Last year, OpenAI stunned everyone with ChatGPT and early this year, their Sora video generation tool almost spelled doom to the world of content creation if not the level of resource requirement was holding it back from being utilized by every user on Earth. But technically speaking, you can run it with a consumer-class rig and GPU. But guess how much you can out of it.
Yes, an “astounding” 240p video that is just 4 seconds long, according to a blog post by GPU CSP Backprop and reported by Jowi Morales from MSN. And that sort of result is achieved with the once “scalper’s bucket of gold” GeForce RTX 3090 with 24GB of VRAM.
As such, Sora is clearly designed with professional and data center GPUs in mind. For instance, NVIDIA’s Hopper-based H100 GPU can be configured with up to 94GB of HBM2e memory, and the H200 offers 141GB of HBM3e, both with higher memory bandwidth.
But the H100 alone would have cost you $30,000 in USD assuming you’re going with the PCIe version. Or you can spend about $6800 for an RTX 6000 Ada Generation card with 48GB VRAM since this is easier to find.
Even if money isn’t an issue, you’ll face copyright and IP challenges. OpenAI frequently deals with lawsuits over content ownership infringement, raising the risk that public users of Sora could become collateral damage.
