[Micron]GDDR5X has Arrived

[AtenRa]

Polaris 10 Performance would be at max equivalent to a ~220-240mm2 GCN part on 28nm. Well below 380X levels.

It is LOW-end, targeting the 950, not even the 960.

If Polaris is close to 120mm2, i believe it could reach 380X level of performance, especially in DX-12 games.

Let me explain,

Samsung 14nm LPE is ~x2.2 denser than 28nm TSMC.

Tonga die with 384bit memory controller is 366mm2
Take out 2x 128bit controllers and you almost have a 300mm2 die.

Shrink that to 14nm and you get almost 120-130mm2 die.

Now add all the new architectural advancements on Pollaris and you could reach R9 380X performance.

 

[AtenRa]

This year i.e. first gen Arctic Islands?

Personally I believe that HBM2 could start to be used on lower level GPUs from H2 2017, even on high-end APUs. So i dont see where AMD would use GDDR5X.

 

[dacostafilipe]

With the added PCB complexity and the new memory controller needed, I don't think GDDRX5 it will be used this year.

 

[Glo.]

I think AMD may have 3 GPUs now that we have seen the zauba listings. I think Polaris 10 refers to GDDR5 based Polaris and Polaris 11 refers to HBM2 based Polaris. We might see 2 Polaris 10 GPUs and 1 Polaris 11 GPU.

http://videocardz.com/58101/nvidia-gp104-and-first-polaris-gpu-supposedly-spotted-on-zauba

http://wccftech.com/amd-polaris-gpus-spotted/

Polaris 10 -

Entry level
C913 (102-C91301-00)
C924 (102-C92401-00)

Mid range
C980 (102-C98001-00)
C981 (102-C98101-00)

Polaris 11-

Flagship
C993 (102-C99398-00)

AMD launched GCN in 2012 with Tahiti, Pitcairn, Cape Verde. I don't think they will change that strategy of 3 GPUs to address the entire price range.

Correction.
C913/C924 - Cut down Polaris 10/full Polaris 10
C980/C981 - Cut down Polaris 11/Full Polaris 11
C993 - Vega 10

Vega 10 is with HBM2. The rest - only AMD knows.

 

[Silverforce11]

That would fit the 2 Polaris chips this year... they are being very coy.

Wonder what Vega 10 is, a big die built on TSMC 16ff?

 

[Glo.]

Cut down Polaris 10: 1536 GCN cores, 128 Bit memory, less than 50W TDP.
Full Polaris 10: 2048 GCN cores, 256 Bit memory 75W TDP. Around 120mm2 die size.
Cut down Polaris 11: 3072 GCN cores, 384 Bit bus, 125W. Around 210 mm2.
Full Polaris 11: 4096 GCN cores, HBM1, 150 W TDP.
Vega 10: 6144 GCN cores, HBM2, 200W TDP.

 

[dacostafilipe]

Cut down Polaris 11: 3072 GCN cores, 384 Bit bus, 125W. Around 210 mm2.
Full Polaris 11: 4096 GCN cores, HBM1, 150 W TDP.

Does it really make sense to have GPU that supports both HBM and GDDR?

 

[ShintaiDK]

HBM isn't the all great savior. GDDR5X is quite competitive for now.

 

[Glo.]

Does it really make sense to have GPU that supports both HBM and GDDR?

Polaris is designed to run both on GDDR5 and HBM.

 

[caswow]

HBM isn't the all great savior. GDDR5X is quite competitive for now.

hbm is in every metric better than hbm except cost. only nvidia is bashing agains hbm wonder why. parasitical behaviour is the worst. they dont even have a product on the market and they managed to talk hbm down. why dont they just stick to microns gddr5x :thumbsup:

it seems like nvidia is starting the damage control because they are not able to make a good card with hbm due to low know how in this field.

 

[raghu78]

Having 3 entirely new GPUs would entirely contradict what Raja and other AMD execs have said about GPU releases in 2016. There would be no point to lie to the public about it either, given how low their stock is and Nvidia operating on their own R&D / timetable regardless of what AMD does.

Let me quote Raja

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"

There is no confirmation that its 2 GPUs only. The 2 versions of FINFET GPUs could mean

1. GDDR5 based GPUs manufactured at GF 14LPP.
2. HBM2 based GPUs at TSMC 16FF+.

AMD have experience with manufacturing big GPUs at TSMC. With Fiji they also have experience with HBM and 2.5D stacking at TSMC . So staying with TSMC for the big die HBM2 GPU makes sense. Yields are also better with TSMC.

110 sq mm - 1024 sp, 128 bit GDDR5, 50-60w GPU.
200-220 sq mm - 2048 sp, 256 bit GDDR5, 110w GPU.
300+ sq mm - 4096 sp, 2048 bit HBM2, 190-200w GPU.

 

[ShintaiDK]

hbm is in every metric better than hbm except cost. only nvidia is bashing agains hbm wonder why. parasitical behaviour is the worst. they dont even have a product on the market and they managed to talk hbm down. why dont they just stick to microns gddr5x :thumbsup:

it seems like nvidia is starting the damage control because they are not able to make a good card with hbm due to low know how in this field.

You do know NVidia uses HBM2 as well. And AMD is very likely to use GDDR5X?

Seems you forgot that in your NVidia rant.

 

[ShintaiDK]

Polaris is designed to run both on GDDR5 and HBM.

I dont think you see the same die with 2 types of memory controllers so different.

But 1 Polaris/Vega die with a GDDR5(X) IMC and one with a HBM. Sure.

 

[caswow]

You do know NVidia uses HBM2 as well. And AMD is very likely to use GDDR5X?

Seems you forgot that in your NVidia rant.

what was the point of the presentation? to tell us how bad hbm is? because thats what comes off of this...

might have been better if they had a better solution...like never.

 

[ShintaiDK]

what was the point of the presentation? to tell us how bad hbm is? because thats what comes off of this...

might have been better if they had a better solution...like never.

You put way more into it than there is due to your attachment.

New technology will be needed.

 

[dacostafilipe]

HBM isn't the all great savior. GDDR5X is quite competitive for now.

Going by your chart it is not. GDDR5X should deliver 2x the GDDR5 bandwidth using the same power. That would still not beat HBM2.

If your chart would show performance/$, then yes, you would be right.

Polaris is designed to run both on GDDR5 and HBM.

I'm not talking about the architecture but about the GPU Die (Polaris 10)

 

[Despoiler]

HBM isn't the all great savior. GDDR5X is quite competitive for now.

Did you even read the presser you posted in your own OP? Micron said GDDR5X is @ 1.35v. The lowest power GDDR5 can go is 1.35v. This is not coincidence. As HBM is 1.2v, that means you just posted an Nvidia graphic of how GDDR5X is not competitive. Not only that, literally no one has said anything about HBM being the solution until the end of time. It was designed to be better than GDDR5, which it is in every respect. This is just like GDDR5 was designed to be better than GDDR4, so on and so forth back through history. I needed a good laugh to start my day.

 

[ShintaiDK]

Did you even read the presser you posted in your own OP? Micron said GDDR5X is @ 1.35v. The lowest power GDDR5 can go is 1.35v. This is not coincidence. As HBM is 1.2v. That means you just posted an Nvidia graphic of how GDDR5X is not competitive. Not only that, literally no one has said anything about HBM being the solution until the end of time. It was designed to be better than GDDR5, which it is in every respect. This is just like GDDR5 was designed to be better than GDDR4, so on and so forth back through history. I needed a good laugh to start my day.

Is HBM better than GDDR5 in cost? I dont think so and I doubt it ever will be.

I think you miss the point. HBM seems to be nothing than a premium product with a relatively short lifespan. Solving nothing but the issues of high end cards. In other words, HBM wont replace GDDR5(X) but rather be another option on the market.

 

[AtenRa]

Where is the price/bandwidth/power comparison of GDDR5X vs HBM2 ??

 

[ShintaiDK]

Where is the price/bandwidth/power comparison of GDDR5X vs HBM2 ??

Remember density too

 

[NTMBK]

Is HBM better than GDDR5 in cost? I dont think so and I doubt it ever will be.

I think you miss the point. HBM seems to be nothing than a premium product with a relatively short lifespan. Solving nothing but the issues of high end cards. In other words, HBM wont replace GDDR5(X) but rather be another option on the market.

GDDR5X is only going to push costs up compared to GDDR5; lower yields, higher pin count and denser traces on the PCB.

 

[ShintaiDK]

GDDR5X is only going to push costs up compared to GDDR5; lower yields, higher pin count and denser traces on the PCB.

You can pretty much say the same about HBM2 over HBM1.

With yields already giving 13Gbps chips I think we will be fine. HBM2 for high end, GDDR5X for performance/mainstream and GDDR5 for low end is how I see it. 4GB low, 8GB middle, 12-16GB high.

 

[raghu78]

Is HBM better than GDDR5 in cost? I dont think so and I doubt it ever will be.

I think you miss the point. HBM seems to be nothing than a premium product with a relatively short lifespan. Solving nothing but the issues of high end cards. In other words, HBM wont replace GDDR5(X) but rather be another option on the market.

GDDR5X will make sense where cost matters. We should see GDDR5X coexist with HBM2 till the end of the decade. Once we go to 5nm and below GDDR5X's bandwidth per sq mm will not be enough to power even the 100-120 sq mm GPUs. At that time it would be HBM2 or HBM3 (which JEDEC is working on).

 

[NTMBK]

You can pretty much say the same about HBM2 over HBM1.

With yields already giving 13Gbps chips I think we will be fine. HBM2 for high end, GDDR5X for performance/mainstream and GDDR5 for low end is how I see it. 4GB low, 8GB middle, 12-16GB high.

A handful of test pieces have hit 13Gbps. That doesn't mean that it is yielding well.

I doubt that GDDR5 is going to start moving into the low end; it's still just too expensive. I'd expect DDR4 to start displacing DDR3 in the low end.

 

[xthetenth]

@R0H1T

Polaris 10 is a very small chip. There's no need for the big increase in bandwidth of GDDR5X (which uses more power than normal GDDR5) for such a product. Low-end, low-cost, GDDR5 is effective enough.

Polaris 11 has to be either HBM2 or GDDR5X (not and/or), since going for GDDR5X means it need a MC, as the MC on HBM2 is on the memory stack. That would double research & development of two separate chip. I don't think AMD is in a financial position to do that and they don't need to. They already have experience with HBM.

So AMD's stack doesn't involve GDDR5X, because they are missing the mid-range chip that needs that bandwidth gains.

What if it's the other way round and they're missing the mid-range chip because it's waiting on GDDR5X? That would create a consistent situation with a high and low launch right now, and a reason for something other than a full stack.