I want to discuss the 2 differing core designs on RDNA2 cards coming in October

oldmanbal

2[H]4U
Joined
Aug 27, 2010
Messages
2,613
First off, take a look at the pic
D-Navi-22-GPU_Navy-Flounder-RDNA-2-Graphics-scaled.jpg

At the top left of the Sienna and Flounder diagrams, we see 2 different designs. The Flounder has a single thread type structure, whereas the Sienna has 2 has a 2 pipe graphics command processor, possibly similar to how Ampere now has a modified dual fp32 processing tree per cuda core. We've seen that this doesn't necessarily represent a 2x improvement, core to core, but a lot of that may be down to the efficiency lapses from going to samsung 10nm (8nm nvidia laff).

Additionally, they are doubling up on their video core next 3.0 for the flagship, which definitely backs up their double up on math in this configuration. However, the dual-pipe graphics engine is a precursor to the actual compute array, which either means double fetches per allocation, or potentially they are scaling up the engine, to actually push double number crunches per instruction. We won't know the truth until we see actual breakdowns at the end of October, but from a design standpoint, this should further differentiate the high end from the mid level cards.

Wot say u
 
The Flounder has a single thread type structure, whereas the Sienna has 2 has a 2 pipe graphics command processor, possibly similar to how Ampere now has a modified dual fp32 processing tree per cuda core. We've seen that this doesn't necessarily represent a 2x improvement, core to core, but a lot of that may be down to the efficiency lapses from going to samsung 10nm (8nm nvidia laff).
Block diagrams are dangerous. The graphics command processor has nothing to do with FP32 processing. In both AMD and Nvidia architectures there are "master" processors that co-ordinate work across the chip and send instructions down to the CUs/SMs. Typically there is one graphics command processor that handles the classic graphics pipeline and one or more compute processors that handle compute workloads.
 
it was from an earlier leak, I'm not someone who creates fabricated info to disseminate, you can peruse my post history.
 
Back
Top