AMD Polaris architecture to succeed Graphics Core Next

SneakyStephan said:
Yeah the midrange market is going to get absolutely flooded with cheap second hand gtx 970s (mine probably one of them, depending on how 1070 benchmarks turn out and the price for an aftermarket cooler one in my country)

Amd are going to have to compete with those because a 150 dollar second hand 970 is pretty appealing if you didn't have a decent gpu yet. (they're already 220 dollars new I saw in the other thread) and probably a lot less than what amd is going to charge for a 390/970 equivalent polaris 10 card
Click to expand...


Dunno about that, only 3.5 GB of (normal speed) memory on those 970s is going to be a put off going forward. But if you can get one for $150 this will appeal to the budget buyer and lift up the baseline.

I'd like to think there will be a Polaris 10 GPU that is significantly faster than a 970 for $300-350. If it's just 390/X performance again this will do nothing in the market.

I'm thinking Polaris 10 = 480 ($250) and 480X ($300).
Polaris 11 = 470, 470X.
 
Papacheeks said:
Do you know if it is Bogus?
Click to expand...

It pretty much has to be. AIB partners would have to be packing up cards by now, so it would be crazy for tit to be delayed this late in the game. Also, they can't hit 850MHz? Really? Frankly, if it turns out to be true, AMD needs to give up and sell all of their IP to the highest bidder. That would be beyond pathetic.
 
dr_rus said:
Rumor: Polaris validation Failed Might Launch in October Now
Well I certainly hope that this is bogus.
Click to expand...

michael-scott-no.gif
 
Coulomb_Barrier said:
Dunno about that, only 3.5 GB of (normal speed) memory on those 970s is going to be a put off going forward. But if you can get one for $150 this will appeal to the budget buyer and lift up the baseline.

I'd like to think there will be a Polaris 10 GPU that is significantly faster than a 970 for $300-350. If it's just 390/X performance again this will do nothing in the market.

I'm thinking Polaris 10 = 480 ($250) and 480X ($300).
Polaris 11 = 470, 470X.
Click to expand...

I've got a buddy with SLI 970s trying to sell them now.

He wants $200 as a friend price, as a friend I told him I'll be able get a 1060ti with at-least-as-good performance and 6Gb of full speed ram for $250 in the fall.

For $150 it's a decent enough upgrade from my 760 though.
 
dr_rus said:
Well, yeah, perf/watt means that you get the same perf for different watt and that's what this roadmap is showing as presumably the only reason Vega is above Polaris there is because of HBM being more energy efficient.

Let's wait for Polaris reveal at least. What I expect from it is Hawaii's performance for 2x less power consumption - and that would be 2x perf/watt.
Click to expand...

You have got that backwards. It's like bhp/litre. If you increase that by 20% it doesn't mean you stick with 200 bhp and make the engine smaller it means your 2l engine can now produce 240 bhp.
 
dr_rus said:
Another issue is that direct comparisons between Polaris and big Vega (since there won't be a big Polaris) may not be possible due to a memory subsystem differences.
Click to expand...
What we do know is that the Rx300 GPUs w/ HBM have MUCH greater efficiencies than the lower-tier offerings -- perf/watt actually goes up as you increase mm^2.
 
E-Cat said:
What we do know is that the Rx300 GPUs w/ HBM have MUCH greater efficiencies than the lower-tier offerings -- perf/watt actually goes up as you increase mm^2.
Click to expand...

This is likely why Vega is in a higher perf/watt category than Polaris on AMDs roadmap.
 
icecold1983 said:
i hope polaris isnt having troubles. its in everyones best interest for AMDs chip to be as good as possible
Click to expand...
If they are still using GF for Polaris and TSMC for Vega, they have much higher chance for setbacks than Nvidia.

The whole AMD rumor mill is weird af since Jaguar APU is manufactured by TSMC but Neo is rumored to be Polaris.
 
Adry9 said:
Regular pricing.

480: $250
390X: $429
970: $309
980: $480
980Ti: $595
1080: $599
Click to expand...

Apparently prices have changed over the last few days. The 980 is now as low as $459 and the Ti $609, so the first one should be slightly better value instead.
 
Inspectah_Deck said:
Pascal is GCN-like: What this means for Maxwell, Gaming & VR:
https://www.youtube.com/watch?v=Qc4RTyGbPv8

Can´t watch currently, because I´m at work, but it should be interesting.

Would be more fitting for the Nvidia thread, but I think the discussion in this thread will be more productive.
Click to expand...

hes not correct. pascal is highly similar to maxwell. pascal still cant perform async compute in the manner amd does because there is no hardware scheduler and it still doesnt seem that SMMs can run graphics and compute tasks simultaneously. hes also making a large leap about the "improved preemeption". the data says compute tasks can be preempted at an instruction level but does not mention whether graphics tasks can be. the latter is the big one. we also dont know if gp104 inherited even the former from gp100.
 
Inspectah_Deck said:
Pascal is GCN-like: What this means for Maxwell, Gaming & VR:
https://www.youtube.com/watch?v=Qc4RTyGbPv8

Can´t watch currently, because I´m at work, but it should be interesting.

Would be more fitting for the Nvidia thread, but I think the discussion in this thread will be more productive.
Click to expand...

Yeah, Pascal isn't GCN like in pretty much anything. The misconception seems to stem from the same number of FP32 units per SM in GP100 as in GCN's CU but it's both a) doesn't mean much and b) not true for GP10x chips.
 
TechPowerUp has posted that their "industry sources" have revealed following Polaris stats:

Polaris 10 - 32 CUs, 5.5Tflops, 8 gigs of ram, no higher than 150W
Polaris 11 - 14 CUs, 2.5 Tflops, 4 gigs of ram, 50W
http://www.techpowerup.com/222450/more-polaris10-and-polaris11-specifications-revealed


32 CUs for P10 is a surprising stat, either AMD is facing yield issues or they plan to release more pricey unlocked chips down the line.



edit - videocardz.com say that they heard the same info, but it was mentioned in the mobile GPU talk
http://videocardz.com/59903/possible-polaris-10-and-polaris-11-specifications-emerge
 
DieH@rd said:
TechPowerUp has posted that their "industry sources" have revealed following Polaris stats:

Polaris 10 - 32 CUs, 5.5Tflops, 8 gigs of ram, no higher than 150W
Polaris 11 - 14 CUs, 2.5 Tflops, 4 gigs of ram, 50W
http://www.techpowerup.com/222450/more-polaris10-and-polaris11-specifications-revealed


32 CUs for P10 is a surprising stat, either AMD is facing yield issues or they plan to release more pricey unlocked chips down the line.



edit - videocardz.com say that they heard the same info, but it was mentioned in the mobile GPU talk
http://videocardz.com/59903/possible-polaris-10-and-polaris-11-specifications-emerge
Click to expand...

One of the posts in the comment lined things up well for what we know so far.

We are here:
370 - 950
380 - 960
390(x) - 970/980
Fury Nano - 980
Fury X - 980Ti/TX

We are moving here:
470 - ?
480 (5.5Tf) - ? ($299?)
? - 1070 (6.5Tf) $379 - $449+
? - 1080 (9tf) $599 - $699+
? - ?
Click to expand...

I made a few corrections.
 
DieH@rd said:
TechPowerUp has posted that their "industry sources" have revealed following Polaris stats:

Polaris 10 - 32 CUs, 5.5Tflops, 8 gigs of ram, no higher than 150W
Polaris 11 - 14 CUs, 2.5 Tflops, 4 gigs of ram, 50W
http://www.techpowerup.com/222450/more-polaris10-and-polaris11-specifications-revealed


32 CUs for P10 is a surprising stat, either AMD is facing yield issues or they plan to release more pricey unlocked chips down the line.



edit - videocardz.com say that they heard the same info, but it was mentioned in the mobile GPU talk
http://videocardz.com/59903/possible-polaris-10-and-polaris-11-specifications-emerge
Click to expand...

If those P10 specs where the 480M maybe que can expect a 480 faster than a 390X. Right?
 
DieH@rd said:
TechPowerUp has posted that their "industry sources" have revealed following Polaris stats:

Polaris 10 - 32 CUs, 5.5Tflops, 8 gigs of ram, no higher than 150W
Polaris 11 - 14 CUs, 2.5 Tflops, 4 gigs of ram, 50W
http://www.techpowerup.com/222450/more-polaris10-and-polaris11-specifications-revealed


32 CUs for P10 is a surprising stat, either AMD is facing yield issues or they plan to release more pricey unlocked chips down the line.



edit - videocardz.com say that they heard the same info, but it was mentioned in the mobile GPU talk
http://videocardz.com/59903/possible-polaris-10-and-polaris-11-specifications-emerge
Click to expand...

These actually seems quite possible. At 5.5TF P10 is unlikely to reach 1070 while a rumored 1060Ti card will be a nice fit as it's direct competition.

The interesting part is the lower P10 model which should end up somewhere around 4TF probably which falls around 380X performance meaning that it'll probably go against GTX960 until GP106 will arrive - and that's where NV won't have much to answer for some months.
 
DieH@rd said:
TechPowerUp has posted that their "industry sources" have revealed following Polaris stats:

Polaris 10 - 32 CUs, 5.5Tflops, 8 gigs of ram, no higher than 150W
Polaris 11 - 14 CUs, 2.5 Tflops, 4 gigs of ram, 50W
http://www.techpowerup.com/222450/more-polaris10-and-polaris11-specifications-revealed


32 CUs for P10 is a surprising stat, either AMD is facing yield issues or they plan to release more pricey unlocked chips down the line.



edit - videocardz.com say that they heard the same info, but it was mentioned in the mobile GPU talk
http://videocardz.com/59903/possible-polaris-10-and-polaris-11-specifications-emerge
Click to expand...


While I prefer AMD, I really think I'm just going to get a 1080 and wait and see on Vega10. I figure I'll be able to sell my 1080 for a good price in 6-10 months anyhow.
 
Hopefully desktop Polaris 10 has the full rumored 40 CU, also according to previous rumors and the CU count rumor of today, only the top two cards (M485X,M470X) should be Polaris and everything else just rebrands.
 
chaosblade said:
Neat. Wonder what it means for the desktop cards.

CBX0IXX.png
Click to expand...

Would be pretty strange to put up a mobile line-up consisting of GPUs that aren't officialy announced yet.

Going by those specs, you'd think they're rebrands. Bonaire has 14 CUs and Tonga has 32. Usually mobile part numbers are higher than their desktop versions, so P11 could end being M480 and P10 M490 series.
 
Renekton said:
As I suspected, AMD is prioritizing an Apple deal to put their power-efficient Polaris into iMacs and MBPs.

Us enthusiasts have to wait.

Edit: mobility gamers will rejoice, this is an underappreciated market
Click to expand...

Doesn't look like there are any Polaris GPUs there at the moment. The lineup is a bucnh of rebrands.

The naming is kinda interesting though as there's a gap between M470X and M485X - this is where P11 will go most likely. And M490/X is where P10 is supposed to be it seems.

Confirmed as rebrands: AMD Quietly Unveils Radeon M400 Series: Starting With Rebadges
 
dr_rus said:
Yeah, Pascal isn't GCN like in pretty much anything. The misconception seems to stem from the same number of FP32 units per SM in GP100 as in GCN's CU but it's both a) doesn't mean much and b) not true for GP10x chips.
Click to expand...
Looking at the ALU-Clusters Pascal resembles GCN the most:

Fermi: 32 ALUs, 128 KB Vector Register (4 KB per ALU), 4 TMUs (8:1 ratio), 64 KB shared-memory (one portion splitted as L1$)

Kepler: 192 ALUs, 256 KB Vector Register (1,3 KB per ALU), 16 TMUs (12:1 ratio) 64 KB shared-memory (one portion splitted as L1$)

Maxwell v2: 128 ALUs, 256 KB Vector Register (2 KB per ALU), 8 TMUs (16:1 ratio), 96 KB shared-memory (no splitt anymore)
2048 active threads.

Pascal GP100: 64 ALUs, 256 KB Vector Register (4 KB per ALU), 4 TMUs (16:1 ratio), 64 KB shared-memory (no split)
2048 active threads.

GCN: 64 ALUs, 256 KB Vector Register (4 KB per ALU), 4 TMUs (16:1 ratio), 64 KB shared-memory (no split)
2560 active threads.


The GP100 has 6 GPCs, with 10 SMs each = 6 (GPC) x 10 (SM) x 64 (ALUs) = 3840 ALUs.
For the GP104 you would need 4 GPC = 4 (GPC) x 10 (SM) x 64 (ALUs)= 2560 ALUs.
It resolves perfectly so I doubt that Nvidia would change the arrangement.

Edit: Speaking of the devil:
https://compubench.com/device.jsp?benchmark=compu15d&os=Windows&api=cl&D=NVIDIA+GeForce+GTX+1080&testgroup=info

If the recognition is right then the GP104 has 20 SMs in total which would mean 128 ALUs per SM.
So this would very much look like Maxwell.
 
dr_rus said:
Doesn't look like there are any Polaris GPUs there at the moment. The lineup is a bucnh of rebrands.

The naming is kinda interesting though as there's a gap between M470X and M485X - this is where P11 will go most likely. And M490/X is where P10 is supposed to be it seems.

Confirmed as rebrands: AMD Quietly Unveils Radeon M400 Series: Starting With Rebadges
Click to expand...
Yup fk me sideways. I just checked the hardware subreddit, you are right.
 
Locuza said:
Looking at the ALU-Clusters Pascal resembles GCN the most:

Fermi: 32 ALUs, 128 KB Vector Register (4 KB per ALU), 4 TMUs (8:1 ratio), 64 KB shared-memory (one portion splitted as L1$)

Kepler: 192 ALUs, 256 KB Vector Register (1,3 KB per ALU), 16 TMUs (12:1 ratio) 64 KB shared-memory (one portion splitted as L1$)

Maxwell v2: 128 ALUs, 256 KB Vector Register (2 KB per ALU), 8 TMUs (16:1 ratio), 96 KB shared-memory (no splitt anymore)
2048 active threads.

Pascal GP100: 64 ALUs, 256 KB Vector Register (4 KB per ALU), 4 TMUs (16:1 ratio), 64 KB shared-memory (no split)
2048 active threads.

GCN: 64 ALUs, 256 KB Vector Register (4 KB per ALU), 4 TMUs (16:1 ratio), 64 KB shared-memory (no split)
2560 active threads.


The GP100 has 6 GPCs, with 10 SMs each = 6 (GPC) x 10 (SM) x 64 (ALUs) = 3840 ALUs.
For the GP104 you would need 4 GPC = 4 (GPC) x 10 (SM) x 64 (ALUs)= 2560 ALUs.
It resolves perfectly so I doubt that Nvidia would change the arrangement.

Edit: Speaking of the devil:
https://compubench.com/device.jsp?benchmark=compu15d&os=Windows&api=cl&D=NVIDIA+GeForce+GTX+1080&testgroup=info

If the recognition is right then the GP104 has 20 SMs in total which would mean 128 ALUs per SM.
So this would very much look like Maxwell.
Click to expand...

Yeah, as I've said - it both doesn't mean much (there are way more important stuff than ALU numbers in an SM) and is wrong for GP10x.
 
dr_rus said:
Yeah, as I've said - it both doesn't mean much (there are way more important stuff than ALU numbers in an SM) and is wrong for GP10x.
Click to expand...

theres speculation on b3d that gp104 might not have inherited the architecture changes shown in the gp100 whitepaper. it might still be 128 alus per sm like maxwell. only a few more days till we know all
 
icecold1983 said:
theres speculation on b3d that gp104 might not have inherited the architecture changes shown in the gp100 whitepaper. it might still be 128 alus per sm like maxwell. only a few more days till we know all
Click to expand...

I think I've said as much here well before any speculation happened on B3D -)
 
icecold1983 said:
pretty disappointing if true.
Click to expand...

Not really. Maxwell's balance was well suited for gaming loads and it makes sense to keep it in the absence of bigger under-the-hood changes. GP100 SM would have a lot of idling silicon in a modern game workload and since it's still very much split between graphics and compute it probably wouldn't do any better in future games either.
 
dr_rus said:
Not really. Maxwell's balance was well suited for gaming loads and it makes sense to keep it in the absence of bigger under-the-hood changes. GP100 SM would have a lot of idling silicon in a modern game workload and since it's still very much split between graphics and compute it probably wouldn't do any better in future games either.
Click to expand...

the increased amt of registers and cache would have been very beneficial
 
You must log in or register to reply here.

Similar threads

MiSTer FPGA Sega Saturn core hits update_all; basically a finished core
19
126
_Ex_ replied
Want to melt your brand new PC with next gen graphics? Nvidia released Zorah Tech Demo!
47
4K
cinnamonandgravy replied
PS6 Rumors Suggest Console Will Use AMD's UDNA Architecture
Rumor  2 3
145
11K
Fake replied
  • Poll
Death Stranding 2 graphics thread - Next-gen doesn’t start until Kojima says so
2 3 4 5 6
287
21K
Larxia replied
Intel Computex 2024 Architecture Thread. Lunar Lake. Arc Battlemage. Lion Cove (P-Core) Skymont (E-Core)
30
3K
winjer replied
Top Bottom