• Competitor rules

    Please remember that any mention of competitors, hinting at competitors or offering to provide details of competitors will result in an account suspension. The full rules can be found under the 'Terms and Rules' link in the bottom right corner of your screen. Just don't mention competitors in any way, shape or form and you'll be OK.

AMD Polaris architecture – GCN 4.0

AMD release an 8 core Zen, with 2048 shader GPU onboard and 8gb HBM2 as shared memory... We can only dream.
 
By the looks of things, AMD shouldn't have really called this 'GCN4'. It looks like each core is much faster than the previous iterations, so they're not really comparable.

If it turns out Polaris 10 is 2304 cores and beats or matches the FuryX (Hitman demo they showed), then clearly Polaris 11 is not simply 'GCN4 Pitcairn'.
 
I think it's naive to think that after waiting all this time for the next gen process, all we're going to get is a slightly improved design in the core and some die shrink benefits. I expect both AMD and Nvidia to throw the kitchen sink at this generation and show what they can really do.
 
AllBodies said:
By the looks of things, AMD shouldn't have really called this 'GCN4'. It looks like each core is much faster than the previous iterations, so they're not really comparable.

If it turns out Polaris 10 is 2304 cores and beats or matches the FuryX (Hitman demo they showed), then clearly Polaris 11 is not simply 'GCN4 Pitcairn'.
Click to expand...

The 2304 part is not a full Polaris 10, it was marked with the value foe cut down Ellesmere. they more than likely decided 2816 and 2304 to differentiate them better, although that one could have been an early engineering test. Just considering Hawaii was cut down to 2500 something for the pro version.

Could be for better die recovery.

But it is essentially Pitcairn 2.0, same shader numbers regardless of them having better performance. :p

I reckon baffin xt will have 380x (stock) performance.
 
Last edited:
Mauller said:
From the looks of things, the current info is pointing ever closer to GCN 4 versions of Pitcairn, hawaii and fiji. Being Baffin/P11, Ellesmere/P10 and Greenland/V10 respectively. With their die sizes being ~120mm^2, ~232mm^2 and ~300-350mm^2 respectively.

although the last is my own estimate based on a slightly larger fiji die on 28nm. As a 4096 shader hawaii would be around 630-50mm^2 with full 1/2 DP, just making a rough estimate based on die area per shader.

Although it will more than likely change around the Ratio of Shaders, TMU's and ROP's again per CU.
Click to expand...


Be careful how you word this, don't use older architecture code name to describe GCN4, of course you know just because it may have a similar number of shaders does not indicate GCN4 has much if anything to do with the older architectures, especially given that the smallest of the GCN4 GPU's listed here scaled up to 28nm would be significantly larger than Pitcairn.
Some people will see the word "Pitcairn" and suddenly they become rebrands.
 
Last edited:
AllBodies said:
By the looks of things, AMD shouldn't have really called this 'GCN4'. It looks like each core is much faster than the previous iterations, so they're not really comparable.

If it turns out Polaris 10 is 2304 cores and beats or matches the FuryX (Hitman demo they showed), then clearly Polaris 11 is not simply 'GCN4 Pitcairn'.
Click to expand...

Thats why they call them GCN 4 and not GCN3+
No confusion at all.
 
was there a longer demo of hitman than the 5seconds they showed on stage?
did the press get to play on it or something?
 
If they're really doing all that with these shader counts, I wonder what kind of performance they're looking at with they just created a Hawaii size chip with around twice the cores (with the alleged 2.3x transistor density).

I wonder if they'll be able to keep pulling performance out of the bag in order the take whatever nvidia throw at them and milk the node for whatever number of years it will take to reach 10nm. (Considering Vega seemingly is supposed the provide yet more efficiency and performance from the architecture even ignoring a bigger chip size, and then even after that they could pull out a fiji size chip)

But I feel I've probably misunderstood and got it all wrong.
 
tavtavtav said:
If they're really doing all that with these shader counts, I wonder what kind of performance they're looking at with they just created a Hawaii size chip with around twice the cores (with the alleged 2.3x transistor density).

I wonder if they'll be able to keep pulling performance out of the bag in order the take whatever nvidia throw at them and milk the node for whatever number of years it will take to reach 10nm. (Considering Vega seemingly is supposed the provide yet more efficiency and performance from the architecture even ignoring a bigger chip size, and then even after that they could pull out a fiji size chip)

But I feel I've probably misunderstood and got it all wrong.
Click to expand...

Vega implemented patent seems likely
http://www.freepatentsonline.com/20160085551.pdf
 
tavtavtav said:
If they're really doing all that with these shader counts, I wonder what kind of performance they're looking at with they just created a Hawaii size chip with around twice the cores (with the alleged 2.3x transistor density).

I wonder if they'll be able to keep pulling performance out of the bag in order the take whatever nvidia throw at them and milk the node for whatever number of years it will take to reach 10nm. (Considering Vega seemingly is supposed the provide yet more efficiency and performance from the architecture even ignoring a bigger chip size, and then even after that they could pull out a fiji size chip)

But I feel I've probably misunderstood and got it all wrong.
Click to expand...

Since we saw Polaris 10 running faster than Fury X in Hitman@1440p, let's assume it was top end Polaris 10.
Now there are leaks that say mid end Polaris 10 comes with 2304 shaders, while top end is unknown, but let's assume top end p10 has same shader count as hawaii: 2816.
So optimistically p10 with same shader count as hawaii is faster than fury x, which is impressive jump in shader performance with GCN 4.0.
Now, looking at it from other side let's assume AMD OCed the **** out of that demo card to make it look impressive. It still gives us quite a jump in shader efficiency, since I don't know any OCed hawaii which could reach Fury X performance levels.
So if we take neutral POV, we can assume that p10 at stock clocks will be on par or a bit slower than fury x in performance and much more power efficient. Since Mr. Koduri said that they will price their new card very very competitively they shouldn't cost an arm and a leg.
I think AMD was bleeding money with fiji cards, and are looking to replace them sooner than later among other lower end cards.

Anyways, that's my theory :D
 
meade16 said:
I was ready to get a 980ti, I'll definitely be waiting.
Click to expand...

Better to buy stuff earlier rather than much later imo seeings as the 980Ti is not far off a year old now, I'll be waiting for the Pascal Ti at least before changing as the midrange level will likely be similar perf, or if better probably not by much at all.
 
Perfect_Chaos said:
Better to buy stuff earlier rather than much later imo seeings as the 980Ti is not far off a year old now, I'll be waiting for the Pascal Ti at least before changing as the midrange level will likely be similar perf, or if better probably not by much at all.
Click to expand...

I'm sure people can hold on from buying nvidia cards for few months, can they? The release, at least for AMD cards is not far off. Its not like everyone is waiting for fastest card. So close to release best advice would be to wait and see what new cards bring in. You can always get yourself that magical 980ti card later for even less money ;)
 
muziqaz said:
Since we saw Polaris 10 running faster than Fury X in Hitman@1440p, let's assume it was top end Polaris 10.
Now there are leaks that say mid end Polaris 10 comes with 2304 shaders, while top end is unknown, but let's assume top end p10 has same shader count as hawaii: 2816.
So optimistically p10 with same shader count as hawaii is faster than fury x, which is impressive jump in shader performance with GCN 4.0.
Now, looking at it from other side let's assume AMD OCed the **** out of that demo card to make it look impressive. It still gives us quite a jump in shader efficiency, since I don't know any OCed hawaii which could reach Fury X performance levels.
So if we take neutral POV, we can assume that p10 at stock clocks will be on par or a bit slower than fury x in performance and much more power efficient. Since Mr. Koduri said that they will price their new card very very competitively they shouldn't cost an arm and a leg.
I think AMD was bleeding money with fiji cards, and are looking to replace them sooner than later among other lower end cards.

Anyways, that's my theory :D
Click to expand...

TBH with the size of the card and heatsink, i don't think they would have overclocked it. Considering it was in a very small ITX case for the demo machines.

But Hawaii is already very close to Fiji in performance, the improvements in GCN 4 probably make P10/Ellesmere XT match it at stock.

Makes me wonder if they went back to GCN1-2 levels of ROPS and TMUS per shader, Since VEGA will more than likely have more than Fiji and should shoot ahead in gaming performance to where fiji should have been if we consider it to have same 8tflops as a base for compute performance.
 
You must log in or register to reply here.
Back
Top Bottom