Hacker News Reader

Comment by Sohcahtoa82 | Hacker News Reader

I'm not even sure you could build a $31,000 desktop computer even if you wanted to without resorting to some ridiculous "expensive for the sake of being expensive" parts. Even quad RTX 3090 Ti's would only set you back $8,000 if you got them at MSRP.

EDIT: Just saw the other comment and I stand corrected.

View on HN

You can run up costs pretty much arbitrarily with big memory and big storage. 2TB of RAM in a workstation will run you at least $30k if not more (it was $40k last time I checked), and you can go as high as 4TB in current systems. And big storage and NVMe arrays, it's almost a matter of "how much you got?", you can really scale capacity arbitrarily large if you've got the cash (although performance won't increase past a certain point).

This was always the dumb bit with the "apple wants HOW MUCH for a mac pro!?!?" articles about the "$50k mac"... it had $40k of memory in it alone, and the "comparable" systems he was building maxed out at 256GB theoretical and 128GB actual. That's great if it works, using a lower spec will push costs down on both sides, but it's not comparable.

Top-of-the-line 512GB LRDIMM DDR4 will run you about $2,500 before tax if you buy name brand Samsung. I know this because that is what is in both of my dual Xeon workstations. It gets pricey when you go through Dell or HP of course.

> 2TB of RAM in a workstation will run you at least $30k

The trick here is to use a board with 32 dimm sockets -- which requires an oddball formfactor-- but it radically lowers the cost of reaching 2TB.

But your point remains, change your target to 4TB ram (which really isn't an absurd amount of ram) and the astronomical costs come back (unless you go to 96 dimm socket systems, which have their own astronomical costs).

>I'm not even sure you could build a $31,000 desktop

A decked out Mac Pro can reach over $50,000 and it's not even that powerful as your 2x 3090Ti example, but that's the Apple tax for you.

Have you tried to price out a comparative PC? If you even can? Because there aren't many that will take as much RAM as that $50K Mac Pro and when you do find a PC that will all of the sudden you realize there isn't much of an Apple tax at all for the equivalent hardware.

Want to argue that Apple should have more variation in their offerings and price points? Sure - I heartily agree. But blithely tossing out a contextless $50K price tag as being some sort of "tax" is just silly.

It’s reasonably easy to build a threadripper workstation with 1TB of RAM (upgrade able to 2TB) and 2 3090’s for $20,000.

LambdaLabs will even sell you a prebuilt around that price with the same specs.

That said, I wouldn’t personally get into a pissing match over x86 Apples. I think their ARM offerings are far more interesting.

>Have you tried to price out a comparative PC? Because there aren't many that will take as much RAM as that $50K Mac Pro and when you do find a PC that will all of the sudden you realize there isn't much of an Apple tax at all for the equivalent hardware

You're deluding yourself with your claim. If you would have done 30 seconds of googling, you would have found out that Dell, HP, Lenovo and Puget Systems, all sell workstations that can beat the top spec Mac Pro or can build one yourself for way cheaper than them but as a business you want the support and no headache so it makes sense to buy prebuilts.

So yeah, there's an Apple tax on it.

https://www.imore.com/53000-pc-competition-apples-53000-mac-...

Quad RTX A6000s would be $24k and that’s what would go in a “workstation”

Venturing off-topic a bit here, but what exactly makes a "workstation" GPU? What's the difference between an RTX A6000 and an RTX 3090?

The A6000 has ECC memory and the 3090 does not. I think that's the chief differentiation between workstations and any other kind of desktop computer. Like a server, they will have ECC everywhere.

The main difference is you get twice the memory. If you don't need that, there is very little reason to get an A6000.

ECC RAM, different cooling setup (blower vs side fans), very different thermal characteristics, 24GB or 48GB, more bus width usually, optimized paths for data load & unload, GPU interconnects for direct GPU to GPU communication, shareable vGPUs between VMs, GPU store & halt, h/w support for desktop state, GPU state hand-off to another machine. It isn't just a "more memory" kind of thing.

> GPU interconnects for direct GPU to GPU communication

Note that nvlink is also available on the 3090, which does also have the same bus width as the A6000.

So interestingly to get multiple RTX 3090 cards using NVLINK you really need the blower style, which NVidia recently stopped making in the consumer line, otherwise you are fighting thermal properties and getting in to exotic cooling solutions. And you also get in to placement in slots on the motherboard on three-slot cards. And then you get into the number of PCIe PSU connectors on most supplies. You also get in to weird corner cases of four way NVLINK on 3090's acting as a "single card" but four way NVLINK on RTX A5000 is still independent cards as far as the OS goes. Regarding NVLINK, 3090 cards only support NVLINK (NVL 3.0 IIRC), rather than NVSWITCH. With NVLINK/NVSWITCH, with two RTX A5000 cards I get a total of 48GB of usable RAM for storing the machine learning models, and so on for multiple cards in the RTX A5K & A6K & AX00 line. But with the consumer cards, even though each card has, e.g. 12GB of RAM on them, your deep learning model is also limited to 12GB of RAM, no matter how many 3090 cards you put in. You're also limited to a maximum of four 3090 cards in total, whereas with the NVLINK & NVSWITCH on the professional grade cards, you are limited only by how many cards you can stuff in your data center - technically it is 16 GPUs on a single switch IIRC, things get funky with the fabric once you progress beyond 16 GPUs on a single switch but that's outside of the scope of this conversation. I may also have some details wrong, my brain is a little fuzzy afer dinner.

In addition to the other stuff people posted you also get to use the certified GPU drivers. Which means they actually tested that the card would work 100% with AutoCAD or whatever

The RTX is really slow for FP64.

Reply on news.ycombinator.com