Area efficiency of HW RT and TensorCores. *spawn

T

troyan

Veteran
pTmdfx said:
Look at RTX 3060. What can one do realistically with ML upscaling and ray tracing on 4K 60Hz, that brings enough IQ gain across all genres for a platform vendor to say “bloody hell let’s get everyone onto this”? Moreover, have one ever thought about the opportunity cost of _not_ spending on these stuff at all? Say… better perf-per-area (PPA) — as a controversial figure here touted — for things not RT and ML, within the customer’s MSP target.
Click to expand...
Nothing beats the area efficiency of HW RT and TensorCores.
 
TopSpoiler said:
Tensor Cores, which exist in SM are not shown in the die shot. Perhaps you misunderstood TPC as Tensor Core?
Click to expand...

Tensor cores are called out in the shot but that’s just their guess. Clamchowder (from chipsandcheese) has a different guess. We don’t know who’s right.
 
trinibwoy said:
Which one? The tensors are annotated in the original shot. You may have to zoom in the see the text clearly on the bottom left GPC. Clamchowder’s reply is in the same thread.

x.com

x.com x.com
Click to expand...
The shot shown in the linked tweet here in the thread is not annotated for Tensor cores, also not in the lower left GPC: https://pbs.twimg.com/media/GMQ2AlnbcAA31vx?format=jpg&name=4096x4096

But thank you for linking the reply, that helps following what people are talking about.

edit: with cleaner vision (coffee might have played a role here) I see what you've been talking about. Sorry.
 
Last edited:
You must log in or register to reply here.
Back
Top