Full AMD Polaris 10 GPU has 2304 Stream Processors

[Silverforce11]

How many times did AMD claim that Tonga only had a 256bit memory bus? Turned out not to be true.

People here have very short term memories, or selective memories to be precise.

Wasn't long ago where PR gave up false hardware spec.

 

[Sweepr]

How many times did AMD claim that Tonga only had a 256bit memory bus? Turned out not to be true.

Shortly after Radeon R9 380X launch they confirmed 384-bit when directly asked, they never lied about it.

AMD confirms 384-bit bus available on Tonga, just not enabled on any product, including 380X. Didn't find a perfect perf/$ slot.

https://twitter.com/ryanshrout/status/672586146300301313?ref_src=twsrc^tfw

Makes zero sense for a Senior Product Manager to explicitly lie about Polaris 10.

 

[tviceman]

How many times did AMD claim that Tonga only had a 256bit memory bus? Turned out not to be true.

Such a bad argument. How many 384 bit Tonga GPU's came out?

RX 480 is a full die P10. It's only 232mm2, yields are fine and production is high. If P10 had a higher performing part with more CU's AMD would have announced and/or released a higher performing part with a higher price.

Let it go.

 

[tviceman]

Deleted

 

[Azix]

polaris 9... 40 CUs. lmao

 

[Silverforce11]

Here's my logic why Polaris 10 has more than 36 CU, see if it makes sense or whether there's anything wrong with this logic.

And no, I don't believe PR statements is ever 100% truth. Recent history should have taught you all that already. So if you don't want to follow logic and go with PR statements, don't bother responding.

1. PS4 Neo, 36 Polaris 10 CU confirmed. Why does this matter? Because you do not ever make a mass volume SOC that ship in the many millions of units by offering the full chip spec. The yield and economics of it as well as production capability do not add up.

The PS4 and Xbox One both had more CU on the chip than what the final hardware shipped with. PS4 has 16 CU enabled, out of a 20 CU design. Similar for the Xbone but less CU overall.

If Polaris 10 was a 36/36 CU chip, it would simply not be able to meet volume demands of the consoles as only the best chips would be used, what happens to the rest that don't make the cut? Discard all of them? Complete waste of $ and chips/wafer/time.

You could argue, the PS4 Neo & Scorpio are custom chips and they can have different configuration to the base architecture chip for PC. Sure it's possible. But this adds cost to the SOC design. It's far simpler and cost effective to take the existing HSA compatible Polaris design and bolt it onto a SOC with a CPU and interlinks.

2. Apple. They want the best chips because of perf/w reasons. To get better perf/w it is easier to achieve using a wider full chip running at lower clocks than a cut chip with higher clocks. This just happened recently with Tonga. Apple full chip, 125W, PC chip, 180W. Similar performance due to full shaders @ lower clocks.

This lines up with #3.

3. Polaris 11, only in the RX 460. No alternate SKU? Where's the harvested silicon going, to the dumpster? To Apple? How likely is it that Polaris 11 is 1024 SP full form? We've already seen Sisoft leaks of Polaris 11 at 1280 SP. These Sisoft leaks were spot on for Polaris 10 SKUs, 2304SP @ 1266mhz and a cut down one 2048 SP.

It stands to reason, Polaris 11 is actually 1280 SP. The RX 460 is using the harvested chip, 1024 SP. The full chips? They are going to Apple.

If AMD can segment Polaris 11 and give the best chips to Apple, surely they can do the same with Polaris 10. What is more likely, Apple demanding the best chips or they are happy to get harvested chips?

 

[3DVagabond]

Such a bad argument. How many 384 bit Tonga GPU's came out?

RX 480 is a full die P10. It's only 232mm2, yields are fine and production is high. If P10 had a higher performing part with more CU's AMD would have announced and/or released a higher performing part with a higher price.

Let it go.

Let what go? I've never claimed it had more CU's. I'm only stating that just because someone at AMD said something it doesn't mean anything certain. We are constantly getting wrong info from these companies.

You need to let your attitude go. Rather than think you should be telling anyone what to do around here.

 

[tviceman]

It's 232mm2. Its a high volume "85% of the market" chip. No chip under 300mm2 has ever led off with a cut down sku. Apple's volume is miniscule to P10's overall production.

Its full die. AMD confirmed it. Other hardware sites have confirmed it. Let's put this horrible fantasy to rest.

 

[Silverforce11]

Btw, for those so keen on PR statements above logic, a reminder.

Logic told me Maxwell was incapable of Async Compute well before anyone found out the truth. I even called out Ryan Smith here for writing a WRONG article on Asynchronous Shaders, with NV PR propaganda making absurdly wrong claims in their published article. Time proved me correct.

Logic also told me that Maxwell was incapable of compute preemption, again, well before NV came out and basically admitted it with Pascal's new and improved preemption.

This is despite NV PR making repeated claims to the contrary. You wanna believe PR? Go ahead.

 

[Silverforce11]

It's 232mm2. Its a high volume "85% of the market" chip. No chip under 300mm2 has ever led off with a cut down sku. Apple's volume is miniscule to P10's overall production.

Did you even look up Apple volume?

Last I checked, 5M Macs per quarter.

Not all will have dGPU, but even if 20% of that, that's a million per quarter.

What's AMD producing on 14nm FF right now?

Let's see:

PS4 Neo & Xbox Scorpio APUs. Likely Nintendo NX (not confirmed yet) too.
Polaris 11, 10.
Zen.
Other custom SOC for Chinese markets.

 

[Arachnotronic]

Did you even look up Apple volume?

Last I checked, 5M Macs per quarter.

Not all will have dGPU, but even if 20% of that, that's a million per quarter.

20% is too high, IMHO. Highest volume sellers by a mile are MacBook Air/MacBook/13-inch MacBook Pro, all use iGPU only. Even 15-inch rMBP comes in both dGPU and iGPU option, and I doubt that most people buy up.

In desktops, all 21.5-inch now using iGPU, only the expensive 27-inch Retina iMacs and Mac Pro use dGPU. I would bet on it being more like 10%.

What's AMD producing on 14nm FF right now?

Let's see:

PS4 Neo & Xbox Scorpio APUs. Likely Nintendo NX (not confirmed yet) too.
Polaris 11, 10.
Zen.
Other custom SOC for Chinese markets.

Doubt that Neo and Scorpio APUs are in production now, Zen probably not either. Custom SoC for Chinese markets probably a long ways off too.

 

[Silverforce11]

Doubt that Neo and Scorpio APUs are in production now, Zen probably not either. Custom SoC for Chinese markets probably a long ways off too.

You can't paper launch consoles, with demands of multi-million units per quarter. Volume would have to be built over a long period of time prior to retail.

There's a range of custom SOC for Chinese market, one of which is a console that's VR capable.

 

[Arachnotronic]

You can't paper launch consoles, with demands of multi-million units per quarter. Volume would have to be built over a long period of time prior to retail.

There's a range of custom SOC for Chinese market, one of which is a console that's VR capable.

Don't know when PS4 Neo is coming, but if XBox Scorpio is launching in December 2017, then there is no way they are stockpiling chips right now. Apple doesn't start running wafers at its fab partners a year in advance of its iPhone launches and those are higher volume than any console ever.

 

[96Firebird]

Honestly, when I first saw this thread show up, I thought it'd die slowly because it was already known that RX 480 was the full die.

Turns out, some still believe it isn't...D:

 

[godihatework]

Btw, for those so keen on PR statements above logic, a reminder.

Logic told me Maxwell was incapable of Async Compute well before anyone found out the truth. I even called out Ryan Smith here for writing a WRONG article on Asynchronous Shaders, with NV PR propaganda making absurdly wrong claims in their published article. Time proved me correct.

Logic also told me that Maxwell was incapable of compute preemption, again, well before NV came out and basically admitted it with Pascal's new and improved preemption.

This is despite NV PR making repeated claims to the contrary. You wanna believe PR? Go ahead.

http://www.pcgameshardware.de/screenshots/1020x/2016/06/Polaris-P10-Die-Shot-pcgh.jpg

did PR shop the die shot as well?

 

[Silverforce11]

http://www.pcgameshardware.de/screenshots/1020x/2016/06/Polaris-P10-Die-Shot-pcgh.jpg

did PR shop the die shot as well?

If that die shot is true...

Polaris is a 2880 SP chip with 576 SP disabled to arrive at the 2304 SP part.

I was thinking that they are FP64 units, but then we all know AMD's GCN does not have separate FP64 Stream Processors. The SPs are linked via a hardware scheduler to operate FP64 instructions at a lower ratio, 1:2 in Hawaii etc.

Thank you for proving me correct, P10 in it's 2304SP configuration is a cut-down chip. Apple took all the full chips, again.

 

[Silverforce11]

Btw, such a large cut 2880 -> 2304 SP would imply GloFo's 14nm FF is utter rubbish.

 

[Arachnotronic]

If that die shot is true...

Polaris is a 2880 SP chip with 576 SP disabled to arrive at the 2304 SP part.

I was thinking that they are FP64 units, but then we all know AMD's GCN does not have separate FP64 Stream Processors. The SPs are linked via a hardware scheduler to operate FP64 instructions at a lower ratio, 1:2 in Hawaii etc.

Thank you for proving me correct, P10 in it's 2304SP configuration is a cut-down chip. Apple took all the full chips, again.

Silverforce11,

AMD has said that RX 480 is the full Polaris 10, so I don't know why you keep saying that RX 480 is cut down.

 

[Silverforce11]

Silverforce11,

AMD has said that RX 480 is the full Polaris 10, so I don't know why you keep saying that RX 480 is cut down.

Listen buddy, I don't give a rats what AMD or NV PR says about hardware specs, they have lied in the past. Your memory is very selective.

IF the die shot is true, Polaris 10 is 2880 SP.

You can argue the die shot is fake, whatever.

 

[Arachnotronic]

Listen buddy, I don't give a rats what AMD or NV PR says about hardware specs, they have lied in the past. Your memory is very selective.

IF the die shot is true, Polaris 10 is 2880 SP.

You can argue the die shot is fake, whatever.

The additional 576 SPs you refer to are very clearly in a different color than the 2304 others. Maybe they're disabled but they could also be something different like double precision units or something.

 

[Abwx]

Silverforce11,

AMD has said that RX 480 is the full Polaris 10, so I don't know why you keep saying that RX 480 is cut down.

Where he has a point is that if it s a 36 CU die then GF yields must be extraordinary good, otherwise they would had launched the 470 as well to get rid as soon as possible of the numerous non fully functional chips, that s exactly what Nvidia did by releasing two cards.

 

[Silverforce11]

For those who want to believe PR, let me test your logic..

NVIDIA says the 1070 has 64 ROPs. Is this the complete truth?

Look at the diagram and look at the actual test result of Rasterizer performance.

^ If you know anything about NV's architecture layout, you would have quickly realized Rasterizers are within a GPC cluster, if it's cut, bye bye ROPs.

Look at it's fillrate performance:

Ohh look at that! Nowhere near the 1080 with full 64 ROPs. It looks to be missing quite a few, like it's only got 48 ROPS usable.

What a coincidence, each GPC has 16 ROPs, four for the full GP104 equates to 64, 3 for the 1070 equates to 48 ROPs.

Do you trust AMD or NV PR?

Need I remind you, 970 fiasco?

Or recently with AMD, 150W RX 480! lol

 

[Saylick]

Those yellow colored units look to be the TMUs. Each block corresponds to 4 TMUs for a total of 144 TMUs, which lines up with what we have for the RX480.

Each block of 64 SPs + 4 TMUs is one CU. 9 x 4 groups of CUs gives you 2304 SPs.

The two 'L' shaped blocks on the top-left and bottom-left are the memory controllers. Looks like 8 groups, so 32-bits per group gives you 256 bit memory bus.

Of course, this is assuming the diagram is even accurate.

 

[Silverforce11]

@Saylick
Thank you, if that is 4 unit blocks, then they are TMUs. In the picture I had from PCGH, it looks like 16 unit blocks.

 

[IEC]

@Saylick
Thank you, if that is 4 unit blocks, then they are TMUs. In the picture I had from PCGH, it looks like 16 unit blocks.

I don't know why they used JPEG compression. That introduced artifacts that made the TMUs look like the blocks.

In any event, the PNG image makes it clear those are TMUs. So we are back to Polaris 10 full fat being 2304 SPs.