 minute compile at it?

> Does It Scale Well With Increasing Cores?

> No, based on the automated analysis of the collected public benchmark data, this test / test settings does not generally scale well with increasing CPU core counts.

If your workload is compiling a single file serially, then sure this might be a great upgrade for you. I went through the copmpile benchmarks and the only other one that has an M1 mac and a 3970x on that site is build2 [0] which shows the m1 is 4x faster on a build that scales well with cores. FFMpeg [1] doesn't list the M1 in the test table, but it does have one benhmark [2] which is again quicker with the threadripper.

[0] https://openbenchmarking.org/test/pts/build2

[1] https://openbenchmarking.org/test/pts/build-ffmpeg

[2] https://openbenchmarking.org/result/2201040-NE-FFMPEGBUI53

One interesting thing would be to know how much of the performance improvement in some of the compile tests is due to the high-bandwidth memory interface on the M1 chip. Since the M1 chip has both different cores AND a quintuple memory bandwidth compared to Ryzen's/Threadripper's off-chip memory, if you consider things like cache misses, one has to wonder what's the contribution of both to the performance increases.

Completely agree, and I don't have an answer to that. However, all of that goes out the window if your macbook starts throttling. I don't doubt that an m1 chip with adequate cooling will outperform an x86 one, but at that point you're not on a laptop anymore

Yes, definitely not true for all programmers. I guess with games, if you cannot get real time with instant compiles development experience with 120fps running 8k in debug mode, there is a room for improvement.

I compared highend 2010 desktop to highend 2021 laptop on my use case (which is arguably much lighter). And for my workloads, the laptop is no more a limiting factor (which is great, but not a leap from 2010 desktop). I can imagine todays desktops could be much better compared to todays highend laptops, but it feels like the most of what I do, there is little reasons to upgrade. Which is also great. With Macs the latest macOS support is removed usually from 5 years old machines, but that feels mostly baseless from technical perspective as with hacks you can still run the latest just fine.

Perhaps it is a nostalgia / romancy. Having a 12 years old machine as you daily workhorse where upgrade was not a leap. That is something I haven’t seen before, and I have been in this industry since 90ies and a PC user since 80ies.

> Yes, definitely not true for all programmers. I guess with games, if you cannot get real time with instant compiles development experience with 120fps running 8k in debug mode, there is a room for improvement.

Day to day that doesn't matter. I work on multiplayer games, and being able to run 8/12/16 clients on one machine at 10 fps at ultra low graphics is far more important than being able to run stupid resolutions. You do bring up another point though - GPU performance. Apologies for the links but I'm on my phone, but [0] and [1] have 3dmark benchmarks for the M1 GPU and an RTX 3090 GPU - the 3090 scores twice what the M1 GPU does. This might not be important to you, but it is to some people.

> And for my workloads, the laptop is no more a limiting factor

Ultimately that's all it comes down to. If you're not limited by the form factor/performance of a laptop power to you, but some of us are!

[0] https://www.notebookcheck.net/Apple-M1-GPU-GPU-Benchmarks-an...

[1] https://www.notebookcheck.net/NVIDIA-GeForce-RTX-3090-GPU-Be...

> the 3090 scores twice what the M1 GPU does

It helps that you can buy a maxed out macbook pro for the price of a single 3090 though.

A 3090 is $2300 right now, even paying the scalper tax.

How much M1 MacBook Pro is that getting you?

That said, I do wish the M1 GPU could game; my downtime hobby is the only thing stopping me from buying one.

>A 3090 is $2300 right now, even paying the scalper tax.

Where can you find that? lowest I can find is 2800 new. There are some pre-owned for 2700 and some open auctions at roughly 2300 but those will land substantially higher than 2300 I would imagine.

I can buy a high spec showhorse for $40k, or a new mid range sedan for $40k. That doesn't mean that the maxed out showhorse will work for commuting to work every day.

The 3090 is such an overpowered (and expensive) card, that another GPU doing half of what the 3090 can do is still damn right impressive. And in a Laptop!

Yeah that's actually a serious compliment to the M1 GPU if the 3090 is merely double the scores?! I expected it should be like 4x or more! That's awesome.

Oh, can you share more? What is your programming environment?

I am debating a move to desktop/workstation with desktop CPU to support smoother Scala/lots-of-containers development experience.

Pandemic made laptops less appealing. With digital nomadism being a question mark for some. Yes, I do know some still continue to practice this lifestyle. I can be only jealous :)

> Oh, can you share more? What is your programming environment?

I work in games and write C++. Compile times are long, and we deal with large amounts of assets (current project is ~50GB of textures, models, animations, sound. This is small. Last project was >500GB).

> With digital nomadism being a question mark for some.

Sure, if you want portability or being able to move around this sort of machine won't work for you, but I live with my partner in a nice city, work remotely and have a separate home office. I have almost 0 need for the portability of a laptop.

Depending on your needs(!), you can just use the laptop as a frontend for a much more powerful machine. I did that for a while with a Sandy Bridge (core i7-2xxx gen) dual socket server in the basement, and it did work quite well. No matter if I was at home or somewhere else - but for dev work ssh into the Linux host OS was enough for me, with an occasional RDP into a Windows VM.

I considered this option with NoMachine and in-house power-machine or cloud workstation. In a way make sense, but ultimately, I decided against because:

Security – having open port to workstation makes me scared; bugs/0-days are discovered every day;

Quality Internet Connection – less tricky as most access would be local LAN, but I am certain that few times I need it remotely to fix a production issue, something will go wrong;

Maintenance vs Cost – in-house machine would be cheaper but will require remote power on/off capability vs in cloud server being rather pricey

But yes, in general, I agree. Having a light, portable device with long battery life and connecting to a remote, beefy machine would be a dream. And I am not imagining it, as this was my workflow back in 2013 when I was working for a bank. If it was possible then, it definitely is possible today.

Have high hopes for https://shadow.tech/

Steam remote play or moonlight/sunshine [0] working great for me in setup like this.

[0] https://github.com/loki-47-6F-64/sunshine

What about real world tests for you specific workflows? My M1 Max runs my daily bread and butter R scripts approximately 4x faster compared to a top of the line Intel Coffee Lake. For me it's a huge improvement. Having over 100GB/s RAM bandwidth per core and a huge cache are a big thing for many latency-sensitive workloads.

I don't actually own an M1 Max, so I can't do a straight up comparison right now sorry. I upgraded from an 8700k to a 3970x, and for compilation the speadups were pretty much on paper gains. Each individual compilation task is ~10% slower, but the parallelism of 32 cores means that overall it's a 5x speedup over my previous CPU.

3 times faster with 64 cores instead of 10, that slower single thread performance is having a big impact.

32 cores, not 64.

Reply on news.ycombinator.com