AMD CDNA3 Architecture Sees the Inevitable Fusion of Compute Units and x86 CPU at Massive Scale

[btarunr]

AMD in its 2022 Financial Analyst Day presentation unveiled its next-generation CDNA3 compute architecture, which will see something we've been expecting for a while—a compute accelerator that has a large number of compute units for scalar processing, and a large number of x86-64 CPU cores based on some future "Zen" microarchitecture, onto a single package. The presence of CPU cores on the package would eliminate the need for the system to have an EPYC or Xeon processor at its head, and clusters of Instinct CDNA3 processors could run themselves without the need for a CPU and its system memory.

The Instinct CDNA3 processor will feature an advanced packaging technology that brings various IP blocks together as chiplets, each based on a node most economical to it, without compromising on its function. The package features stacked HBM memory, and this memory is shared not just by the compute units and x86 cores, but also forms part of large shared memory pools accessible across packages. 4th Generation Infinity Fabric ties it all together.



AMD is claiming a 500% (or 5 times) AI compute performance/Watt uplift over CDNA2, thanks to the combination of 5 nm processor for the compute dies, an advanced 3D chiplet packaging technology, 4th Gen Infinity Fabric, new math computing formats, Infinity Cache on the compute dies, and a unified memory architecture. The company is working toward a 2023 debut of CDNA3.

View at TechPowerUp Main Site

 

[Mathragh]

"The Future Is Fusion" finally coming to complete fruition!

 

[R0H1T]

Only a decade late!

 

[Mathragh]

Only a decade late!

Well, better late than never
And certainly better than not having tried at all.

 

[Anc13ntEvil]

"AMD is claiming a 500% (or 5 times) AI compute performance/Watt uplift over CDNA2"

I don't understand why so many people don't understand how percentages compared to multiplication works...

500% is SIX TIMES, not five. If you have $10 and I increase it by 100%, you would then have $20. 100% is double, 200% is triple, and so on, yet I constantly see people get this simple math wrong. Sorry not sorry but this annoys the crap out of me.

It also bugs the crap out of me when companies list something like "1.3 times the speed" (or power).
30%...just say 30%.

 

[DeathtoGnomes]

"AMD is claiming a 500% (or 5 times) AI compute performance/Watt uplift over CDNA2"

I don't understand why so many people don't understand how percentages compared to multiplication works...

500% is SIX TIMES, not five. If you have $10 and I increase it by 100%, you would then have $20. 100% is double, 200% is triple, and so on, yet I constantly see people get this simple math wrong. Sorry not sorry but this annoys the crap out of me.

It also bugs the crap out of me when companies list something like "1.3 times the speed" (or power).
30%...just say 30%.

All compute starts at 0 not 1, its pretty binary.

 

[InVasMani]

Me from two days ago..."If AMD had a 5600X with larger 3D stacked cache or also paired APU chiplet that didn't castrate PCIE lanes in the process it would be pretty popular for a budget AM4 build. In fact the larger cache could be on the APU and if the chip could utilize the cache between either with maybe a slight latency penalty worst case scenario in doing so it would be really nice. Both CPU/GPU would have a set amount of cache that's equal and at the same time can tapped into by the other in a pinch with only a bit more added latency when and where needed."

Seems AMD thought similar, but I like that they took it steps further with HBM and Zen 4 is of course a step up for such a APU designed in tandem with 3D stacked cache. It seems they went with the pooled and shared unified memory as well! I'm not sure what the new "MATHS" is, but suspect it's referring to AVX related and/or FP precision stuff bit o this bit o that!

 

[R-T-B]

"AMD is claiming a 500% (or 5 times) AI compute performance/Watt uplift over CDNA2"

I don't understand why so many people don't understand how percentages compared to multiplication works...

500% is SIX TIMES, not five. If you have $10 and I increase it by 100%, you would then have $20. 100% is double, 200% is triple, and so on, yet I constantly see people get this simple math wrong. Sorry not sorry but this annoys the crap out of me.

It also bugs the crap out of me when companies list something like "1.3 times the speed" (or power).
30%...just say 30%.

That's not how "uplift" works.

A 100% uplift would be double the performance.

100% of the performance would indeed be the same.

 

[eidairaman1]

"The Future Is Fusion" finally coming to complete fruition!

SoCA, blade servers/workstations

All compute starts at 0 not 1, its pretty binary.

MACHINE CODE DUNANANA

 

[Minus Infinity]

That's not how "uplift" works.

A 100% uplift would be double the performance.

100% of the performance would indeed be the same.

Yes the subtlety of one single word and you are indeed correct.

 

[ModEl4]

I wonder if the work that it's done on the CDNA3 unified memory architecture will bring some benefits for future Zen5 based APUs on the memory controller/ cache/ V-cache system integration or everything is sorted out in the APU space with HUMA later iterations.

 

[Imaamdfanboy]

Nope...500% is indeed 5 times more.
If you have two apples for comparison they are both equall.Lets says they are both exactly same size then they are 1 to 1.This is why they do the 1.5x or maybe 2.3x explanation these days.The one's are equalls so the points after is the presentation.So 1.5 will then be 50 present faster or if you like half more times faster than one.If you use presentation it not like normal maths witch comes down to 1*1 = 1.It excaly 0.
So 2.3 times wil hê 130% môre which relates to 1.3 times more powerful.