Remix.run Logo
khalic 7 hours ago

LPDDR6... that's not enough for speedy inference

hoherd 6 hours ago | parent | next [-]

Oh thank heavens, computing hardware that the AI companies will not buy all of the supply of.

CharlesW 6 hours ago | parent | prev | next [-]

It depends on the number of channels. The Xring O3 appears to have 4×24-bit channels, so 113.8 GB/s. The iPhone 17 Pro's memory bandwidth is ~76.8 GB/s. Seems fine?

GeekyBear 2 hours ago | parent | next [-]

According to Mark Gurman, the base M6 is bumped up to 200 GB/s of memory bandwidth with a single core performance bump of 15%.

rpdillon 6 hours ago | parent | prev [-]

If I recall correctly, Strix Halo gets about 450G/s. The M series processors get between 400 and 800G/s, and an actual NVIDIA card is like 5.5T/s. 113G/s is pretty slow for inference.

dust42 6 hours ago | parent | next [-]

The CPU is for mobile phones. An Nvidia H200 has 4.8TB/s. An RTX 5090 has 1.8TB/s. Both use ~700W - not comparable with a phone.

trvz 6 hours ago | parent | prev [-]

Strix Halo is 256 GB/s.

simlevesque 7 hours ago | parent | prev | next [-]

Not every computing platform is expected to do fast inference.

rjzzleep 7 hours ago | parent | prev [-]

Maybe,but memory bandwidth is much better than on the DGX Spark at least.

khalic 7 hours ago | parent [-]

I hate this RAMpocalypse so much

fauigerzigerk 6 hours ago | parent [-]

On the other hand, this year's RAMpocalypse could be 2028's RAMbundance ;P

khalic 5 hours ago | parent [-]

RAMvana was right there!!