Mainstream is about to collapse. The exploitation nonsense is faltering. Open source is emerging as the only legitimate player.
Nvidia is just playing conservative because it was massively overvalued by the market. The GPU use for AI is a stopover hack until hardware can be developed from scratch. The real life cycle of hardware is 10 years from initial idea to first consumer availability. The issue with the CPU in AI is quite simple. It will be solved in a future iteration, and this means the GPU will get relegated back to graphics or it might even become redundant entirely. Once upon a time the CPU needed a math coprocessor to handle floating point precision. That experiment failed. It proved that a general monolithic solution is far more successful. No data center operator wants two types of processors for dedicated workloads when one type can accomplish nearly the same task. The CPU must be restructured for a wider bandwidth memory cache. This will likely require slower thread speeds overall, but it is the most likely solution in the long term. Solving this issue is likely to accompany more threading parallelism and therefore has the potential to render the GPU redundant in favor of a broader range of CPU scaling.
Human persistence of vision is not capable of matching higher speeds that are ultimately only marketing. The hardware will likely never support this stuff because no billionaire is putting up the funding to back up the marketing with tangible hardware investments. … IMO.
Neo Feudalism is well worth abandoning. Most of us are entirely uninterested in this business model. I have zero faith in the present market. I have AAA capable hardware for AI. I play and mod open source games. I could easily be a customer in this space, but there are no game manufacturers. I do not make compromises in ownership. If I buy a product, my terms of purchase are full ownership with no strings attached whatsoever. I don’t care about what everyone else does. I am not for sale and I will not sell myself for anyone’s legalise nonsense or pay ownership costs to rent from some neo feudal overlord.
Multi threading is parallelism and is poised to scale to a similar factor, the primary issue is simply getting tensors in and out of the ALU. Good enough is the engineering game. Having massive chunks of silicon laying around without use are a mach more serious problem. At the present, the choke point is not the parallelism of the math but actually the L2 to L1 bus width and cycle timing. The ALU can handle the issue. The AVX instruction set is capable of loading 512 bit wide words in a single instruction, the problem is just getting these in and out in larger volume.
I speculate that the only reason this has not been done already is because pretty much because of the marketability of single thread speeds. Present thread speeds are insane and well into the radio realm of black magic bearded nude virgins wizardry. I don’t think it is possible to make these bus widths wider and maintain the thread speeds because it has too many LCR consequences. I mean, at around 5 GHz the concept of wire connections and gaps as insulators is a fallacy when capacitive coupling can make connections across all small gaps.
Personally, I think this is a problem that will take on a whole new architectural solution. It is anyone’s game unlike any other time since the late 1970’s. It will likely be the beginning of the real RISC-V age and the death of x86. We are presently at the age of the 20+ thread CPU. If a redesign can make a 50-500 logical core CPU slower for single thread speeds but capable of all workloads, I think it will dominate easily. Choosing the appropriate CPU model will become much more relevant.