News AMD Navy Flounder and Sienna Cichlid Specifications Leak in ROCm Up

Endymio

Reputable
BANNED
Aug 3, 2020
725
264
5,270
> " if the GPU runs on TSMC's latest 7nm process, we could see RTX 3080-levels of performance, but with much better efficiency. ..."

How exactly is that conclusion reached? Isn't that the same node the 3080 is manufactured on?
 

dwards

Reputable
Jan 20, 2016
5
0
4,510
Yes, but I thought the two nodes were essentially equal in performance (and density), but that NVidia switched to Samsung because the thermals were better?
No, I remember reading that Samsung’s node can hold 80 somethings per square mm, while TSMC’s can hold 96, which is exactly 20% more, so I assume that Nvidia could’ve made GPUs with better thermals if it used TSMC’s node instead of Samsung’s
 
Nvidia went with samsung due to availability from what I'veread, sorry no references to offer as it was a while ago. TSMC is super busy.
It was a combination of factors. 8nm had a higher yield and depreciation cost so it was much cheaper per chip than the 7nm (Speculation) This decision was made 2 years ago. That's when you start to plan out the layout of the chip based on the fab rules. THIS IS ALL SPECULATION. But I cannot find fault in Adored's analysis.

View: https://www.youtube.com/watch?v=tXb-8feWoOE&t=0s


Plus 8nm Samsung was available. AMD and all the other seized that silicon capacity early on.
 

MasterMadBones

Distinguished
Yes, but I thought the two nodes were essentially equal in performance (and density), but that NVidia switched to Samsung because the thermals were better?
Samsung 8nm is roughly equivalent in to early-mid 2019 TSMC 7nm in terms of performance and yield, but efficiency is much worse. I'm not sure truly how efficient either of these nodes is, but when density is 20% higher and efficiency is 15% higher on TSMC, thermal density should be roughly the same. The absolute amount of heat that needs to be dissipated from Samsung die is larger though, which has resulted in large and elaborate cooling solutions on the FE cards.
8nm had a higher yield and depreciation cost so it was much cheaper per chip than the 7nm (Speculation) This decision was made 2 years ago
This is comparing Samsung 8nm to Samsung 7nm. Samsung 7nm can be considered a 'broken' node.
Plus 8nm Samsung was available. AMD and all the other seized that silicon capacity early on.
Nvidia tried to bully TSMC into lowering prices but they didn't budge because demand for 7nm was already high. AMD took up a large part of the capacity that Nvidia was looking to get in the meantime.
 
Last edited:

Giroro

Splendid
256-bit memory bus sounds like memory bandwidth could be too bottle-necked to reach 3080 performance, even if they use best available GDDR6X.
But, then again The Xbox Series S is targeting 1440p120 using only 10GB total system memory... so maybe Navi2 is some miracle architecture with way better memory efficiency than anybody thought was possible.
 
256-bit memory bus sounds like memory bandwidth could be too bottle-necked to reach 3080 performance, even if they use best available GDDR6X.
But, then again The Xbox Series S is targeting 1440p120 using only 10GB total system memory... so maybe Navi2 is some miracle architecture with way better memory efficiency than anybody thought was possible.

It's the caching architecture. There is a very large cache on Navi. And to be honest this is the first step to MCM Navi chips that scale using a similar type of infinity fabric.


When you render to a small block, only a small subset of data is needed because blocks are relatively small. And you can only execute so many at a time. This limits the memory needed. BUT and this is a biggggggggg but, this is highly dependent on the look ahead predicting what memory is going to be needed for which tile and having it ready. That is some VERY tricky architectural challenges there. This might be possible as final texture render doesn't happen till later stages of the pipe. It could be easy to trip up the cache system. Textures are what eat up memory.
 
  • Like
Reactions: Giroro

Chung Leong

Reputable
Dec 6, 2019
493
193
4,860
But, then again The Xbox Series S is targeting 1440p120 using only 10GB total system memory... so maybe Navi2 is some miracle architecture with way better memory efficiency than anybody thought was possible.

The Series X's GPU uses a 320-bit memory bus. I don't think there's any miracle architecture. AMD is either using a new type of memory or it's going with a 512-bit bus. Inferring a 256-bit bus based on the number of texture caches seems suspect to me. Textures aren't the only major consumer of memory bandwidth on a chip that does ray-tracing.
 
Last edited:
The Series X's GPU uses a 320-bit memory bus. I don't think there's any miracle architecture. AMD is either using a new type of memory or it's going with a 512-bit bus. Inferring a 256-bit bus based on the number of texture caches seems suspect to me. Textures aren't the only major consumer of memory bandwidth on a chip that does ray-tracing.

That's part of the lighting/shading stage which is one of the last elements after mapping

If AMD operates anything like turing ray bounces will be calculated in parallel to shaders them combined. Will this is another buffer (z buffer, render buffer, lighting buffer, it isn't huge. Like I said render blocks are small.
 

MasterMadBones

Distinguished
256-bit memory bus sounds like memory bandwidth could be too bottle-necked to reach 3080 performance, even if they use best available GDDR6X.
But, then again The Xbox Series S is targeting 1440p120 using only 10GB total system memory... so maybe Navi2 is some miracle architecture with way better memory efficiency than anybody thought was possible.
There was a 384-bit version of Navi 21, but it was apparently cancelled. Performance between 256 and 384 appeared to the roughly the same, although it's likely that there are other differences between them.