![[H]ard|Forum](/styles/hardforum/xenforo/logo_dark.png){kind=logo}
Disney has a huge fleet of the 2019 studios which were all still Intel's and are by all accounts dated.
Yes but there a giant amount of people not working on the 3d-sfx stuff at Disney (the editors for example are probably on mac, graders, sounds), who would be mostly on MAC.
For the type of work most buying these things for - the 192GB shared is more useful than ~TB of normal RAM. Keep in mind that 192GB of ram on this system means you get that much for the GPU as well. The functional VRAM is more important for a lot of these workloads.
Sure, I realize that. For me, the Studio would be a perfect replacement for the Intel Mac Pro, if I had the 2019 version, as I’m in video editing. For ML, it’s a downgrade. Not to mention that the Intel version could have Linux installed bare metal whereas the M{x} chips don’t have full native ARM support.
I'm running Debian decently hassle free via Parallels on my M2 Max MacBook Pro. No, there isn't full native ARM support with many packages, but it is what it is. Although most standard x86_64 binaries will run just fine if you deal with Rosetta.
Pooling 2 Quadro RTX 8000 get you to 96GB
Totally knew it was coming, but not quite what I hoped for.
I suspect Apple is just being pragmatic — it can't keep selling a four-year-old system just because there are some situations where the extra memory will be useful. Better to get something out now than to wait until there's enough RAM to completely satisfy the machine learning crowd. I just hope the M3 Ultra (or ideally, an M3 Extreme) significantly boosts the RAM ceiling.
Agreed. TBH if this had expandable ram (even tiered between SoC and off-chip), I'd be starting to think my way through it - need to check on BSD support for M1/2 though
It's actually still $700. It's only 400 if you order it with the mac pro hehI wonder if the wheels still cost $700!
NVM, I see that it's a $400 option when ordering. Wow what a bargain!
For AI you'd genuinely be better off with an Ampere Altra running Fedora or Ubuntu with the Nvidia card of your choice stuffed into it, or even a Raptor Computing Blackbird running Power9. Nvidia drivers work fine on both, albeit only for CUDA/OpenCL on the ppc64le side. I also have questions about Apple's PCIe implementation - to my knowledge eGPU support wasn't implemented because of an M1 bug with memory-mapped I/O in ranges GPUs heavily use, which would have required altering aspects of the L2 cache to properly fix. In any case, I would not drop a minimum of seven grand to find out the answers...
Agreed entirely. I do want a blackbird personally - but I’m praying they build a power10 model soon. And I’d slap an AMD GPU I. For graphics and then nvidia for the ai work.
No, I'm pretty sure you can throw it into a stock standard ATX case. See here. I've got an eight core Power9 in a Blackbird and the 64 core Altra still tempts me. The Blackbird has a 16x PCIe4 slot and an 8x as well - I am tempted to delegate the 16x to an Nvidia card, then perform some adapter fu and plug my Radeon into the lower 8x slot by way of a 16x adapter mounted in the case. Maybe next year. I will say that getting Quake II RTX running on the Radeon RX 6600 on Power9 was quite satisfying!
Neat. Now if I only had a use for a mega-ARM server
Just like if I only had a use for a Power9 box. How do you like the Blackbird? I'd be tempted to go with the full Talos board (just for the extra slots - make it a combined workstation and mega storage server, given the case it would have to go in), but I might also be insane. TSMC isn't even confident in its ability to meet the demand for the iPhone 15 and iPad Pro with the M3 refresh until some time in mid-2025, forget about the Macbook Pro lineup and the Mac Pro or Studio options.
I bought it for several reasons - flush with pandemic cash, curious about non-x86 desktop architectures, had a lingering soft spot for PowerPC, and wanted to see what it was like to buy a machine that had nowhere to go but *nix I installed myself. It’s been a great educational experience - I’ve learned a ton about navigating build systems to insert optimizations, and that the current status of Linux makes it possible to use even outré hardware as a competent desktop. It has its issues - SIMD is relatively weak, video support always trails the leading edge, and sometimes you take a gamble and force SIMD translation in GCC to coax more performance out. But I love it, and predict I’ll be running this as a Linux desktop daily driver for years before I put it to work as a quintessential Server in the Corner until it dies of old age.Neat. Now if I only had a use for a mega-ARM server
And huh, didn't know that about Power10. Valid point on the size - those suckers are HUGE - but I'd still hoped that it would show up. I feel weird buying a 7 year old architecture and processor. Did some reading on it - fascinating issues there, including the fact that it's only high-end parts (and looks likely to stay that way - Power is a shrinking market, after all.)
I don't think we've seen anything official from TSMC on this front, so I wouldn't read too much into the reported problems just yet. I won't be shocked if iPhone 15 supplies are tight, and if we don't see most M3 Macs until 2024. But I also don't want to proclaim doom and gloom because of sketchy rumors from sites like DigiTimes.
TSMC already had to apologize to their board because they yields are so bad they are only charging Apple per successful chip not per wafer as per their normal pricing.
Fair enough — still, I'd like to see how this pans out. Even the iPhone 14 Pro production delays didn't last long in practice.
Based on the things I have seen so far from the various behind-the-scenes documentaries on the technology they are currently using Lenovo Threadripper systems with multiple NVLinked workstation GPUs and for the AR aspect they have multiple people with multiple tablets just sort of holding them up and wandering around that way, looks awkward AF.
I don't think the op-ed says much that we don't already know, and it's difficult to say just how much the sacrifices will truly impact performance in some cases. The Mac Pro might be limited to 192GB of RAM, but it's not going to bottleneck as badly as some workstations might.
Apple silicon apparently can’t even talk to GPUs outside the SoC
From what I've heard - and I may be getting something at least a little wrong, as I'm going on some hearsay - it's a bug in common with the Raspberry Pi 4(!), where the memory controller can't reliably perform MMIO to memory ranges that non-integrated video heavily leverages. I'm not sure if this is directly connected to the bug that bit the Synquacer Socionext Developerbox a while ago, but that's another ARM machine with hobbled PCIe compatibility (From what I remember that one's only certified to use a GeForce GT 710 via nouveau with a source patch applied, and no hardware 3D). M1's problem can't be cleanly worked around without a silicon fix - Jeff Geerling's done a lot of investigation on the Pi side, and without rearchitecting the only potential workarounds involve kernel hacking too ugly and unreliable to ever make it into the kernel source tree. I'd be interested to know how the AS Mac Pro fares when the Asahi Linux devs get their hands on one - maybe I'll start following Hector Martin over on Mastodon soon.
Yeah Apple tends to do a lot of research on anything they launch. They know their market very very well and they have treasure troves of data to back that up. I don’t think they have released a product into a vacuum in a very long time, the AR goggles being a noticeable exception because it’s completely new for them. So if they are launching these Mac Pro’s they already know who they are selling too and exactly what they want them for.
I have accounting running that now on some M1 based MBP’s and Win 11 ARM in Parallels is actually running the software I needed them too.
I don’t know about “bug” exactly, but ARMv8 is designed for SoC first and foremost and externals aren’t really on that menu.
Looks like you're right - it's a mess, and a broken PCIE root complex that many have implemented the same as how Broadcom did it on the Pi. I spent the last half hour or so digging through a lot of the details for fun - it's whacky as shit. Short answer - Radeon cards MIGHT work with a shit-ton of fiddling - on Linux or BSD (most likely linux). Poorly. Nvidia is right out with the closed source drivers, and the open source ones have massive limitations as you noted, and it's gonna be buggy as hell for a very long time until they either hack in fixes or just wait for a new generation of silicon.
Uh, no. That’s not how that works
. At least if you’re talking about GPUs.
This is also before you start digging into the weeds of kernel page sizes on non-x86 architectures, where ppc64le since Power8 or so has supported 4KB and 64KB kernel sizes, and most OSes on the architecture only officially support the 64KB sizes. That wasn't a problem for the pre-GCN radeon driver, but it took years for amdgpu to properly manage the gap. Heck, RDNA2 hardware didn't work on ppc64le until kernel 6.1 or so, and there was an oogy regression on Navi/RDNA1 hardware until recently too. And Apple Silicon only(?) works with 16KB kernel sizes, which could also throw a wrench into drivers. Nouveau flat out doesn't work on non-4KB kernel page sizes and its maintainers don't seem fussed to do anything about it yet. At least Intel's fixing it with the xe driver, though that hasn't been mainlined and may never support the HuC microcontroller for Alchemist parts, crippling its video encoding capabilities.
Chimera?
.
Chimera is a non-GNU Linux spearheaded by the former developer of Void Linux on ppc64le, intended for aarch64, ppc64le, x86_64, and RISC-V. It's a Linux kernel and FreeBSD userland, and by all accounts it's coming together pretty well. I'm looking forward to jumping to it once it reaches beta - more info is here.Chimera?
I’ll admit you have me debating now which would be most useful as a non-x86 workstation, the Power9 box or an Altra. I think Power9, but I’m not positive. Also a mental debate on running FreeBSD or Linux on them.
That being said, my workstation needs at home tend to be either super heavy lifting with video and VMs (which my Threadripper handles - or my x299 box), or all terminals and web browsers. Vms on power9 won’t do me any good (or arm), and the rest can be handled by, well, anything