TechpowerupPolaris 10@ 2048sp 5.5 TFLOP/sPolaris 11@896sp 2.5 TFLOP/s TDP 50W)

[csbin]

http://www.techpowerup.com/222450/more-polaris10-and-polaris11-specifications-revealed

Industry sources revealed to TechPowerUp some pretty interesting specifications of AMD's two upcoming GPUs based on the 4th generation Graphics CoreNext "Polaris" architecture. The company is preparing a performance-segment GPU and a mainstream one. It turns out, that the performance-segment chip, which the press has been referring to as "Ellesmere," could feature 32 compute units (CUs), and not the previously thought 40.

Assuming that each CU continues to consist of 64 stream processors (SP), you're looking at an SP count of 2,048. What's more, this chip is said to offer a single-precision floating point performance of 5.5 TFLOP/s, as claimed by AMD. To put this into perspective, the company had claimed 5.2 TFLOP/s for the "Hawaii"/"Grenada" based FirePro W9100, which launched earlier this February, and that SKU featured all 2,816 SP present on the chip. So this chip is definitely faster than most "Hawaii" based SKUs.

While "Hawaii" based SKUs feature TDP of no less than 250W, the new chip has a TDP rated no higher than 150W. AMD could pull off a "single 8-pin power connector" feat like NVIDIA, with quite some headroom to spare. The chip features a 256-bit wide GDDR5/GDDR5X memory interface, and 8 GB could be its standard memory amount. The first SKUs based on this chip could feature 7 Gbps GDDR5 memory.


AMD will upgrade the feature-set to include HVEC/H.265 hardware encode/decode acceleration, DisplayPort 1.3, and HDMI 2.0a outputs.


The smaller "Polaris" chip scheduled for 2016, which the press has been referring to as "Baffin," could feature 14 compute units, working out to a stream processor count of 896. It will be a mainstream chip, succeeding the "Tobago" silicon, which drives the current R7 360 series SKUs, although it wouldn't surprise us if it outperformed bigger chips, such as the "Trinidad" based R7 370 series. This chip has its peak single-precision floating-point performance rated at 2.5 TFLOP/s. Its TDP is rated at just 50W, and it is expected to feature a 128-bit wide GDDR5 memory interface, holding 4 GB of memory.

 

[Arachnotronic]

So Ellesmere should be about as fast as Hawaii but be significantly more efficient. I think this will make a lot of people happy.

 

[airfathaaaaa]

if anything i would be happy about the small one.. if true it has only 0.8 tflop less than 380 for 50w ......now that would be a pretty nifty little card for a lot of people

 

[railven]

Price will be the deciding factor, if you ask me.

I'm sure it will have improvements for DX12, but I can see a bunch of AMD buyers passing it over if it doesn't beat 390X handily. Hawaii like performance @ ~$200 is 2014/2015 numbers, if it's more than $250 it's definitely being skipped.

Might put weight to AMD trying to get Vega out faster.

 

[digitaldurandal]

I will be pleasantly surprised if true.. however, equal performance is not enough to get me to buy.

 

[lixlax]

To get 5,5TFLOP/s with 2048 SP's, the chip should operate at ~1,35GHz.

 

[Glo.]

Cut down versions - sure. Not the full parts.

P.S. P11 has 16 CU's at least. And that accounts for 1024 cores.

 

[airfathaaaaa]

dont forget about the launch of 4870.. the rumor was about a 480 shaders gpu while it came with 800...
(not saying it will happen again but given that we practicly have heard EVERYTHING something is bound to be right)

 

[xthetenth]

dont forget about the launch of 4870.. the rumor was about a 480 shaders gpu while it came with 800...
(not saying it will happen again but given that we practicly have heard EVERYTHING something is bound to be right)

Yeah. At this point it's like trying to navigate with a compass that has a disc for a needle. Sure it's pointing in the right direction but it's also pointing in every other direction and how do you know the difference?

 

[Head1985]

To get 5,5TFLOP/s with 2048 SP's, the chip should operate at ~1,35GHz.

And 2048sp is cut SKU.Full will be 2304SP or more..
2304SP at 1.4Ghz is 6.5Tflops..just like GTX1070.

 

[Krteq]

From VC

Our comment: We received the exact same specifications just few days ago. Although it was noted that those are specifications for mobile GPUs, so you may want to take this into consideration.

 

[Glo.]

http://forums.anandtech.com/showthread.php?t=2472871

 

[Vesku]

If these are mobile SKUs then AMD should be able to offer a P10 based desktop 1070 alternative. The speculated P11 SKU is pretty close to 7950 in TFlops.

 

[railven]

http://forums.anandtech.com/showthread.php?t=2472871

$170 is that for full Polaris 11 or cut down?

Unless AMD is going to stack these cards on top of each other, that would put (assuming this is Polaris 11 full) Cut P10 perhaps $230, and full P10 perhaps $300??

 

[Vesku]

$169-$229-$289/299?

Leaving this here to check against actual launch pricing.

 

[railven]

More pieces falling into place! Soon the fireworks will really start!

 

[MrTeal]

2048 seems a little low if the die size truely is 232mm². P10 to Hawaii is the same size ratio as GP104 to GM200, and nVidia packed in 83% of the CU's into GP104 that GM200 had while moving from a 384-bit to 256-bit bus. If the P10 is only 2048 SP's that 73% of Hawaii while moving from a 512-bit to 256-bit memory bus, that would be pretty poor in comparison.

Not impossible of course, but it does seem like pretty poor scaling especially given that Hawaii isn't a DP-stripped part to begin with.

 

[crisium]

Perhaps the 2048SP is the cutdown and the 2304SP is full. Of course if these are mobile, which can vary quite a bit and be further cutdown from desktop, there is still hope for a 2560SP Polaris.

"The first SKUs based on this chip could feature 7 Gbps GDDR5 memory." If true, another baffling decision by AMD. Nvidia chooses cutting edge 8 Gbps GDDR5 for the 1070 - nevermind the 10Gbps GDDR5X 1080. Does AMD desire slower performance? Going with 5/6 on the 290/390 was fine since it was 512-bit, but now AMD really needs bandwidth. It frustrates me to see them willingly choose inferior speeds. These are gonna be bandwidth starved. Even if AMD compression can match Nvidia's somehow, they are still looking at a bandwidth deficit to the 1070 for no other reason than deliberate choice. If true, I am disappoint. Although this is an improvement over the 6 Gbps rumours which would truly make AMD idiotic.

 

[coercitiv]

nVidia packed in 83% of the CU's into GP104 that GM200 had while moving from a 384-bit to 256-bit bus.

If the P10 is only 2048 SP's that 73% of Hawaii while moving from a 512-bit to 256-bit memory bus.

This also ignores the small yet notable density advantage of "14nm" over "16nm".

 

[Flapdrol1337]

Perhaps the 2048SP is the cutdown and the 2304SP is full. Of course if these are mobile, which can vary quite a bit and be further cutdown from desktop, there is still hope for a 2560SP Polaris.

"The first SKUs based on this chip could feature 7 Gbps GDDR5 memory." If true, another baffling decision by AMD. Nvidia chooses cutting edge 8 Gbps GDDR5 for the 1070 - nevermind the 10Gbps GDDR5X 1080. Does AMD desire slower performance? Going with 5/6 on the 290/390 was fine since it was 512-bit, but now AMD really needs bandwidth. It frustrates me to see them willingly choose inferior speeds. These are gonna be bandwidth starved. Even if AMD compression can match Nvidia's somehow, they are still looking at a bandwidth deficit to the 1070 for no other reason than deliberate choice. If true, I am disappoint. Although this is an improvement over the 6 Gbps rumours which would truly make AMD idiotic.

It doesn't matter if it's a smaller and cheaper chip.

 

[Glo.]

One more thing, isn't 896 GCN core GPU the same as M470X? And it also has around 50W TDP.

I smell that someone is trolling hard.

 

[kraatus77]

2048 seems a little low if the die size truely is 232mm². P10 to Hawaii is the same size ratio as GP104 to GM200, and nVidia packed in 83% of the CU's into GP104 that GM200 had while moving from a 384-bit to 256-bit bus. If the P10 is only 2048 SP's that 73% of Hawaii while moving from a 512-bit to 256-bit memory bus, that would be pretty poor in comparison.

Not impossible of course, but it does seem like pretty poor scaling especially given that Hawaii isn't a DP-stripped part to begin with.

Because it's not full p10 part imo, that will be 2560 shaders same as gp104 but in much smaller package. don't forget it will have ACEs too which nv still lacks thus saving them some mm^2 space.

 

[sm625]

Polaris SPs could each contain many more transistors vs Hawaii SPs.

 

[dark zero]

INBF the presented Polaris 11 is the R5 440 and the Polaris 10 is R7 460 and Vega 10 is R9 490 onwards. No rebrands available.

But on mobile Polaris 11 could be R7 460M and Polaris 10 R8 480M. I feel that Vega 10 will have a Mobile versión too taking the spot of 490M.

Also.. Expecting to see Fiji reducing prices a LOT.

 

[crisium]

It doesn't matter if it's a smaller and cheaper chip.

It's still less performance, even ignoring competition. 14% more bandwidth always translates into at least some more performance, period, regardless of compression. Perhaps it's outrageously expensive, but even then I'd think you'd want performance to look as good as possible relative to the 1070 regardless of price. Consider it a pet peeve of mine when the AMD team agrees to make a chip slower than possible regardless of the other rational behind it. Obviously 8Gbps is there in the wild since Nvidia is using it on 1070, so it is possible.

Maybe they will surprise me. But these are my reactions to the current rumour.