Polaris 10 and 11 confirmed to be GDDR5 based

[Blitzvogel]

Exactly. I live in Texas.

Psssh Texas isn't that bad. If you can't afford AC, maybe you shouldn't be stretching your dollar to afford PC gaming :whiste:

 

[3DVagabond]

Psssh Texas isn't that bad. If you can't afford AC, maybe you shouldn't be stretching your dollar to afford PC gaming :whiste:

not sure if serious?

 

[raghu78]

So all the talk about bringing VR down in cost and giving gamers a lot of perf/$ isn't just hype. Ofc as soon as Raja said that, it should have been plain obvious that Polaris is meant to be cheap and hence, only GDDR5.

I think I understand what's going on when they demo Hitman 1440p, outside looking into the boat scene with a ton of NPCs hidden.

That people, is the Primitive Discard Accelerator in action. On 390X, that scene will drop below 60 fps easy (I know, I have the game). But Polaris doesn't even take in those geometry into it's pipeline to process and then Z-cull, it discards it before rendering even occurs. This results is huge improvements to minimum FPS when the scene complexity is bottlenecking the engine.

It may well only be 390X class performance normally, but it will perform better at MIN FPS in GPU bound scenarios. That's what they played, those tricksters at AMD...

Just like their perf/w comparison at 60 fps vsync lock. They would have massively improved power gating and so if the GPU isn't running max load, power usage drop a lot.

However the curious thing is why there's so few SP, 232mm2 Polaris 10 with 14ff density can easily fit much more than that unless they have really changed the design so that each SP takes many more transistors.

Raja Koduri has already stated that shading efficiency and throughput has increased significantly.

http://venturebeat.com/2016/01/15/a...-to-full-graphics-immersion-with-16k-screens/

I am quite optimistic about the full Polaris 10 SKU (2560 sp) beating Fury X.

 

[Blitzvogel]

not sure if serious?

I am serious. I live in Texas myself.

 

[SPBHM]

I am serious. I live in Texas myself.

maybe electricity is super cheap there, but in a lot of places the cost of running an AC is pretty high.

 

[3DVagabond]

I am serious. I live in Texas myself.

Well, I lived in Tx. and when the outside temp is north of 100° extra heat sources aren't really a good thing.

 

[IEC]

maybe electricity is super cheap there, but in a lot of places the cost of running an AC is pretty high.

Electricity is super cheap in parts of Texas. And a newer home with proper insulation and a heat pump means the extra cost of running the AC is pretty negligible.

 

[MrTeal]

Folks on the high end aren't the target for Polaris 10 to upgrade. It's a very small two chip, 120mm2 and 232mm2, these classes are entry level and low-midrange.

Typically in the past, mid-range chips were ~300mm2 or even larger.

Folks on high-end 28nm, will have Vega 10 and 11 to look forward to.

Vega 11 should be the upper-midrange, ~400mm2 (my guess). Vega 10 the biggest one.

You think Vega 11 will be that large? If P10 does slot in at 232 mm², I would expect Vega 11 to come in at ~325-350, with Vega 10 being roughly Hawaii-sized. 400 on the small one and 500+ on the big one just seems too big too early, as well as leaving a pretty massive gap in die sizes and performance. The HBM2 controller in Vega should be physically smaller than P10 as well, so 400mm253 might actually get you close to double the shaders of P10.

For the people declaring 2304 shaders is a cut down die (some with a lot of conviction) is there any actual indication of that, or is it just speculation?

 

[IEC]

You think Vega 11 will be that large? If P10 does slot in at 232 mm², I would expect Vega 11 to come in at ~325-350, with Vega 10 being roughly Hawaii-sized. 400 on the small one and 500+ on the big one just seems too big too early, as well as leaving a pretty massive gap in die sizes and performance. The HBM2 controller in Vega should be physically smaller than P10 as well, so 400mm253 might actually get you close to double the shaders of P10.

For the people declaring 2304 shaders is a cut down die (some with a lot of conviction) is there any actual indication of that, or is it just speculation?

Well, it's speculative, but 2304 would be an odd number to be a "full" die.

 

[MrTeal]

Well, it's speculative, but 2304 would be an odd number to be a "full" die.

I suppose, though no moreso than 11CU in each of 4 shader engines to give 44CU and 2816 shaders for Hawaii.

 

[Blitzvogel]

Well, I lived in Tx. and when the outside temp is north of 100° extra heat sources aren't really a good thing.

I'll be honest, running the computer was never really the problem for me. It was the 32 inch LCD HDTV I used as a monitor in a generally small bedroom with the door closed to prevent noise propagation (I had two other roommates) that was an issue to a degree. Sucker gets quite warm, which was nice in the winter, but certainly an issue in the summer months. I was also in the south corner of the house which got hit by sun from sunrise to sunset.

 

[itsmydamnation]

I suppose, though no moreso than 11CU in each of 4 shader engines to give 44CU and 2816 shaders for Hawaii.

it also means they didn't do that great a job at scaling down on GF/sammy 16nm which form all accounts ahs about a 10% advantage in that area vs TSMC.

A Hawaii with only 256 of memory interface, 2304 shaders and reduced L2 would be around 310mm. If you took that and go to 230mm thats only a 26% reduction as a straight shrink, so unless the architectural improvements cost a serious amount of transistors (like 2 billion) it would seem like Polaris 10 has to have more CU's then 36.

 

[SPBHM]

I'll be honest, running the computer was never really the problem for me. It was the 32 inch LCD HDTV I used as a monitor in a generally small bedroom with the door closed to prevent noise propagation (I had two other roommates) that was an issue to a degree. Sucker gets quite warm, which was nice in the winter, but certainly an issue in the summer months. I was also in the south corner of the house which got hit by sun from sunrise to sunset.

a TV like that is probably under or around 200W, which should be lower than a gaming PC

 

[Blitzvogel]

a TV like that is probably under or around 200W, which should be lower than a gaming PC

My PC at the time (Phenom II x4 + Radeon 5850) never felt like it was putting out too much heat, even when gaming. The TV on the other hand (still have it), feels quite warm, especially on the backside after it's been running a short while. I'll try to remember to check it's power supply/consumption properties on the back panel tomorrow when my wife isn't asleep, as it's in our bedroom.

 

[poofyhairguy]

Psssh Texas isn't that bad. If you can't afford AC, maybe you shouldn't be stretching your dollar to afford PC gaming :whiste:

The AC is already being pushed by the 65 inch plasma I play on.

 

[ShintaiDK]

Even if you dont live in a hot place. When you live in a BR2015 home (hair away from passive home). You also care about heat generation

http://bygningsreglementet.dk/english/0/40

 

[JDG1980]

http://ranker.sisoftware.net/show_r...e0d2e3d4e7d0e8cebc81b197f297aa9abccff2ca&l=en

Polaris 10 - Ellesmere

36-40CU, 2304-2560 sp, 256 bit GDDR5 @ 6Ghz , 8GB.

I don't see how Polaris 10 could avoid being bottlenecked by a 192 GB/sec memory bus. One possibility that comes to mind is that the card actually uses GDDR5X at a 12Ghz effective rate, but SiSoft doesn't recognize that memory type yet so it's interpreting it as GDDR5 at half the clock rate. 384 GB/sec would make a lot more sense for a card with Hawaii to Fiji level performance. It's also possible that the memory controller is designed for GDDR5X but the engineering sample shown here had to use GDDR5 due to lack of availability at this time.

 

[Head1985]

Vega 10 (Hawaii replacement) ~ GP100
Vega 11 (Tahiti replacement) ~ GP104
Polaris 10 (Pitcairn replacement) ~ GP106
Polaris 11 (Cape Verde replacement) ~ GP108

 

[MangoX]

So P10 and P11 are looking to be much more about energy efficiency than pure brute power.

If we could get 390X performance at 100 watts that would be awesome.

 

[itsmydamnation]

I don't see how Polaris 10 could avoid being bottlenecked by a 192 GB/sec memory bus. One possibility that comes to mind is that the card actually uses GDDR5X at a 12Ghz effective rate, but SiSoft doesn't recognize that memory type yet so it's interpreting it as GDDR5 at half the clock rate. 384 GB/sec would make a lot more sense for a card with Hawaii to Fiji level performance. It's also possible that the memory controller is designed for GDDR5X but the engineering sample shown here had to use GDDR5 due to lack of availability at this time.

tonga only uses 256bit bus, so increase in GDDR clock + improvements to memory compression to handle 25% more shaders.

 

[Kuiva maa]

I don't expect big Polaris to beat Fiji with such a small die-unless we are all wrong about its size . I feel AMD will replace its entire range besides fury/fury X/nano now with Polaris, while making a push to mobile at the same time. The question is whether big Polaris will be fast enough to be considered by Hawaii users that want to upgrade to 14nm.

 

[Head1985]

polaris 10 is just 7870 replacement.
polaris 11 is ultra low-end
I really dont know why the hell people expect hawaii replacement from 230mm2 SKU?Its just 7870 in 14nm for god sake.

If you want hawaii/Tahity replacement wait for 2017 and vega with HBM2.
Vega 10 (Hawaii replacement) ~ GP100
Vega 11 (Tahiti replacement) ~ GP104
Polaris 10 (Pitcairn replacement) ~ GP106 - around GTX980 performance much less TDP
Polaris 11 (Cape Verde replacement) ~ GP108 - around GTX960 much less TDP

 

[Mercennarius]

maybe electricity is super cheap there, but in a lot of places the cost of running an AC is pretty high.

Energy is pretty cheap in most of the state. Where I live my 2,400sqft home's energy bill is usually under $100 a month. And I use plenty of energy (hot tub, A/C on most of the year, plenty of electronics on all the time, etc.).

 

[Paratus]

I think some of you have touched on a concern I have about the move from 28nm to first gen 16/14nm.

Since both vendors were forced to push their high end chips to the 600mm^2 reticule limit and they expect to be on 16/14nm for an extended length of time, (like 28nm) I don't think we are going to see amazing performance upgrades out of the gate.

Look at the previous transition: 40nm - 28nm

• Cayman (40nm) / 2.64B Xtors / 389mm^2 / 6.8m Xtors/mm^2
• Tahiti (28nm) / 4.31B Xtors / 365 mm^2 / 11.8m Xtors/mm^2

We got a 74% increase in transistor density.

Now let's look at the 28nm progression.

• Tahiti / 4.31B Xtors / 365 mm^2 / 11.8m Xtors/mm^2
• Hawaii / 6.2B Xtors / 438 mm^2 / 14.1m Xtors/mm^2
• Fiji / 8.9B Xtors / 596mm^2 / 14.9m Xtors/mm^2

By the end of the 28nm era AMD was getting 120% the Xtors density as 40nm. We can't assume they will have the new process that optimized right off the bat.

If we assume a similar 74% increase in density going from 28nm to 16/14nm a Fiji class card would be at least 341mm^2. To get a significant increase in Xtors and performance, (more than just new architecture and a clock bump) means we need something the size of Hawaii on the new process.

I'm concerned this means we won't really see a good performance increase until the second round of chips. Performance per watt and per mm^2 should still go through the roof though.

 

[ShintaiDK]

You cant really use 40->28nm compare. Since that transition gave cheaper transistors. 28nm->14/16nm is same cost at best. But its quite obvious that high end cards are for 2017.