[Ars] AMD confirms high-end Polaris GPU will be released in 2016

[littleg]

16 million Scoville units is equal to pure capsaicin. My guess is that they're hinting at the name of the next card.

The AMD Radeon 8-methyl-N-vanillyl-6-nonenamide. Doesn't exactly trip off the tongue

 

[Glo.]

That's all a hell of a stretch. 8-methyl-N-vanillyl-6-nonenamide is capsaicin, the active ingredient in hot sauce. Pure cap is 16M scoville units, which is the defacto "Heat Rating" for peppers. Raja and Chris Hook have been tweeting back and forth about hot sauce for over a month.
https://twitter.com/GChip/status/684603003307859968

About the other points
1. All the 232mm² came from a linked in profile, which while not nothing is a pretty flimsy evidence, especially considering it's not linked to any GPU in particular. It's possible they will launch a GPU at around that size though.
2. See above. 8.6 doesn't mean anything. They might launch a 4096 core GPU (though @ 232mm² would be really, really good scaling), but it has nothing to do with hot sauce.
3. SiSoft is just always wrong, in its memory configurations, as discussed before. Fury's regularly show up with that same memory config.

http://ranker.sisoftware.net/top_ru...c7facbed85b888aed6ebdafc99fcc1f1d7a499a9&l=en
Reading anything into them is silly. Besides, there is absolutely zero chance of AMD shipping a >Fiji class GPU with 3GB of VRAM with some goofy 2GB+1GB asymmetric topology.

Why do you think, you have to explain to me all this? I take everything of this into consideration, believe me.

 

[3DVagabond]

The AMD Radeon 8-methyl-N-vanillyl-6-nonenamide. Doesn't exactly trip off the tongue

Google searching for the card would be hilarious though.

 

[MrTeal]

The AMD Radeon 8-methyl-N-vanillyl-6-nonenamide. Doesn't exactly trip off the tongue

Heh, pure cap isn't something you want on your tongue in any form.

 

[swilli89]

Heh, pure cap isn't something you want on your tongue in any form.

Actually pure cap on the tongue can be an enlightening experience

 

[3DVagabond]

The AMD Radeon 8-methyl-N-vanillyl-6-nonenamide. Doesn't exactly trip off the tongue

Engineer humor. lol

 

[MrTeal]

Why do you think, you have to explain to me all this? I take everything of this into consideration, believe me.

Because these things tend to get repeated as fact. Just look at all the people in this thread stating that Polaris 11 is a 232mm² die.

 

[3DVagabond]

Because these things tend to get repeated as fact. Just look at all the people in this thread stating that Polaris 11 is a 232mm² die.

There are always going to be people who lack reading comprehension. Trust me most of us realize this is all guesswork based on the interpretation of the leaks/rumors we're seeing.

 

[tential]

There are always going to be people who lack reading comprehension. Trust me most of us realize this is all guesswork based on the interpretation of the leaks/rumors we're seeing.

Not to mention the explanation helps other people too.

 

[JDG1980]

Because these things tend to get repeated as fact. Just look at all the people in this thread stating that Polaris 11 is a 232mm² die.

Sure, it's possible that the LinkedIn guy could be lying, or that this is just an internal lab project that never goes public, but it's not the way to bet. 232mm^2 is a fairly normal size for a low-midrange GPU die. That's slightly bigger than Pitcairn (212mm^2) and on par with Nvidia's GM206 (227mm^2). It's safe to say that AMD will release a FinFET GPU in this size range - if not in 2016, then next year.

That's the big question here. We know one of AMD's new GPUs, Polaris 10, will be fairly small; Anandtech's Ryan Smith estimated it was a bit smaller than Cape Verde and GK107. Giving some room for error, let's say somewhere between 100-130mm^2, with the lower to mid portion of the range being more likely. Polaris 11, on the other hand, is still unknown. Raja Koduri has indicated that it's meant to take back the high end from Nvidia, which would seem to indicate a bigger chip than 232mm^2. A 14nm FinFET chip that size could conceivably beat GTX 980 Ti if it hits the jackpot on clock rate boosts and architectural improvements, but it would be a stretch. If AMD wants to win this round (especially once Nvidia has GP104 ready), it needs a new chip that is around the 350mm^2 size range. So the potential options seem to be something like this:

• Small chip (Polaris 10, ~120mm^2) and mid-size chip (Polaris 11, 232mm^2) coming in 2016. Performance for the small chip is Pitcairn-Tonga range depending on TDP, and for the large chip, it tops out around Fiji/GM200 levels. The focus would then be on perf/watt and winning laptop and AIO contracts. Gaming enthusiasts would have to wait until 2017 for the big GP104-fighter. (One other drawback: this would leave AMD short-handed in the high-end FirePro department as well).
• Small chip (Polaris 10, ~120mm^2) and large chip (Polaris 11, ~350mm^2) coming in 2016. In this scenario, Polaris 11 would handily outclass anything on 28nm from either vendor, and at least trade blows with GP104. It would also probably have strong Double Precision support, and serve as AMD's flagship chip in FirePro cards. The only problem with this scenario would be that it leaves the mid-range devoid of options. We might see a three-tier salvage die scenario, which happened with Tahiti and GK104 near the beginning of the 28nm node. Presumably, in this scenario, the 232mm^2 chip would be coming in 2017 to fill in the midrange gap.
• The final possibility is that the statements from Raja Koduri were misinterpreted, and there are going to be three new GPUs this year, not two. Presumably two would be Polaris 10 and 11, and the third might be the rumored Vega. That would correspond with the fact that we've known about three "Arctic Islands" codenames - Baffin, Ellesmere, and Greenland - for quite some time now. A three-GPU lineup would allow AMD to cover all bases, but I'm concerned that this is just wishful thinking.

 

[MrTeal]

Sure, it's possible that the LinkedIn guy could be lying, or that this is just an internal lab project that never goes public, but it's not the way to bet. 232mm^2 is a fairly normal size for a low-midrange GPU die. That's slightly bigger than Pitcairn (212mm^2) and on par with Nvidia's GM206 (227mm^2). It's safe to say that AMD will release a FinFET GPU in this size range - if not in 2016, then next year.

That's the big question here. We know one of AMD's new GPUs, Polaris 10, will be fairly small; Anandtech's Ryan Smith estimated it was a bit smaller than Cape Verde and GK107. Giving some room for error, let's say somewhere between 100-130mm^2, with the lower to mid portion of the range being more likely. Polaris 11, on the other hand, is still unknown. Raja Koduri has indicated that it's meant to take back the high end from Nvidia, which would seem to indicate a bigger chip than 232mm^2. A 14nm FinFET chip that size could conceivably beat GTX 980 Ti if it hits the jackpot on clock rate boosts and architectural improvements, but it would be a stretch. If AMD wants to win this round (especially once Nvidia has GP104 ready), it needs a new chip that is around the 350mm^2 size range. So the potential options seem to be something like this:

• Small chip (Polaris 10, ~120mm^2) and mid-size chip (Polaris 11, 232mm^2) coming in 2016. Performance for the small chip is Pitcairn-Tonga range depending on TDP, and for the large chip, it tops out around Fiji/GM200 levels. The focus would then be on perf/watt and winning laptop and AIO contracts. Gaming enthusiasts would have to wait until 2017 for the big GP104-fighter. (One other drawback: this would leave AMD short-handed in the high-end FirePro department as well).
• Small chip (Polaris 10, ~120mm^2) and large chip (Polaris 11, ~350mm^2) coming in 2016. In this scenario, Polaris 11 would handily outclass anything on 28nm from either vendor, and at least trade blows with GP104. It would also probably have strong Double Precision support, and serve as AMD's flagship chip in FirePro cards. The only problem with this scenario would be that it leaves the mid-range devoid of options. We might see a three-tier salvage die scenario, which happened with Tahiti and GK104 near the beginning of the 28nm node. Presumably, in this scenario, the 232mm^2 chip would be coming in 2017 to fill in the midrange gap.
• The final possibility is that the statements from Raja Koduri were misinterpreted, and there are going to be three new GPUs this year, not two. Presumably two would be Polaris 10 and 11, and the third might be the rumored Vega. That would correspond with the fact that we've known about three "Arctic Islands" codenames - Baffin, Ellesmere, and Greenland - for quite some time now. A three-GPU lineup would allow AMD to cover all bases, but I'm concerned that this is just wishful thinking.

For sure, AMD might well launch a 232mm² die. As you say, it's not at all unlikely. That's basically the size of Pitcairn and GK106. There's nothing even reasonably concrete on that though, and you listed a very possible scenario where the 232mm² die could actually be the final size but it would end up in a product we see launch a bit later similar to how Pitcairn launched after Tahiti and Cape Verde. It's pretty much exactly what nVidia did last round where GK106 launched six months after GK104. It would also fit better with launching a GDDR5 P10 and a HBM2-based larger die, following with a midsized GDDR5X part as availability improves towards the end of 2016. I'm probably in the minority here, but I remain unconvinced we'll see HBM2 down to 232mm² die sizes.

Who knows though, it could all be my internal rationalizing, since while I love my 290CF, the idea of a 100 and 232mm² dies on AMD's side and an end of year GP104 launch is the stuff of nightmares for me.

 

[raghu78]

Something which is a distinct possibility is misinterpretation of Raja Koduri statements.

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

"Yes. We have two versions of these FinFET GPUs. Both are extremely power efficient. This is Polaris 10 and that’s Polaris 11. In terms of what we’ve done at the high level, it’s our most revolutionary jump in performance so far."

Raja states two versions of these FINFET GPUs. Why wouldn't he say two 14nm FINFET GPUs. So it could mean two types of designs - Polaris 10 for 14LPP entry/mid-range products with GDDR5 and Polaris 11 for 16FF+ flagship products with HBM2. TSMC has the better yields and 16FF+ has a 6 month TTM lead over 14LPP which again helps in volume ramp and yield learning. Moreover TSMC has experience with large GPUs for many process generations.

So it all comes down to what was meant by Raja Koduri. Anyway we will know within 3-4 months. My guess is this is the lineup.

Polaris 10

Ellesmere - 1024 sp, 128 bit GDDR5, 110 sq mm
Baffin - 2560, 256 bit GDDR5, 232 sq mm.

Polaris 11
Greenland - 4096 sp, 2048 bit HBM2, 350 sq mm. (Radeon/Firepro)

 

[Mahigan]

Since the Baffin circuit board doesn't mention GDDR5, it is likely using HBM on an interposer. 232mm2 would allow for far more than 2,560 SIMDs as well. Hawaii was 438mm2 and it had 2,816 SIMDs. If we assume a 50% die reduction with 14LPP, give or take, ~219mm2 is what Hawaii would take up. If we assume a 50% power reduction that's 137.5W (of course Hawaii made use of GDDR5 while Baffin appears to be HBM).

Baffin XT could have:
- 2,816 SIMDs (with higher shader efficiency)
- 64ROps
- 176 TMUs
- 4GB HBM
- New 4x Geometry Processors (Primitive Discard Accelerators)
- New Command Processor (Command buffer/Instruction prefetching for increased DX11/Single threaded performance)
- TDP lower than 130W
~219mm2 (and since it is likely 232mm2 then you could see room for the new Multimedia cores, larger caches etc).

That would make it quite the card for the price point in terms of performance per watt.

For Polaris 11, we could see a Greenland XT look like this:
- 4,096 SIMDs (with higher shader efficiency)
- 64ROps
- 256 TMUs
- 8GB HBM2
- New 4x Geometry Processors (Primitive Discard Accelerators)
- New Command Processor (Command buffer/Instruction prefetching for increased DX11/Single threaded performance)
- TDP lower than 160W
~350mm2 (Room for the new Multimedia cores, larger caches etc).

Architecture overhauls could account for +45% performance per clk over Hawaii and Fiji respectively.

We could also see changes in the ROp count but that would take a shader engine overhaul. We might only see that with Vega 11.

Neither of these two would compete with GP100. This leads me to believe that Vega 11 is intended as an answer to GP100.

 

[haedon]

How was the performance of polaris 11 compare to NVIDIA Titan X?

 

[Mahigan]

How was the performance of polaris 11 compare to NVIDIA Titan X?

It all depends on what a game stresses. By all accounts, FuryX is not ROps bound because if it were, it wouldn't be catching up or even surpassing a stock GTX 980 Ti when running 4k.

You could see this Polaris 11 move past a Titan X with ease. Heck that Baffin XT core could also give a Titan-X quite a run.

Sounds crazy but it really isn't.

 

[Magee_MC]

Something which is a distinct possibility is misinterpretation of Raja Koduri statements.

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

"Yes. We have two versions of these FinFET GPUs. Both are extremely power efficient. This is Polaris 10 and that’s Polaris 11. In terms of what we’ve done at the high level, it’s our most revolutionary jump in performance so far."

Raja states two versions of these FINFET GPUs. Why wouldn't he say two 14nm FINFET GPUs. So it could mean two types of designs - Polaris 10 for 14LPP entry/mid-range products with GDDR5 and Polaris 11 for 16FF+ flagship products with HBM2. TSMC has the better yields and 16FF+ has a 6 month TTM lead over 14LPP which again helps in volume ramp and yield learning. Moreover TSMC has experience with large GPUs for many process generations.

So it all comes down to what was meant by Raja Koduri. Anyway we will know within 3-4 months. My guess is this is the lineup.

Polaris 10

Ellesmere - 1024 sp, 128 bit GDDR5, 110 sq mm
Baffin - 2560, 256 bit GDDR5, 232 sq mm.

Polaris 11
Greenland - 4096 sp, 2048 bit HBM2, 350 sq mm. (Radeon/Firepro)

Raja said "Yes. We have two versions of these FinFET GPUs. Both are extremely power efficient. This is Polaris 10 and that’s Polaris 11. In terms of what we’ve done at the high level, it’s our most revolutionary jump in performance so far."

What if he was saying that there are two versions of EACH FinFET GPU like a FuryX and a Fury Nano. Each new GPU has one version tuned for power efficiency and another version tuned for processing power. A Polaris 10 chip would be like the Nano version and Polaris 11 would be the X version.

 

[Pinstripe]

Greenland (Flagship) has been renamed to Vega and won't probably release before 2017. The two FF chips Raja was refering to for this year are Polaris 10 & 11.

 

[Azix]

Why do you think TSMC has better yields?

He said two versions of these finfet GPUs. We know one of them is 14nm, so would that not suggest they have two versions of the same 14nm? Just saying.

doing both processes is unnecessary effort but we'll see.

Something which is a distinct possibility is misinterpretation of Raja Koduri statements.

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

"Yes. We have two versions of these FinFET GPUs. Both are extremely power efficient. This is Polaris 10 and that’s Polaris 11. In terms of what we’ve done at the high level, it’s our most revolutionary jump in performance so far."

Raja states two versions of these FINFET GPUs. Why wouldn't he say two 14nm FINFET GPUs. So it could mean two types of designs - Polaris 10 for 14LPP entry/mid-range products with GDDR5 and Polaris 11 for 16FF+ flagship products with HBM2. TSMC has the better yields and 16FF+ has a 6 month TTM lead over 14LPP which again helps in volume ramp and yield learning. Moreover TSMC has experience with large GPUs for many process generations.

So it all comes down to what was meant by Raja Koduri. Anyway we will know within 3-4 months. My guess is this is the lineup.

Polaris 10

Ellesmere - 1024 sp, 128 bit GDDR5, 110 sq mm
Baffin - 2560, 256 bit GDDR5, 232 sq mm.

Polaris 11
Greenland - 4096 sp, 2048 bit HBM2, 350 sq mm. (Radeon/Firepro)

 

[3DVagabond]

Why do you think TSMC has better yields?

He said two versions of these finfet GPUs. We know one of them is 14nm, so would that not suggest they have two versions of the same 14nm? Just saying.

doing both processes is unnecessary effort but we'll see.

Yeah. I think people are trying to add too much crypticism to what Raja said. He's an engineer not a PR person. He'll be pretty black and white.

 

[Kenmitch]

Lenovo Yoga 510-15ISK with r7 m460 2gb is launching in April. Not much info yet as to if it's a rebrand or possibly baby Polaris.

http://videocardz.com/58323/amd-radeon-m400-series-to-be-released-in-april

Using the Polaris demo results baby Polaris would most likely smack down the 960m I'd imagine. At least looking at some quick results of the GTX 950 vs 960m makes me believe so.

 

[raghu78]

Why do you think TSMC has better yields?

The fact that Apple A9X which is a 147 sq mm SoC is manufactured at TSMC 16FF+ instead of 14 LPP. The other reason is TSMC seems to have won 100% of next gen A10/A10X production. TSMC also confidently stated they will have almost 70% FINFET foundry market share in 2016. This points to strong customer adoption/demand for what is the best foundry FINFET process in terms of yields and electrical characteristics.

 

[Glo.]

Erm, so far from what has been floating around here and there, Polaris 11 is supposed to be around 220mm2 GPU, and Polaris 10 should be around 115 mm2.

 

[Glo.]

The fact that Apple A9X which is a 147 sq mm SoC is manufactured at TSMC 16FF+ instead of 14 LPP. The other reason is TSMC seems to have won 100% of next gen A10/A10X production. TSMC also confidently stated they will have almost 70% FINFET foundry in 2016. This points to strong customer adoption/demand for what is the best foundry FINFET process in terms of yields and electrical characteristics.

A9X was produced on 16nm FF+ is because at time when it was needed to be produced there was no process from Samsung to make such powerful chip. There was only 14 nm LPE. And Apple produced A9 on both manufacturers because of volume that had to buy from TSMC, and A9X was not enough alone.

 

[Glo.]

Ellesmere - 1024 sp, 128 bit GDDR5, 110 sq mm
Baffin - 2560, 256 bit GDDR5, 232 sq mm.

Polaris 11
Greenland - 4096 sp, 2048 bit HBM2, 350 sq mm. (Radeon/Firepro)

Baffin: 4096 GCN cores, HBM2. - Polaris 11.
Ellsmere - 2048 GCN cores, 128 Bit memory bus, GDDR5 - Polaris 10
Greenland - 6144 GCN cores, HBM2, - Vega 10.

If we are playing with predictions.

 

[Krteq]

Lenovo Yoga 510-15ISK with r7 m460 2gb is launching in April. Not much info yet as to if it's a rebrand or possibly baby Polaris.

http://videocardz.com/58323/amd-radeon-m400-series-to-be-released-in-april

Using the Polaris demo results baby Polaris would most likely smack down the 960m I'd imagine. At least looking at some quick results of the GTX 950 vs 960m makes me believe so.

Those are rebrands.

By the way, Banks and Weston are 28nm GPUs used in embeded market and are based on GCN Gen 2 (Banks) and GCN Gen 3 (Weston).