Reminder: there are no video outputs on these chatbot data center processors driving up the prices of graphics cards.
So they can’t even sell as used GPUs to crash the consumer GPU price market when the AI bubble pops.
This is a reminder that businesses aren’t “money focused calculation machines that optimize for the maximum possible profit.” They don’t worry about every little dollar, they just print money and use it to control you.
Raising prices for you is the goal, not a byproduct of some other smarter plan.
Some people don’t need the rest of this post, and it’s very long, so I’ll put it in a comment.
To be more specific here, GPUs are really, really good at linear algebra. They multiply matrices and vectors as single operations. CPUs can often do some SIMD operations, but not nearly as well or as many.
Video games do a lot of LA in order to render scenes. At the bare minimum, each model vertex is being multiplied by matrices to convert from world space to screen space, clip space, NDC, etc which are calculated based on the properties of your camera and projection type.
ML also does a lot of LA. Neural nets, for example. are literally a sequence of matrix multiplications. A very simple neural net works by taking a vector representing an input (or matrix for multiple inputs), multiplies that by a matrix representing a node’s weights, then passes the result to an activation function. Then does that a bunch more times.
Both functions want GPUs, but both need different things from it. AI wants GPUs with huge amounts of memory (for these huge models) which are optimized for data center usage (using cooling designed for racks). Games want GPUs that don’t need to have terabytes of VRAM, but which should be fast at calculating, fast at transferring data between CPU and GPU, and capable of running many shader programs in parallel (so that you can render more pixels at a time, for example).
This doesn’t mean it would be near useless to just add video outputs to neural net cards though.
Used data center GPUs might be equivalent to a low end or outdated GPU with extra VRAM, but there would be so many of them on the market, you’d see stuff like games being optimized differently to make use of them.
Nvidia sold many of their data center GPUs as full server racks. The GPUs aren’t in a form factor to use with a traditional PC and simply cannot slot into a PCIe slot because they don’t have that kind of interface. Look up the DGX B200, which is shipped in a form factor intended for rack mounting and has 8 GPUs alongside two CPUs and everything else needed to run it as a server. These GPUs don’t support video output. It’s not that they just don’t have an output port. They don’t even have the software for it because these GPUs are not capable of rendering graphics (which makes you wonder why they are even called “GPU” anymore). They cannot be plugged into a PCIe slot because there is no interface for it.
I try not to call them GPUs, though it’s hard to avoid.
But I didn’t know they’re not even capable of rendering graphics at a deeper level than just not having a video output.
It sounds like you definitely know some stuff I don’t, but wouldn’t it be smart for these companies to bid a bit more if they could, to make these builds with more resellable parts instead of using these crazy server rack combo platters?
I still think it’s an economy controlled top down by the authorities that makes this “profitable,” and when you boil it down it’s just a fancy mathy story to distract from them making special stuff for themselves they don’t want to share with us
Their customers don’t care about if they are resellable. They just want GPUs.
We aren’t their customers, and I mean this in the most literal sense possible. You can’t buy these. They only sell them to big companies.
Yes, it’s shit.