Dual Polaris speculation/discussion

i might be reading it wrong, does not seem to be doing better than 2 RX 480s CF. hardly worth making if that is true.
 
SighTurtle said:
i might be reading it wrong, does not seem to be doing better than 2 RX 480s CF. hardly worth making if that is true.
Click to expand...
I wouldn't expect it to be better than two rx480s, dual gpu board are usually clocked lower and more aggressively throttled.

More importantly I'm wondering how this reflects on Vega's release timeframe, makes no sense to make this if Vega is around the corner
 
Ieldra said:
I wouldn't expect it to be better than two rx480s, dual gpu board are usually clocked lower and more aggressively throttled.

More importantly I'm wondering how this reflects on Vega's release timeframe, makes no sense to make this if Vega is around the corner
Click to expand...
It's entirely possible there is a refresh of Polaris. Not saying that's what's happening here, but fabbing at TSMC, adding some faster memory (easy 10-15%), or some other fix could be significant.

The boards likely make sense in the professional/server market to increase density. Why they haven't been released until now, if that's the case, is anyone's guess. It still seems strange they didn't have a Polaris based Fury design. Fury is already ~$300 and I'd expect Polaris+4GB HBM1 to be comparable in performance, depending on title, and likely cheaper. Leaving open that dual GPU on interposer design I speculated a while back. AMD are still making use of the Fiji dice in the profession market with SSG. No idea what the sales are there, but that's still an intriguing product.

As for the Vega timeline I'd agree it's interesting, but the part may just be a byproduct of the professional/server sector. Even if Vega is superior, a partner may still be deploying servers based around Polaris. Especially if Vega is primarily HBM2/interposer based. I wouldn't be surprised if we only see interposer based Vega variants. APUs for the low to mid segments, high end HBM2 discrete boards, and maybe Polaris lingering in the low to mid discrete market. That being the case a dual Polaris likely still has some life left.
 
Link doesn't seem to be working for me. Do you have a screenshot?
 
Anarchist4000 said:
It's entirely possible there is a refresh of Polaris. Not saying that's what's happening here, but fabbing at TSMC, adding some faster memory (easy 10-15%), or some other fix could be significant.

The boards likely make sense in the professional/server market to increase density. Why they haven't been released until now, if that's the case, is anyone's guess. It still seems strange they didn't have a Polaris based Fury design. Fury is already ~$300 and I'd expect Polaris+4GB HBM1 to be comparable in performance, depending on title, and likely cheaper. Leaving open that dual GPU on interposer design I speculated a while back. AMD are still making use of the Fiji dice in the profession market with SSG. No idea what the sales are there, but that's still an intriguing product.

As for the Vega timeline I'd agree it's interesting, but the part may just be a byproduct of the professional/server sector. Even if Vega is superior, a partner may still be deploying servers based around Polaris. Especially if Vega is primarily HBM2/interposer based. I wouldn't be surprised if we only see interposer based Vega variants. APUs for the low to mid segments, high end HBM2 discrete boards, and maybe Polaris lingering in the low to mid discrete market. That being the case a dual Polaris likely still has some life left.
Click to expand...


Another great suggestion, a 390X type revamp, but this soon? I am unsure about that. Also the double price would be unlikely for a refresh. People would bitch and moan. Dual gpu or SSG are the two most likely it seems, however I have another theory now after sleeping on this all.

Vega 11 is after Vega 10 road map, but we all know road maps are not always black and white. Vega 11 would also be mostly a Polaris replacement (again recently released), Vega 10 would almost undoubtedly be priced higher than 2x that of a Baffin XT 4gb shipment....

Are AMD going to play the weirdest game of chess ever, releasing two mid range architectures sequentially in form of Polaris then Vega 11? With Vega 11 being the natural RX570-RX580 with appropriate pricing and performance to match a 1070 price range? Given the lack of R&D and high end success lately, this may strangely make some sense after much consideration.... Yesterday I didn't even consider it.

Zauba shipment would've listed dual baffin xt or similar, it's a known chip.. they almost always list known chips in dual config in past. This has no mention... perhaps it is Vega 11?

Or is it a 475/485 type scenario....?

Going to be an interesting month whatever they drop, it will be telling of how things are going.
 
N4CR said:
Or is it a 475/485 type scenario....?
Click to expand...
If a Polaris refresh it could be a bit different. As a 490/X didn't previously exist a significant refresh could fall there. With the WSA having GF make 470/480 and TSMC the 490 might be reasonable. That would rely on the current Polaris being rather hobbled by GF. If they managed to make it 40-50% faster it could be a 490. Not sure how well that correlates to the AOTS scores though. Faster ram for 1.2x and clocked towards 1.8GHz it could approach that. Still seems more likely a dual GPU or Vega.
 
Anarchist4000 said:
If a Polaris refresh it could be a bit different. As a 490/X didn't previously exist a significant refresh could fall there. With the WSA having GF make 470/480 and TSMC the 490 might be reasonable. That would rely on the current Polaris being rather hobbled by GF. If they managed to make it 40-50% faster it could be a 490. Not sure how well that correlates to the AOTS scores though. Faster ram for 1.2x and clocked towards 1.8GHz it could approach that. Still seems more likely a dual GPU or Vega.
Click to expand...
I thought it was already surmised that TSMC would offer no benefit.
 
JustReason said:
I thought it was already surmised that TSMC would offer no benefit.
Click to expand...

There would be a small benefit but also a small downside. Its certainly not going to clock much different tho. And it would come with a big bill to create 16FF+ masks etc. GP107 does quite well on 14LPP.
 
There is also another tradeoff as we see with Nvidia moving to 14nm and that is the performance envelope subtle change, notice that on TSMC the optimal voltage for Nvidia is around 1.06-1.09V, while on Samsung it is around 1.075-1.11V, it does not look much but has implications and pros-cons.
Cheers
 
JustReason said:
I thought it was already surmised that TSMC would offer no benefit.
Click to expand...
I doubt it would be huge, but I'd expect some gains. More parts like that 1475MHz@100W card that was reviewed a while back. AMD should know if it's worth it because GCN gets fabbed at TSMC for console parts. So they should have had an idea what differences could exist at the very least. Whether or not that's worth it who knows, but they may have wanted better parts for the professional market or a 490 style part. Again I doubt this is the case, but certainly seems feasible.

CSI_PC said:
There is also another tradeoff as we see with Nvidia moving to 14nm and that is the performance envelope subtle change, notice that on TSMC the optimal voltage for Nvidia is around 1.06-1.09V, while on Samsung it is around 1.075-1.11V, it does not look much but has implications and pros-cons.
Cheers
Click to expand...
Worth mentioning power is a square of voltage. Small change, but not entirely insignificant. The real gains would be if the tolerances improved on the process. In the case of Nvidia the change was subtle, but it was also a relatively small part. If there was a minimal difference, I'd expect Nvidia to be fabbing more chips there to cut costs already.
 
Anarchist4000 said:
I doubt it would be huge, but I'd expect some gains. More parts like that 1475MHz@100W card that was reviewed a while back. AMD should know if it's worth it because GCN gets fabbed at TSMC for console parts. So they should have had an idea what differences could exist at the very least. Whether or not that's worth it who knows, but they may have wanted better parts for the professional market or a 490 style part. Again I doubt this is the case, but certainly seems feasible.


Worth mentioning power is a square of voltage. Small change, but not entirely insignificant. The real gains would be if the tolerances improved on the process. In the case of Nvidia the change was subtle, but it was also a relatively small part. If there was a minimal difference, I'd expect Nvidia to be fabbing more chips there to cut costs already.
Click to expand...
Yep but I feel the voltage tolerances are stronger with the Samsung technology.
Worth saying there is also leakage/waste energy/variance to consider as well along with influence on silicon localised die thermal hotspots (quite a lot of EE papers cover this), and yeah totally agree part of this is also tolerances.

Cheers
 
You must log in or register to reply here.
Back
Top