[joshuamorton]

A big part of it is the type of graphics card. FirePro (the kind in workstations) are much more expensive than normal general purpose graphics cards, which are again more expensive than custom SoCs which can be developed barebones for a specific use. Reason being that a calculation error that results in a dead pixel is fine when you are playing a video game. It'll be there for 1/60s and then be gone, never seen again. The same error in a disney film or a 3d scene rendered for a poster needs to be pixel perfect, so workstation class cards have a much lower threshhold for error, and cost accordingly.

(Also, and this is just conjecture: its possible that Apple intentionally overprices their pros as a sort of "look this is premium" cost, while the goal of an xbox is mass market sale).

[nailer]

> Reason being that a calculation error that results in a dead pixel is fine when you are playing a video game. It'll be there for 1/60s and then be gone, never seen again. The same error in a disney film or a 3d scene rendered for a poster needs to be pixel perfect, so workstation class cards have a much lower threshhold for error, and cost accordingly.

I've heard this a lot, and I don't doubt that it's correct, but could you explain why? Like do consumer class GPUs have less accurate floating point, or do their embedded algos contain hacks to produce less accurate results faster?

[khedoros1]

In the past, I remember being told that the fixed-pipeline cards could push a ton of vertices, but that they couldn't hand the same volume of textures and rendering effects as a consumer card.

One of the first things on the FirePro and Quadro Wikipedia pages is that the actual graphics chip in the hardware is the same as consumer levels. Aside from hardware tweaks like ECC RAM and possibly different display connectors, I think that the biggest differences are the workload that the drivers are optimized for and reliability guarantees for the results that the hardware produces.

[joshuamorton]

Unfortunately I'm not familiar enough to give you a conclusive answer. One big thing is that most consumer-class GPUs use normal memory, whereas workstation-class use ECC memory, which can correct for bitflips that might occur during normal operation.

[AlphaSite]

HBM is ECC by default, so that distinction is slowly vanishing.

[martinald]

That's not true. Loads of stuff is offloaded to GPU, which would cause a complete crash if it didn't work. GPUs aren't just layering rasters on top of each other.

[joshuamorton]

It is though, here's someone else saying essentially the same thing: https://superuser.com/questions/690388/why-do-workstation-gr....

Drivers that prioritize accuracy, and ECC ram fall well within the realm of "lower threshold for error", and price discrimination is covered by my parenthetical.

That said, the reason that custom SoCs are often less expensive than mass market cards is just because they have fewer things on them. If you don't need an FPU (which you obviously do here, but as an example), you can leave it off, which saves silicon and saves money. General purpose cards needs general purpose things, but a custom chip can often leave out certain components, although I don't know which ones those might be in this case.

[snuxoll]

It is true. Workstation graphics cards often have ECC memory, consumer cards don't. The biggest difference in most use though is better fp64 support on workstation cards, important for CAD applications as well as some use cases with 3D animated films or special effects.

[pixel_fcker]

Animation software works just fine on geforces (no-one uses AMD for professional work) - you don't need fp64 for anything. Rendering for film is 99.9% CPU-based, and even if that weren't true, all the GPU renderers on the market don't make use of fp64 anyway.

The only reason to use workstation class cards is because the software is certified to run on them (I.e. You won't get support from AD if you use a GeForce instead of a quadro). The only real reason used to be memory but with 11GB in a 1080ti even that's less true these days.

[lowbloodsugar]

Actually the difference between a top consumer graphics card and a professional card is just the drivers. A few years ago the president of nVidia went as far as to say that we was running a software company that monetized their software with "dongles" (the card). It used to be the case that cutting or adding a link would turn a $500 consumer card into a $2000 pro card: the link would be detected by the driver to enable pro features.

Also, aren't Disney movies raytraced?

[pixel_fcker]

> Also, aren't Disney movies raytraced?

Yes.