Speculation on the next Navi configurations

neeyik

TS Evangelist
Staff member
Having just stuffed myself with food, and feeling to full to do any real work, I spent a few minutes thinking about where AMD can take Navi, in terms of the configurations (SPs, ROPs, etc). This led to a few minutes of setting it out in Excel, and voila:

https://imgur.com/a/zWqkGdN



All of this is, of course, not based on any fact, other than the overall Navi layout lends itself to be scaled quite nicely. The area and power values are total guesstimates.

Random speculation demands random thoughts and suggestions :)
 

neeyik

TS Evangelist
Staff member
The newly released RDNA whitepaper has some interesting bits of detail:
"Some variants of the dual compute unit expose additional mixed-precision dot-product modes in the ALUs, primarily for accelerating machine learning inference. A mixed-precision FMA dot2 will compute two half-precision multiplications and then add the results to a single-precision accumulator. For even greater throughput, some ALUs will support 8-bit integer dot4 operations and 4-bit dot8 operations, all of which use 32-bit accumulators to avoid any overflows.
This is clearly to compete against Nvidia's Turing GPU and its INT4/INT8 capability, but it doesn't look like there will anything with dedicate tensor units.

The SIMDs use separate execution units for double-precision data. Each implementation includes between two and sixteen double-precision pipelines that can perform FMAs and other FP operations, depending on the target market. As a result, the latency of double-precision wavefronts varies from as little as two cycles up to sixteen. The double-precision execution units operate separately from the main vector ALUs and can overlap execution."
Like with the Vega 20 Radeon Pro M150/M160 cards, the Navi Radeon Pro range will be up 1/2 FP64 rate (compared to the desktop and Vega 10/earlier Pros that were all 1/16 rate).
 

MattS

TS Evangelist
Honestly, I just hope anything which is released will be able to compete with next gen Nvidia cards or atleast the rtx 2080 ti.. Hopes up.
 

neeyik

TS Evangelist
Staff member

It's interesting to compare the Navi 14 die with the Navi 10 one:





The images suggest that Navi 14 has:

1 Shader Engine (vs 2 for Navi 10)
2 Asynchronous Compute Engines, ACEs (same as Navi 10)
6 Workgroup Processors, WGPs per ACE (vs 5 per ACE in Navi 10)
2 Compute Units per WGP (same as Navi 10)

So a full Navi 14 die would have 24 CUs, so there's scope for something like a 5500 XT. It also looks like the discrete 5500 has 88 TMUs (vs 160 for a full Navi 10) and 32 ROPs (half that of Navi 10).

Most of these specs put up against the likes of the GeForce GTX 1660/1650 - similar shaders and TMUs but less ROPs than the 1660.

The bad news is the TDP: 150W!

Either way, chalk up one speculation fail for me! :laughing:
 

neeyik

TS Evangelist
Staff member
Well there goes another speculation fail! There is indeed an RX 5500 XT model, but it's not a 24 CU processor. At least my 5600 XT is pretty close to the 5500 XT, in terms of power, die size, and a few other specs. The crystal ball ain't dead just yet!
 

jobeard

TS Ambassador
Sorry; just flat ignorant but WHAT is " Navi 14 "

Never seen or been "Navi". Your patience appreciated.