> run free LLMs locally at native speed

This reads like a hallucination. What does native speed even mean?

kyxsc 36 minutes ago | parent | next [-]

for example, models running at like 100-150 tokens/second (or faster!) vs 15 t/s

(fable/sol are ~60 t/s, and OpenAI just announced their Cerebras partnership(?) for "ultrafast" mode of 750 t/s)

models aren't able to run that fast right now on our consumer/prosumer hardware. M5 Max for example has a memory bandwidth of 600 GB/s. a 5090 has 3x that, so running the same model on a 5090 is that much faster (provided the model is within 30GB).

running a bigger model on an M5 Ultra is still much slower than running it on a Blackwell chip with sufficient vram, CUDA being a major difference. if apple can bridge this gap, interesting things will happen... and just imagine if M7 Ultra has comparable speeds to Blackwell (or even Rubin)!

	▲	toasty228 3 minutes ago \| parent [-]
		Meanwhile the GB300 used by hosted llms: GPU Memory Bandwidth: 7.1 TB/s Interconnect Bandwidth: 900 GB/s bidirectional https://pi3g.com/nvidia-gb300-specifications-including-memor... If you think M7 will hit even 15% of these speeds you're very optimistic.

▲

lmpdev 34 minutes ago | parent | prev [-]

I assume they mean same t/sec as a SOTA cloud model