AMD: Navi Speculation, Rumours and Discussion [2019-2020]

Status
Not open for further replies.
Erinyes said:
Why do you say 5600XT is on the verge of losing performance due to B/W? The 5700 with 33% more B/W is only faster by approx 6-8%. If anything, it shows it has plenty of bandwidth.
Click to expand...
Because the plentifullness of RX5700 is diminished by 25%.
 
Erinyes said:
Which is even better right? It shows that even with all the added features, its still 1X power. Or are you implying something different?
Click to expand...

I havent listened to the hotchip presentation but Microsoft put the XBSX at 6x the raw performance of the XBO with the same power consumption of the XBOX. The SoC is limited by other components and units to use all these 12 TFLOPs to get to the 10x more performance. And when the SoC is limited power consumption will go down. Makes it more efficient than a discrete card with more bandwidth or even higher clocks.
 
troyan said:
The SoC is limited by other components and units to use all these 12 TFLOPs to get to the 10x more performance
Click to expand...
No lol it's statically power allocated.
You get 12TF under any condition (just like you get >900GFLOPS of CPU SIMD under any condition).

This isn't PS5 that uses SmartShift v2.
 
Considering the 5500Xt, 5600XT, 5700, 5700XT, there is evidence that Navi scales, but it is imbalanced somewhere. I.e. if we consider pure shading power, going from the 5500XT to the 5600XT there is a more than linear scaling: 1.31-1.38x the shading power results in average 1.45x more performance (5600XT vs 5500XT) but this is quite probably due to the geometry/ROP ratio (I think geometry is the reason there) that is in the 1.6-1.69x range. As told before by analyzing the 5600XT/5700 situation, bandwidth seems to play a lesser role there. So the performance gain stays in between these two numbers (shading/geometry ratios).

Note: Techpowerup numbers @1080p
 
Well this is why I compared lower end cards such as 5500XT-5600XT, judging by the behaviour of 5700-5700XT and NV cards, there is no CPU limitation at that resolution for these cards (this may be maybe an issue for 2080 and above class cards, even if they still show considerable gain). Also, a CPU limitation should show more on more powerful cards, while we have clearly a more-than-linear scaling regarding shading power, and almost a linear scaling with geometry power between 5600XT and 5700 and 5700XT. This seems not an indication about a CPU limitation.
 
Jawed said:
You should ask the people who claim it does.
Click to expand...

We'll you're the one who said there's no evidence that Navi scales. What evidence are you referring to exactly?
CarstenS said:
Because the plentifullness of RX5700 is diminished by 25%.
Click to expand...

Plentifulness? What does that mean? And that still dosen't in any way explain your claim that 5600XT is on the verge of losing performance due to B/W.
troyan said:
I havent listened to the hotchip presentation but Microsoft put the XBSX at 6x the raw performance of the XBO with the same power consumption of the XBOX. The SoC is limited by other components and units to use all these 12 TFLOPs to get to the 10x more performance. And when the SoC is limited power consumption will go down. Makes it more efficient than a discrete card with more bandwidth or even higher clocks.
Click to expand...

No idea what you're even basing this on. Suggest you read the hotchips presentation first. The GPU runs at a fixed, constant clock of 1.825 Ghz.
CarstenS said:
Try running FS2020 at ultra settings. Kinda not anymore.
Click to expand...

The discussion was pretty much on the TPU data. Does one single outlier of FS2020 even matter here?
Skinner said:
Maybe this has been explained earlier in this thread, but why wouldn't Navi21 use HBM2E?
Click to expand...

Mostly cost reasons. HBM is more expensive than GDDR and dosen't make a lot of sense in most consumer cards. Which is why you dont see Nvidia putting it in their halo cards either.
 
Erinyes said:
One would hope they could with Navi 3x. I'd hate to see them using vanilla G6 for the next gen.
Click to expand...
AMD's path to higher performing memory doesn't need to be the same as nvidia's.

HBM2 will be 4 years old next January. HBM2E is still coming and HBM3/HBMnext is slated for late 2022.
In the meanwhile, lots of different branches of research have been made to either replace interposers or make them significantly cheaper to produce.
 
Erinyes said:
We'll you're the one who said there's no evidence that Navi scales. What evidence are you referring to exactly?


Mostly cost reasons. HBM is more expensive than GDDR and dosen't make a lot of sense in most consumer cards. Which is why you dont see Nvidia putting it in their halo cards either.
Click to expand...

But they did in VII, and now that they maybe really need it in the fastest SKU, makes me wonder. Its also a bit more powerfriendly?
 
Skinner said:
but why wouldn't Navi21 use HBM2E?
Click to expand...
Because it's not really a net win until you plop 4+ stacks.
The fastest one is 2.54Gbps which isn't really *that* fast.
ToTTenTranz said:
In the meanwhile, lots of different branches of research have been made to either replace interposers or make them significantly cheaper to produce.
Click to expand...
Yeah. between CoWoS-STAR and CoWoS-R it's nicer all-around.
Skinner said:
But they did in VII, and now that they maybe really need it in the fastest SKU, makes me wonder.
Click to expand...
AMD is here to make money, not memes.
Skinner said:
Its also a bit more powerfriendly?
Click to expand...
You don't need it when you're ahead in efficiency rofl.
 
Question when 2 CUs can manage 256 FB32 opperations. That means for 80Cus 10.240 Operations. That means also AMD has now doubple FP32?

attachment.php
 
You mean RPM (rapid packed math), which they've had for a while now, since 2016 if not earlier, or something else?
 
Status
Not open for further replies.
Back
Top