←back to thread

1080 points antipaul | 1 comments | | HN request time: 0.213s | source
Show context
andy_threos_io ◴[] No.25066484[source]
Can somebody explain how is M1 going to work on larger GPU tasks (rendering, encoding etc) with it's memory bandwidth M1 total memory bandwidth is 68.2 GB/s (128 bit LPDDR4x-4267)

AMD Radeon Pro 5600m memory bandwidth is 394.2 GB/s (2048 bit HBM2)

https://www.techpowerup.com/gpu-specs/radeon-pro-5600m.c3612

replies(5): >>25066523 #>>25066548 #>>25066569 #>>25066693 #>>25067438 #
tedd4u ◴[] No.25067438[source]
They were waving their hands around a bit saying that in addition to on-package RAM with high bandwidth, they have done optimizations inside driver, OS and APIs to eliminate 1 or even 2 or more RAM to RAM copies for common tasks. So maybe there is a higher effective bandwidth when comparing to Wintel machines. It was very vague. Would love to read a detailed article comparing the systems.
replies(1): >>25083718 #
1. wmf ◴[] No.25083718[source]
Intel and AMD theoretically also support zero-copy but I wonder if it gets used in reality since most apps are probably optimized for discrete GPUs.