AMD: Speculation, Rumors, and Discussion (Archive)

That picture looks massively distorted, judging from the capacitors to the left.

 

Yes, it's far from perfect. It's based on the link here: http://unlocked.newegg.com/sites/default/files/amd-rx-460-470-hands-on-3.jpg

Anyway, If we know the size of the GDDR-chip, which is distorted the same way, we should get a decent size estimation of the P10-chip.

 

14x12 mm.

BTW, I measured ~266 mm² for the GPU.

 

Wasn't it 23x mm² according to rumor mill until recently?

 

Too much margin of error on this picture, as the die is rotated of 45°C vs the GDDR ones, you need correct this too. and you will need to know the focal length . maybe by using a 3D software VFX ( Blender + CAM correction for camera mapping ) who can correctly calculate the focal, correct the aspect ratio the image and apply correction. .

 

Yes. 232 mm2 according the Linkedin AMD-engineer.
If the GGDR5-chip is 168mm2 and even though the image of the card is distorted and could be showing somewhat wrong proportions... I think it safe to say that P10 is at least 232 mm2

Now. If we take Tonga which is 366 mm2, and shrink it to the 14 nm lpp process, it should be roughly half the size yes? That makes it 186 mm2. Tonga is a 2048 shader chip.
RX 480 only have 256 more stream processors. Isn't that a bit low for about 25% more die area?

Then, If P10 in fact is bigger than 232 mm2...

 

I usually use Vanishing Point tool in Photoshop to align the measurements to the image's perspective.

 

http://www.anandtech.com/show/9886/amd-reveals-polaris-gpu-architecture/2
That stuff has a die size cost. Also, certain elements of the chip (e.g. GDDR5 PHYs) won't scale down linearly.

 

Actually, on a more refined measurement, I'm getting closer to the 232 mm² figure. The problem with the picture of the board is that the edges of the GPU die are very obscured and hard to accurately distinguish them from the underfill material.

 

What if AMD is saving fully enabled parts for next year?

i.e.
480X - 40CU
480 - 36
470X - 32 ???
470 - 28 ???
460X - 20 ???
460 - 16

They'd maybe get more yields out of introducing non-X parts now, and then when the node matures, they can bring out fully operational chips.

One might say it was the reverse of 28nm strategy (aside from tonga situation).

 

Yes. You are probably right on this one.

 

Not selling things you have is a pretty poor way to make money.

 

That depends on exactly how many of fully working chips you have. To explain it, one would need the assumption of significant yield problems with GF's 14LPP process. But if there aren't enough working 40CU chips, you can't sell them properly, because in the price range below 300$ demand is usually high. Or you price the thing ridiculously high to mask the low availability. But this tends to work only for halo parts. For mainstream/performance GPUs it could have negative impacts on the image of the rest of the line. Furthermore, it could be the case that a 40CU version running at 1.3+GHz (if 1.267GHz for the RX480 are true) is already severely bandwidth limited in quite some scenarios, so it isn't really worth it without GDDR5X, which is currently in short supply (again assuming Polaris10 MC supports GDDR5X, which may be not the case).

 

You're lowering the barrier to usable chips for a particular SKU, so you end up with increased product volumes. Next year, people will expect faster products instead of rebadged, side-ways products (or even lower spec). Process will inevitably mature anyway, so they'd be able to source more fully enabled chips.

Perception.

 

Sure, there's a theoretical edge case where it would sort of make sense but by that point were getting into conspiracy land.

AMD having parametric yields so bad to not even announce the fastest SKU on a <250mm part would take a Fermi-on-40nm level disaster and the 1070 seems to be doing fine on bandwidth.

You're also crippling your competitiveness and racking up storage fees for something that could be done with a simple clock increase down the road.

 

Storage fees? They're just setting the standard usable production at fewer functional CUs. If more work, then so be it - laser cut. It's not like it hasn't happened before.

AMD is intent on doing a price war of sorts, remember. More chips that can be used for a product line now. When fully enabled parts crosses some yield threshold (or they're looking at the next year's product line-up), then they'll have something to show to OEMs or to folks holding off.

There's already rumour of "10nmFF" being not much of an upgrade as well. For all we know, it's gonna be another 4-5 years to hit 7nm, and we don't know what sorts of things AMD is planning to do to GCN in the interim. Stretching out the Polaris 10/11 masks for products over the years might be useful for their er... bottom line/budgets.

 

According to most leaks, P10 seems to score around 390/x in benchmarks. RX 480 is probably? also the full P10-chip which makes it a great upper mid range card at a very good price.

But what about the HighEnd?
Is RTG:s plan to leave the performance/HighEnd completely to Nvidia for a whole 6 months?
Or maybe the GP104:s performance took them by a complete surprise?

Performance/watt seems to be the main focus, not necessary performance.

 

This only makes sense if they can't get more than a couple of fully working dies out of a wafer and there's no sign of that changing in the next month or two, but you have bigger problems to deal with than upcoming refreshes, if so. Else there's no reason to not release a premium fully enabled SKU, I'm not convinced that not increasing prices makes for a price war just yet.

 

Forgive my lack of faith in Global Foundries.

Anyways, just a random idea.

 

Probably.

I'd expect they plan to recapture market share by catering to OEMs with products with good performance/$ and performance/power ratios.

That, and start binning low power parts for future mobile SKUs, - that's where the bulk of the market is anyway.

Cheers