Xbox One (Durango) Technical hardware investigation

[Gubbi]

Defects don't magically disappear and allow you to maintain yields when you increase the SOC size.

They do when you build in redundancy which you fuse off if a defect is detected.

They have a CPU core reserved for this. They probably have two CUs reserved as well. The eSRAM is easily made redundant wrt. defects. The bits that are most vulnerable are all the interconnect stuff where you really can't build in redundancy, here PS4's and XB1's SOC are similar.

Cheers

 

[Strange]

They do when you build in redundancy which you fuse off if a defect is detected.

They have a CPU core reserved for this. They probably have two CUs reserved as well. The eSRAM is easily made redundant wrt. defects. The bits that are most vulnerable are all the interconnect stuff where you really can't build in redundancy, here PS4's and XB1's SOC are similar.

Cheers

Why wouldn't the PS4's SOC benefit from the same increase in yields through the exact same processes like redundancy leading to (again) better yields?

Isn't 18 CUs a good indicator that it might actually be 20 CUs and 2 for redundancy?

And Jaguar cores come in fours. I don't think you can build redundancy in that.

You're attributing lots of things to the Xbox One SOC to improve it's yields while ignoring that the PS4's SOCs can also benefit from the exact same reasons. In the end they pretty much cancel out each other and you're back to square one- SOC size dictating your yield percentages.

 

[kalelovil]

Isn't 18 CUs a good indicator that it might actually be 20 CUs and 2 for redundancy?

No. The PS4 and Xbox One APUs are new dies, how AMD have chosen to configure previous discrete products is of no direct bearing on them. Modern GPU design is very modular.

18CUs and 12CUs are in fact slightly more balanced layouts for GCN than the 20CUs and 14CUs used for Pitcairn and Bonaire respectively, since GCN groups CUs into CU arrays of 3-4 CUs sharing read-only L1 caches.
18CUs and 12CUs divide cleanly into 6 CU arrays, while in Pitcairn and Bonaire two of the arrays have a fourth CU hanging onto them increasing cache contention.

 

[Strange]

No. The PS4 and Xbox One APUs are new dies, how AMD have chosen to configure previous discrete products is of no direct bearing on them. Modern GPU design is very modular.

18CUs and 12CUs are in fact slightly more balanced layouts for GCN than the 20CUs and 14CUs used for Pitcairn and Bonaire respectively, since GCN groups CUs into CU arrays of 3-4 CUs sharing read-only L1 caches.
18CUs and 12CUs divide cleanly into 6 CU arrays, while in Pitcairn and Bonaire two of the arrays have a fourth CU hanging onto them increasing cache contention.

Ah I see. Thanks for the info.

 

[Gubbi]

Why wouldn't the PS4's SOC benefit from the same increase in yields through the exact same processes like redundancy leading to (again) better yields?

Isn't 18 CUs a good indicator that it might actually be 20 CUs and 2 for redundancy?

Of course Sony will do the same.

And Jaguar cores come in fours. I don't think you can build redundancy in that.

They could fuse off a core. Enumerate the cores 0-6 with core 0-3 in the full cluster.

You're attributing lots of things to the Xbox One SOC to improve it's yields while ignoring that the PS4's SOCs can also benefit from the exact same reasons.

No I did not, that is all in your head.

I fully expect the Sony SOC to employ the same redundancy.

In the end they pretty much cancel out each other and you're back to square one- SOC size dictating your yield percentages.

Size of the part of the die that can't easily be covered with redundancies, yes. And I'd expect this size to be roughly equal, hence yield to be roughly equal despite the larger MS die.

Cheers

 

[Strange]

Of course Sony will do the same.



They could fuse off a core. Enumerate the cores 0-6 with core 0-3 in the full cluster.



No I did not, that is all in your head.

I fully expect the Sony SOC to employ the same redundancy.



Size of the part of the die that can't easily be covered with redundancies, yes. And I'd expect this size to be roughly equal, hence yield to be roughly equal despite the larger MS die.

Cheers

So you're telling me that a 400mm^2 SOC will have roughly the same yields as a 300mm^2 SOC because the eSRAM is essentially "defect proof" due to redundancies.

Even if I give you that we're obviously forgetting that the eSRAM with 1.6 billion transistors is rated to be only @ around 60 mm^2 max, which doesn't even make up for the size difference.

 

[BeyondTed]

They do when you build in redundancy which you fuse off if a defect is detected.

They have a CPU core reserved for this. They probably have two CUs reserved as well. The eSRAM is easily made redundant wrt. defects. The bits that are most vulnerable are all the interconnect stuff where you really can't build in redundancy, here PS4's and XB1's SOC are similar.

Cheers

Am I understanding correctly that you mean there are 9 CPU cores and 8 active? Or is it 8 CPU cores and 7 active?

 

[3dilettante]

Even if I give you that we're obviously forgetting that the eSRAM with 1.6 billion transistors is rated to be only @ around 60 mm^2 max, which doesn't even make up for the size difference.

They could opt for something less dense. The data arrays for AMD's Jaguar L2 alone would be about half that dense, not including any of the interface.
AMD has at times blown their L3 density, which is the closest reference we have to a pool like the eSRAM. That was a different node and product, but building awesome large on-die memories is not what AMD is known for. I thought it interesting that Microsoft made that call anyway.

I'm watching out for a die shot to see if AMD has shaken its legacy.

 

[bkilian]

So you're telling me that a 400mm^2 SOC will have roughly the same yields as a 300mm^2 SOC because the eSRAM is essentially "defect proof" due to redundancies.

Even if I give you that we're obviously forgetting that the eSRAM with 1.6 billion transistors is rated to be only @ around 60 mm^2 max, which doesn't even make up for the size difference.

Um, if that is true, then the PS4 APU size we're positing is the wrong size, unless you're claiming that the same size CPU and a smaller GPU results in more mm^2? Ooh.. Maybe it's the secret sauce...

 

[Strange]

Um, if that is true, then the PS4 APU size we're positing is the wrong size, unless you're claiming that the same size CPU and a smaller GPU results in more mm^2? Ooh.. Maybe it's the secret sauce...

Go look at Grall, he's the one who came up with the numbers. I just pointed out the yield issue. :smile:

 

[Phil]

I understood that they were just merely numbers to demonstrate a hypothetical scenario / educated guess of how yields, die size and a reasonable assumption on DDR3 vs GDDR5 costs might turn out for each approach. I think the point was merely to point out that the DDR3 approach, despite the bigger die size, would end up being more costly. If the die sizes are even closer, that would only strenghten the case that DDR3 is the less expensive solution. Or that's at least how I understood it to be ment.

 

[MarkoIt]

Am I understanding correctly that you mean there are 9 CPU cores and 8 active? Or is it 8 CPU cores and 7 active?

8 CPU cores, 7 active. Jaguar comes only in 4 cores modules.

 

[Strange]

But the spec sheets specifically said 8 core IIRC, so I'm not sure if they'll disable any.

 

[XpiderMX]

But the spec sheets specifically said 8 core IIRC, so I'm not sure if they'll disable any.

Isn't one of the 8 cell SPU cores disabled to increase production yield?

 

[3dilettante]

Most of the spec sheets for that typically say 8 cores (SPEs+PPE) or say 7 SPEs. They do neglect to mention that one of the SPEs is reserved.

They could just bite the bullet and not accept chips with 7 cores. I could imagine some kind of Gakai or cloud instance salvage scheme, where the cloud machine has no need for one of the cores that would have gone towards apps. I'm not sure the volume would be there, but who knows.

 

[bkilian]

8 CPU cores, 7 active. Jaguar comes only in 4 cores modules.

I don't know about the PS4, but the XBox One has all 8 cores active.

 

[TheWretched]

Isn't one of the 8 cell SPU cores disabled to increase production yield?

Yes, but Sonys initial PS3 spec sheet mentions this specifically. It doesn't specify any disabled cores for PS4.

i.e. here http://www.scei.co.jp/corporate/release/pdf/050517e.pdf

Cell Processor
PowerPC-base Core @3.2GHz
1 VMX vector unit per core
512KB L2 cache
7 x SPE @3.2GHz
7 x 128b 128 SIMD GPRs
7 x 256KB SRAM for SPE
* 1 of 8 SPEs reserved for redundancy
total floating point performance : 218 GFLOPS

 

[XpiderMX]

Yes, but Sonys initial PS3 spec sheet mentions this specifically. It doesn't specify any disabled cores for PS4.

i.e. here http://www.scei.co.jp/corporate/release/pdf/050517e.pdf

Cell Processor
PowerPC-base Core @3.2GHz
1 VMX vector unit per core
512KB L2 cache
7 x SPE @3.2GHz
7 x 128b 128 SIMD GPRs
7 x 256KB SRAM for SPE
* 1 of 8 SPEs reserved for redundancy
total floating point performance : 218 GFLOPS

Thanks

Off-topic. Lol at the GPU flops

 

[Aeoniss]

I don't know about the PS4, but the XBox One has all 8 cores active.

And given all the press releases for both Xbox One and PS4 say "8 CPU cores" and in the PS4's case they state 1.84TF from the GPU its not likely that the figures we have (8 jaguar cores, 18 CU's) are going to be deactivated for yield purposes.

Or am I off on that?

 

[Shifty Geezer]

Thanks

Off-topic. Lol at the GPU flops

PS4's no more powerful than PS3. and XB1 is less powerful than PS4. So XB1 is less powerful than PS3!

Stands to reason, obviously.