Polaris 10 and 11 confirmed to be GDDR5 based

[JDG1980]

ARM confirmed it not too long ago when they were promoting new 28 nm designs. This was in 2015.

ARM has to come up with some marketing excuse why they're using an obsolete process. I'd be more interested to hear from someone who doesn't have a vested interest in what they're saying. Any recent, impartial sources saying FinFET is still substantially more expensive than 28nm was at the same point in its production cycle?

 

[raghu78]

If we ignore core clock speed.

HD6970=176GB/sec
HD7870=154GB/sec

390X=384GB/sec
Polaris 10=192GB/sec (If leaks are true)

If Polaris 10 got GDDR5X, then its a whole other discussion.

R9 380X (Tonga) showed that it could match R9 280X (Tahiti) with just 63% of the bandwidth.

https://www.techpowerup.com/reviews/ASUS/R9_380X_Strix/

With Polaris we are talking about
1. New Command Processor
2. New Geometry Processor (with new Primitive Discard Accelerator)
3. Improved 4th gen GCN shaders/CUs
4. Improved memory compression
5. New memory controller.
6. New L2 cache.

http://www.anandtech.com/show/9886/amd-reveals-polaris-gpu-architecture/2

We are looking at even better memory compression combined with memory controller and L2 cache improvements. I think AMD knows better than any of the people on this forum how to address the bottlenecks and build a well balanced GPU. I am thinking Polaris 10 is going to surprise us. Anyway I am willing to wait and see instead of writing it off.

 

[jpiniero]

It's not substantially more expensive, otherwise neither company would bother and there would be another round of 28 nm GPUs. But it's only competitive on a per transistor basis, and obviously a 16FF/14FF die is going to be much smaller.

 

[airfathaaaaa]

how do we end up talking about core clocks while discussing memory bandwidth bottlenecks?
also why we keep saying "if" polaris gets gddr5x? they develop the damn memory they will have it not on polaris 11 obviously

 

[Qwertilot]

That reasoning doesn't really hold up - 20nm was an 'ignorable' advance, but these nodes are offering a ~2 fold power reduction, so at this point both NV and AMD are basically looking at die shrinking or entirely shuttering their GPU businesses.

They'd only both not die shrink now if it wasn't possible for a monopoly on the entire dGPU market to profitably support doing so. Some way off that I'd think.

All ARM doing 28nm designs proves is that 28nm is still cheaper (no surprise!) - they do designs for a lot of different markets, some very cost sensitive.

The if for GDDRX is based on timing - it won't be ready if they're launching this stuff soon.

 

[Hidemind]

Polaris 11 and 10 could replace all current graphics cards from $100 to $650 including Fury-X and especially Nano.

Certainly not.

P10 lacks the bandwidth in order to compete with Fiji, as well it lacks the raw fillrates.

P10 with a 256 bit memory interface, 6 Ghz GDDR5, more L2-Cache and improved DCC, will never reach the effective bandwidth of Fiji.

You should all reduce your expectations for Polaris in general. Do you really think, that a 36 CU P10 just slams a 64 CU Fiji? Even if the Graphics Command Processor, the Geometry Engines and the Primitive Engines got a huge update, P10 will be most likely match Hawaii. Sry, folks, Fiji will be here for a while, or do you think, that Fiji Gemini was introduced in order to sink into obscurity right away?

Those guys who think P10 exceeds Fiji really don't have a clue.

Ah, yes, the same goes for P11. P11 will most likely be a Pitcairn/Tonga successor. I doubt P11 reaches Tonga XT, maybe it can compete with Tonga Pro, but don't expect too much again.

Maybe the general efficiency will reach Maxwell level, meaning the a 36 CU P10 should overcome GM204 by a small margin. A 16 CU P11 could be a good Pitcairn and Tonga Pro replacement, but nothing more.

And there are some hints regarding the target clock speeds in the leaked drivers. Expect clock speeds ~1 Ghz. It looks like the underlying architecture for the CUs was trimmed for power efficiency and minor driver ignoring changes. And some low-level power gating will be introduced, according to the drivers.

/* This function is for Polaris11 only for now,
* Powerplay will only control the static per CU Power Gating.
* Dynamic per CU Power Gating will be done in gfx.
*/

 

[MrTeal]

This is a lot speculation based solely on the speeds reported by SiSoft which are notoriously inaccurate. We had the same thing in a previous thread when someone was positing a 3GB, 1GHz, 2048-bit three stack Polaris based on a Sandra entry, despite the fact that the same MB on the same day posted the same memory config with the description of being Fury. Hell, when I run Sandra benches on my system with 290s it not only reports the wrong memory speed, it says they're DDR3. Those are GPUs that have been out for years. I put zero faith in SiSoft's reporting of GPU features.

 

[Hidemind]

This is a lot speculation based solely on the speeds reported by SiSoft which are notoriously inaccurate. We had the same thing in a previous thread when someone was positing a 3GB, 1GHz, 2048-bit three stack Polaris based on a Sandra entry, despite the fact that the same MB on the same day posted the same memory config with the description of being Fury. Hell, when I run Sandra benches on my system with 290s it not only reports the wrong memory speed, it says they're DDR3. Those are GPUs that have been out for years. I put zero faith in SiSoft's reporting of GPU features.

I would never do speculations based on those unreliable sisoft numbers.
Tonga shows some impressive efficiency gains in comparison to Tahiti according to newer games. Take "The Division" and Hitman, where Tonga sometimes achieves more than 20% higher framerates. Tonga has a pretty impressive FPS/GFLOPs ratio, which comes close to Maxwell. Maybe Polaris achieves another 10~20% efficiency gain? A 36 CU P10 should compete pretty nicely with a GM204 and Hawaii.

But I would never expect Fiji performance levels. In order to reach Fiji, P10 definitily needs more bandwidth, which only GDDR5X could provide. So a GDDR5 6Ghz P10 shoudl remain at ~Hawaii level, with GDDR5X maybe a little bit above.

 

[poofyhairguy]

The only people interested in that performance range are AMD users and 970 owners.

Aka the largest group of gamers on Steam. The 970 is the sweet spot, where the consumer GPU money is made.

 

[Adored]

Aka the largest group of gamers on Steam. The 970 is the sweet spot, where the consumer GPU money is made.

Yes but not at $500-$600, which is the kind of price Nvidia needs to charge in order to keep their finances going they way they have been going.

 

[msi2]

Yes but not at $500-$600, which is the kind of price Nvidia needs to charge in order to keep their finances going they way they have been going.

Are the 970 at 500-600$?

 

[Adored]

Are the 970 at 500-600$?

That was my point. Yes there are a bunch of people with 970's, but they won't be upgrading to this or any card at $500.

970 owners will upgrade to a faster card near the price point they paid.

 

[3DVagabond]

You would pay over $500 for 15%-20% over a 980? Surely these people already own 980's...

Look at all of the GM104 cards that sell. All they offer over the previous gen is efficiency.

 

[AtenRa]

Cayman (HD6970) at 40nm = 389mm2 with 2.62B transistors

Transistor per mm2 = 6,73

Pitcairn XT (HD7870) at 28nm = 212mm2 with 2.8B transistors.

Transistor per mm2 = 13,2

Hawaii (R9 290X) at 28nm = 438mm2 with 6.2B transistors.

Tranistor per mm2 = 14,15

Now Polaris 10 possible die size of 232mm2 at 14nm.
14nm Density over 28nm is close to 2.2x

That will put the Polaris 10 die size between 6 and 7 Bil transistors at 232mm2.
This will essentially make Polaris 10 a Hawaii equivalent from a transistor count.

Add all the new architectural upgrades and I strongly believe Polaris 10 could be very close to Fiji (At current performance) at 1080p and perhaps 1440p. Fiji should distance it self at 4K though.

 

[3DVagabond]

980 Ti is probably more like 25-30% over 980 vanilla.

In any case, I recall Robert Hallock saying their new uarch should be compatible with both GDDR5 and HBM. Do we know if GDDR5X is compatible with vanilla GDDR5 memory controllers? I'm assuming it would be.

This is just a conspiracy theory on my part, but this is how I see it:

Polaris 11 = small GPU + GDDR5
Polaris 10 = medium GPU + GDDR5
Vega 11 = Polaris 10 rebrand with HBM
Vega 10 = proper Fiji successor

Take Hallock with a grain of salt when he's answering live questions. If he doesn't know he will often just reply with what he thinks or assumes.

Between 390 and 390X performance from Polaris 10 sounds about right. AMD would be cannibalizing Nano/Fury/Fury X sales if Polaris 10 was faster, so I doubt that will be the case.

They need to be more concerned with nVidia than Fiji.

They take advantage of Ethereum mining and they'll sell every one they make.

 

[raghu78]

Cayman (HD6970) at 40nm = 389mm2 with 2.62B transistors

Transistor per mm2 = 6,73

Pitcairn XT (HD7870) at 28nm = 212mm2 with 2.8B transistors.

Transistor per mm2 = 13,2

Hawaii (R9 290X) at 28nm = 438mm2 with 6.2B transistors.

Tranistor per mm2 = 14,15

Now Polaris 10 possible die size of 232mm2 at 14nm.
14nm Density over 28nm is close to 2.2x

That will put the Polaris 10 die size between 6 and 7 Bil transistors at 232mm2.
This will essentially make Polaris 10 a Hawaii equivalent from a transistor count.

Add all the new architectural upgrades and I strongly believe Polaris 10 could be very close to Fiji (At current performance) at 1080p and perhaps 1440p. Fiji should distance it self at 4K though.

I think 7 billion transistors is a good estimate. I am of the opinion the architectural improvements are going to be very significant and we will see a big increase in perf/sp and perf/CU. I think Polaris 10 will be faster than Fury X for 1080p/1440p. At 4k I can see bandwidth holding it back and Fury X being the faster chip.

 

[Blitzvogel]

I'm guessing a cut down Polaris 11 will be the new and badly needed sans-PCIE powerconnector king. AFAIK the Radeon 7750 (and it's "newer" derivatives) are still the fastest sans-PCIE graphics cards AMD has to offer, which is crazy since they are pathetic in comparison to the non-externally powered GTX 950s and even GTX 750s to a fair degree. Wife's PC has a 7750 in it, and it's not that bad at all, but I feel the itch to upgrade it, even if she never uses the damn thing

The faster AMD get's there stuff out, the better. The drivers need to be spot on and coming out fast too. Been somewhat interested in starting a build-to-order gaming PC business, and new AMD graphics that are competitive with Nvidia would certainly be a good thing to offer customers.

 

[USER8000]

I'm guessing a cut down Polaris 11 will be the new and badly needed sans-PCIE powerconnector king. AFAIK the Radeon 7750 (and it's "newer" derivatives) are still the fastest sans-PCIE graphics cards AMD has to offer, which is crazy since they are pathetic in comparison to the non-externally powered GTX 950s and even GTX 750s to a fair degree. Wife's PC has a 7750 in it, and it's not that bad at all, but I feel the itch to upgrade it, even if she never uses the damn thing

The faster AMD get's there stuff out, the better. The drivers need to be spot on and coming out fast too. Been somewhat interested in starting a build-to-order gaming PC business, and new AMD graphics that are competitive with Nvidia would certainly be a good thing to offer customers.

The R7 360E is the fastest card AMD have to offer in that bracket - from what I gather due to the lower clockspeeds it is in-between an HD7770 and an HD7790 and has around the same power consumption as the GTX750.

The fastest bus power card from AMD was this HD7850:

http://www.techpowerup.com/185003/triplex-shows-off-slot-powered-radeon-hd-7850.html

Edit to post

Actually,I was wrong,the HIS R7 360E seems to be similar in performance to a GTX750:

http://www.madshrimps.be/articles/a...oler-OC-2GB-Video-Card-Review/7#axzz443EgliNG

 

[jpiniero]

The faster AMD get's there stuff out, the better. The drivers need to be spot on and coming out fast too. Been somewhat interested in starting a build-to-order gaming PC business, and new AMD graphics that are competitive with Nvidia would certainly be a good thing to offer customers.

This is totally off topic, but what's with people thinking they can make money on selling white boxes? This isn't the early 90s.