- Joined
- Mar 21, 2016
- Messages
- 2,508 (0.79/day)
I'm not overly keen on the approach Nvidia went about this. If I were AMD I'd be looking at taking a CCX approach and have two very distinctively different chiplet designs geared heavily for FP32 and and another toward FP64. They could then bin each of the designs and have a bridge I/O die between them which they could then mix and match based on binning SKU's for quality and defects they need to laser cut. It would be better in terms of die sizes and harvesting components along with broader selection of custom tailored performance that more suitably meets. The I/O bridge chip could have basic functionality general desktop and HTPC use, but also act as bridge between chiplet's that could be more customer tailored for stronger or lighter FP32 or FP64 duties. Below I kind of mocked up different chiplet design's they might come up. I really think a tensor core could be part of the bridge chip as well if it makes sense to do so. Basically utilize additional CCX's for additional GPU resources geared more heavily toward task specific use cases thru the bridge chip. The bottom right corner is Nvidia's stock design, but the other three are different arrangements that could've been designed as individual CCX die chiplet's and then binned. You want more FP32 use a pairing of the top left that type of thing stronger FP64 top right more of a balance, but less tensor cores bottom left. To a degree they are already designed similarly with a large monolithic die, but they are more constrained by that design at the same time than actual CCX chiplet's that were designed more favorable for one or the other and bridged together seamlessly by another chip in a sensible manner.
How compelling ampere will be is difficult to say. I do like that DLSS is improving and becoming more flexible though it's still quite developer dependent it doesn't just work with everything so while it's a great option to have when available it's of no real use otherwise. I'm a much bigger fan of performance or image quality enhancements I can just outright utilize whenever without being reliant on a developer to implement it's functionality. We've seen how well that works with mGPU for CF/SLI I mean hey developers are "lazy" or more appropriately from their own perspective time is money.
How compelling ampere will be is difficult to say. I do like that DLSS is improving and becoming more flexible though it's still quite developer dependent it doesn't just work with everything so while it's a great option to have when available it's of no real use otherwise. I'm a much bigger fan of performance or image quality enhancements I can just outright utilize whenever without being reliant on a developer to implement it's functionality. We've seen how well that works with mGPU for CF/SLI I mean hey developers are "lazy" or more appropriately from their own perspective time is money.
A little bit of a AMD STANboi...on my Intel CPU & Nvidia GPU...is what it is I did quite like AMD's CPU's prior to bulldozer and it's current CPU from the release Ryzen have been highly "compelling" even if they weren't 100% perfect and continued to extend upon that aspect. They've even managed to make Intel's products drastically more compelling than the dogsh*t products they'd been overpricing and selling to consumers taking advantage of everyone they could.Also, nice new account with posts only about calling people AMD fanboys bro. Welcome to TPU.
Last edited:
