>integrated graphics with 4070 performance at the fraction of the size and power consumptio
How the fuck is Nvidia and Intel gonna compete against this?

>>717839173 (OP)
[citation needed]

The problem with these APUs is they're always so goddamn expensive there's no point versus even a shitty dedicated GPU. The original hope was a cheap APU would be good enough if you didn't care about bleeding edge triple A slop and instead they've just jacked the prices up so high on these I'm never even sure who is buying this shit.

>>717839173 (OP)

>Source: My ass

>>717839276
The AI in the name isn't there just as a buzzword. These are for VRAM-hungry AI applications. This is the cheapest way to have a shit ton of VRAM (though it's both system RAM and VRAM in this case).

>>717839701
What is the use case for this versus something like a 5090? Are people just using dedicated APU boxes for local LLMs? What sort of productivity work is being done on that versus a proper workstation? Is it just schizos afraid of chatGPT phoning home or something?

>>717839173 (OP)
If what you said is true, they won't - Nvidia is currently under investigation for violating US sanctions and distributing product to black markets, and Intel has spent 8 years committing and un-committing to building plants to develop products that use infrastructure more advanced than 14nm (which all of the competition moved through almost a decade ago).

>>717839934
Whatever happened to "CHINA CAN'T GET THESE AI CHIPS BECAUSE ITS A MATTER OF NATIONAL SECURITY"

>>717839269
>>717839658
https://youtu.be/uYLwDkGZOJk
https://youtu.be/GHlgHmdCVXo

>>717839173 (OP)
for larger language models (32B or above) 4070-tier memory bandwidth simply isn't good enough unless it's an moe. Even then, if you can afford one of these with more than 128 gb ddr5 ram you can generally afford a $7k pro 6000 which will run models at literally 3x the speed.

Another thing to consider is the quants. GGUF quants are fine for GPUs but with CPUs you will be compute bound, apparently there are new variants called ternary quants designed for CPUs but they're very new.

>>717840330
LLMs don't need low latency.

>>717840113
it was always retarded, how the fuck are you gonna stop smuggling when they're right next to taiwan?