WirkoSome of the cache has evaporated overnight, or something...

https://twitter.com/i/web/status/1610974871575744514

ToTTenTranzOn one hand it's strange that they'd make such a big mistake on something like an official presentation slide, and then repeat that same mistake across all 3 models.
On the other hand, that slide only mentions "total cache", and doesn't specify whether it's L3+L2+L1 or something else.

At the same time, they haven't really said much about that RDNA3 iGPU with 6 WGPs running at ~3GHz (despite being the highest-clocked RDNA3 GPU to date and the only one at N4).
The Radeon 780M seems to have >150% of the processing throughput of the 680M in Rembrandt while the total RAM bandwidth only goes up from 6400MT/s to 7500MT/s.
This is only a 17% boost which is hardly enough to keep up with the faster CPU cores which will undoubtedly demand more bandwidth by themselves.

One thing that could make sense is that the new RDNA3 iGPU actually has 16MB of Infinity Cache all to itself. It would explain the 40MB "total cache" typo, and provide an adequate boost to effective bandwidth to the substantially more powerful iGPU.

Or that really was just a mistake in those slides and this iGPU is going to be massively bottlenecked by memory bandwidth.

MawkzinLooking to the Ryzen 5 7640 the igpu has 8 CU this time instead of 6 CU, so maybe with rdna3 it uses 3 wgp with 4 CU each instead of 2 wgp with 6 CU each in the rembrand series.

TumbleGeorgeMaybe we has been a some difference between models for China and for...outer world where part of them will be artificially limited.


PS. I have an idea why limitations. Because official support of more RAM, maybe even more cache from Windows is more expensive. Recently, in China, they are betting on their own operating systems...

ToTTenTranzOn one hand it's strange that they'd make such a big mistake on something like an official presentation slide, and then repeat that same mistake across all 3 models.
On the other hand, that slide only mentions "total cache", and doesn't specify whether it's L3+L2+L1 or something else.

At the same time, they haven't really said much about that RDNA3 iGPU with 6 WGPs running at ~3GHz (despite being the highest-clocked RDNA3 GPU to date and the only one at N4).
The Radeon 780M seems to have >150% of the processing throughput of the 680M in Rembrandt while the total RAM bandwidth only goes up from 6400MT/s to 7500MT/s.
This is only a 17% boost which is hardly enough to keep up with the faster CPU cores which will undoubtedly demand more bandwidth by themselves.

One thing that could make sense is that the new RDNA3 iGPU actually has 16MB of Infinity Cache all to itself. It would explain the 40MB "total cache" typo, and provide an adequate boost to effective bandwidth to the substantially more powerful iGPU.

Or that really was just a mistake in those slides and this iGPU is going to be massively bottlenecked by memory bandwidth.

THANATOSWGP is a Dual Compute unit(CU). Phoenix has 6 WGPs or in other words 6 dual compute units, people like to say 12CU. There is no difference compared to Rembrandt.
There should be 2 shader engines(arrays) with 3 WGPs per shader engine, just like what you see in that block diagram. Ryzen 5 7640 will lose a WGP per shader engine, which leaves 4WGPs or 8CUs.
It looks like in Rembrandt the 6CU version has half of shader engines deactivated, that's why only 1/2 of CU is left.


Phoenix has only 16MB of L3 cache physically, the same as Rembrandt.
There is no artificial difference based on region. AMD marketing team did a poor job by not noticing they wrote the wrong amount of cache in some press slides, which caused this mess about L3 cache.


There was no mention anywhere about Phoenix having Infinity cache. 40MB is a typo.

That processing throughput increase is in reality only 25% higher (3GHz vs 2.4GHz). The rest of that increase comes from dual issue CUs, and for that to work, you need VOPD instructions and even then you can't use It for everything, there are limitations and also shaders need to be optimized for It. Even AMD says RDNA3 architecture is only 17.4% faster per clock than RDNA2. There is a reason why they say N31 has only 6144 stream processors, which is only 20% more than Navi21.
BTW, If AMD really couldn't meet the required BW, then they wouldn't use 6 WGPs, If they performed so much better.

Chip & Cheese made a great RDNA3 architectural analysis. Go, check It out.