Comment by sylens | Hacker News Reader

Comment by sylens | Hacker News Reader

Apple’s high end desktop strategy has been confusing to me for several years now. They made a big deal out of the iMac Pro, even though that seemed like an obvious quick fix for power users since the new Mac Pro wasn’t ready at the time. However, I never understood why Pros would want to pay more for an an all-in-one they couldn’t self service or upgrade when they probably already had monitors they preferred using.

Then they finally released the 2019 Mac Pro at a cost that I thought was prohibitive to all but the highest end professionals. I don’t remember previous Mac Pros being targeted this exclusively; heck, somebody in my freshman dorm in 2006 had the cheese grater Pro at the time for music production.

When they unveiled the Studio, I thought that they were replacing the Pro line with that product…but then they confirmed the Pro is still coming. I really do not know what differentiation there will be between the Studio and the new Pro. I’m curious to see what they do with expansion and dedicated GPUs, but considering they don’t play ball with Nvidia anymore it’s probably still going to be a niche product

In 2006, computing was less generous. A music producer probably needed a Mac Pro for their workflow to be efficient. Since then, Intel has massively reduced the delta between laptop and desktop chips. Coupled with the increase in computing ability of our processors while most tasks do not increase their needs (that music producer does not need more computing than in 2006) and desktops slowly faded to nothing but the most demanding tasks (ML, 8k video production, science, etc.) and Apple has gone upmarket with the Mac Pro to chase this market.

With that change, laptops have become the computer market. Look at Apple's iMac offering; they no longer have multiple models and the insides are now laptop parts. The Mac Studio is simply the desktop computer that gets the Macbook Pro chips. Since they don't want to engineer a cooling solution behind a monitor every time, we get a simple box with a ginormous fan.

The Mac Pro was supposed to be the computer with the truly desktop-only ARM chip. Even that seems cancelled now. I don't know what Apple is planning.

That's an interesting point. Do you know a good reference comparing laptop vs desktop cpu performance? I'm always confused by comparisons that inevitablty are affected by other laptop specific components (differences in mgpu / motherboard chipsets etc).

I guess you could look at the Mac Benchmarks of Geekbench: https://browser.geekbench.com/mac-benchmarks

Let's forget the whole Apple Silicon thing, and look at 2020. The difference in single core is minuscule, but the big difference is multicore. Unless you need those extra cores, those computers are equivalent.

Now compare a 2010 Mac Pro and the 2010-2012 consumer laptops. Single core is 100%, multi-core is at least 400%. That's a big delta and that's only 2010.

If you go back into the past, it would surely ebb and flow but surely stay in that zone. I'd expect a laptop to be half the speed of a big pro desktop all the way to the prehistoric times when all computers were beige and laptops charged extra to get rid of screen ghosting.

That's a really good point. I think it would be far more likely to see that same freshman with a Mac Studio in their dorm now instead of a Mac Pro even if it was at a similar price point as the Studio - just smaller and more practical.

The MBP was plenty sufficient for most music production tasks even in 2004 when I got my first one. If I were running a pro studio with a 32+ channel mixing desk, I could see needing more, but I was able to manage 16 channels of live recording and at least 64 channels of live mix on my 2004 TiBook.

Berklee College of Music [0] requires at MBP with 16 GB of memory, 512 GB or disk space, which I think has been available since at least 2007.

[0] https://www.berklee.edu/sts/macbook-pro

Of course my point is not super strict about dates, it's about trends. I'm sure you were fine in 2004, but the more you go back into the past, the more a desktop is required.

And the more you move into the future the more a laptop is fine.

> […] and the insides are now laptop parts. The Mac Studio is simply the desktop computer that gets the Macbook Pro chips.

«Laptop parts» connect M1 CPU clusters to RAM via a 512-bit wide (M1 Max) and a 1024-bit wide (M1 Ultra) memory bus yielding 100 Gb/sec (per a CPU core), 200 Gb/sec (per a CPU cluster) and 400 (M1 Max)/800 (M1 Ultra) Gb/sec cumulative data transfer speeds. GPU and ANE cores access the main memory via the same wide memory bus. L1 caches (I+D) are 128kB+64 kb (efficiency cores) and 192+128 kb (performance); L2 caches are 24 Mb (M1 Max) and 48 Mb (M1 Ultra), and L3 cache is 48 Mb (Max) and 96 Mb (Ultra). Auxiliary tasks (disk encryption, video encoding/decoding etc) are off-loaded to dedicated processing units which is reminiscent to that a typical mainframe architecture with channel processors and channel programmes rather than that of a laptop.

Most «hi-end» servers today do not come anywhere close to such configurations leave alone the form factor and the net system efficiency. 512-bit wide memory bus is currently only found inside hi-end GPU cards, and the compute off-loading / accelerating units are an add-on feature for a server that may or may not be available.

Only POWER10 and Telum systems from IBM offer similar (albeit far more scalable and feature rich) designs and at a completely different price scale.

> 512-bit wide memory bus is currently only found inside hi-end GPU cards,

AMD EPYC CPUs just went from 512-bit DDR4 to 768-bit DDR5, per socket.

… which is another testament to the fact that a wide memory bus is an attribute of server or hi-end components rather than «laptop parts» that one can throw into a bucket, stir vigorously, cook on a slow fire and get a M1 Ultra with a 1024-bit wide memory bus.

I ought to have used «typically» rather than «currently» as wide memory buses were coming and going more mainstream in the server space, whilst not yet being widely available today. The hardware transition cycle will take a few years, and as of today «currently» still holds true, semantically and factually, as cloud platforms and data centres run nearly entirely exclusively on Xeon something Lake and EPYC 3rd Gen CPU's, and neither has such a wide memory bus.

> … per socket.

Does that mean that 2x socketable EPYC CPU's (say, 512-bit wide each) will be able to access RAM via a 1x 1024-bit wide memory bus? If yes, how will access to RAM work, i.e. access multiplexing – memory banks? time interleaved access? something else?

> Does that mean that 2x socketable EPYC CPU's (say, 512-bit wide each) will be able to access RAM via a 1x 1024-bit wide memory bus?

No. All recent dual-socket platforms are NUMA, meaning a CPU only has fast access to its local memory, and accessing memory attached to the other CPU socket requires going through a slower link between sockets.

NUMA-aware operating systems and applications try to avoid allocating memory on a different NUMA node than the processor cores that will be using that memory. On AMD EPYC, the internal structure of the IO die (memory controller) means it is sometimes advantageous to enable the option for the system to report more than one NUMA node per socket.

I am no stranger to NUMA since SGI days, and AMD EPYC is a ccNUMA.

If I read it correctly, 512- and 768-bit wide memory bus is available to access the socketed CPU's local memory only (which is relatively small), and access to the shared memory (which is large to very large) seems to be still via a switched Infinity Fabric memory controller.

How wide is the Infinity Fabric nowadays? I can't easily find the conclusive evidence.

> I don’t remember previous Mac Pros being targeted this exclusively; heck, somebody in my freshman dorm in 2006 had the cheese grater Pro at the time for music production.

Well, even before the rise of covid-induced work-from-home, Apple has been shepherding almost all of their users onto laptops.

That college student's 2006 Mac Pro might have had 1GB RAM [1] - but these days, Apple has a laptop for users wanting 64GB RAM. And it isn't a suitcase-sized mobile workstation either.

People wanting more power than that for things like ML and scientific computing will often be shifting their workload to the cloud or a cluster anyway - it's only people doing interactive tasks like CAD and high-end video editing who are buying the Mac Pro.

[1] https://youtu.be/l72MsGZQA8Q?t=418

The switch to working from home made high-end workstations a bit more relevant. While offices had fast and reliable internet connections, the quality of home internet is often much worse. You can't rely on the cloud as much from home as from office.

I develop bioinformatics methods and software, and I see significant productivity gains when I can run tests with full datasets locally. So far I've been using an iMac with 128 GB RAM, but I'd like to upgrade to something with 256 GB or even 512 GB. As the Mac Studio is pretty underwhelming from my perspective, the options are basically a hypothetical Mac Pro and a Linux box.

> While offices had fast and reliable internet connections,

most offices I’ve been around have internet that tops out no more than a little faster than a decent home connection, but have dozens of people sharing them.

A while back I built a linux box for bioinformatics and was happy with the performance [1]. You might need to look for suitable main boards to support your RAM requirements, though.

[1] https://www.ecseq.com/blog/2020/Whole-Exome-Analysis-on-a-Ma...

Reply on news.ycombinator.com