Pemalite said:
This is false.
For doubling of CU's that would mean 36 CU's x2 = 72 CU's.
Graphics Core Next tends to have a hard limit at 64 CU's.
And a doubling of CU's for only a 1.35x performance boost? Ouch. Vega 7 actually has less CU's than Vega 64 and offered a substantial performance increase for roughly the same TDP.
As for TDP's: https://www.anandtech.com/show/10663/analyzing-sonys-playstation-4-pro-announcement
But TDP isn't actually power consumption...
https://www.youtube.com/watch?time_continue=13&v=0wNoCnPxTp4
There are some serious technical limitations we need to consider as well.
It's not just about size and cost.
Vega 7 is 331mm2.
Vega 64 is 486mm2.
Correct. But consoles are also at the mercy of PC commodity prices... That is, when DRAM goes up in price... Then that price also increases for consoles.
Conversely... Because Microsoft and Sony are buying fab capacity from the likes of companies such as TSMC... When there isn't much spare capacity, then TSMC can charge an extra premium.
I would give more credit to Anandtech when it comes to low-level stuff like that.
At the moment, 64 CU's is a GCN hard limit. - Is it entirely impossible to circumvent? Of course not.
But by going past that limit, inefficiencies come into play... And to get around those inefficiencies, some serious re-engineering work needs to be done to Graphics Core Next at a very low level, something AMD is not likely to spend significant time and money on to achieve when their next-generation GPU architecture is coming next year.
I.E. What is possible is mirroring the GPU and using an interconnect for communication. - The chip clusters themselves are still limited to 64CU's though.
It would be no different than getting two Vega 64 GPU's and running them in Crossfire. - You technically have 128 CU's to play with. - But each chip is still 64 CU.
Slightly ahead of the more sensible Anaconda specs.
There are 64 Stream processors per Compute Unit. That is a hard limit.
There is a maximum of 64 Compute Units. That is a hard limit.
There is a maximum of 64 Render Output Pipelines. That is a hard limit.
There is a maximum of 256 Texture Mapping units. That is a hard limit.
There is a maximum of 4 Geometry units. That is a hard limit.
There is a reason why Vega doesn't exceed those limits that were hit by Fiji many years earlier.
Navi's main focus will likely be extracting as much as it can from GDDR6 and implementing/improving on a few new tricks introduced with Vega... And hitting the Polaris price point/TDP.
It's a Polaris replacement... And not a high-end GPU, it's main focus is going to be efficiency.
Vega 7 is 7nm and didn't double anything except on the memory side of the equation.
Or... They simply might go with GDDR5X or GDDR5 which if taken wide enough can be faster than GDDR6 anyway.
All comes down to price/capacity in the end, it's a balance the console manufacturers will need to weigh.
Although, by the time 2020 rolls around, GDDR6 should be very price competitive even with GDDR5 and certainly have a capacity advantage.
|
The TDP i mentioned is 330 a typo it's the same number that mentioned on Anand tech (310w)
Ok, but we don't have an exact architecture lay out of Navi, AMD might have something on their sleeve like more Rop's on single CU. It's the same design GCN but it's still "Navi" a next gen mainstream GPU to replacing Polaris. Even PS4 to PS4 Pro can have a graphic boost by 2,5 times more.
what's the point of 7nm if there is no performance gain. You seem pessimist, while historical data pointed different way.
If you " theory " is true, PS5 performance will not have a generational leap, it will be even smaller leap compared to PS3 to PS4 gap, while PS3 to PS4 already considered small.
If you said using Zen2 is big improvement , you said on others thread, it's not. CPU is not affecting rendered pipeline directly, it might rid of bottle neck but it will not affecting the shaders count etc.
You also need to remember what happen with Polaris , it's able to gain more performance compared to previous Pitcairn mainstream GPU and comparable to Hawaii high end GPU or even to Nvidia Maxwell (GTX 970- GTX 980).
I believe we are on the times where 2016 high end GPU will be a mainstream GPU like what happen with polaris. So 14 teraflop is achievable .Hell even 2016 AMD able to bring 12,5 teraflop with vega 64.
Last edited by HollyGamer - on 12 March 2019
