PlayStation 4 (codename Orbis) technical hardware investigation (news and rumours)

[Shortbread]

Wiki is mistaken.

Radeon 7870 has 20CU's [most likely PS4 has them as well, without that buffer zone Sony wouldn't have that "phenomenal yields" Tretton mentioned] and it has size od 212mm2 and total transistor count of 2.8 billion. CPU section and other modules will not add much to that numbers.

Although I agree with you that Wiki is possibly mistaken... however, we don't know enough (officially) what Island family, what Sony-PS4 specific modification were made and whatever else that's undisclosed. Please could you link us to Trenton 2.8 billion transistor mention... I just find that odd he would mention that all, at any time.

 

[oldschoolnerd]

Taking the fillrate benchmarks of hardware.fr as reference (which use of course the ROPs to access the memory, read or write benchmarks limited by the memory bandwidth through the TMUs don't exist), the HD7850 or 7870 (also 32 ROPs and a 256Bit GDDR5 interface made by AMD) achieve about 93% of the peak bandwidth with writes and about 91% of the peak bandwidth with blending (mixed reads and writes, but longer bursts).

That's quite a lot higher than the competition...they only manage about 75%. I was wondering if Orbis would encounter more bus contention because of the cpu needing access ram down the same bus as the GPU? It is a factor that isn't present in the example you give.

 

[DieH@rd]

Tretton mentioned only "phenomenal yields" in the interview with the Fox. 2.8bil and 212mm2 are official numbers for Radeon 7870, posted by me as a starting point of PS4 chip size discussion.

 

[Love_In_Rio]

I just had another thought on this that's kinda out there but the numbers match up with the older Devkits being 10% slower.

What if the PS4 GPU is a modified HD 7870 & the yields was so good that they decided to use all 20 CU's?

18 CU's is 10% slower than 20 CU's.

Or something more plausible like final dev kits including Curacao instead of Pitcairn, being Curacao a 10% more efficient.

 

[Arwin]

Interesting, but is there anything to substantiate this? I'd think that if this were true Sony would be blasting it from the rooftops.

This.

 

[Shortbread]

Tretton mentioned only "phenomenal yields" in the interview with the Fox. 2.8bil and 212mm2 are official numbers for Radeon 7870, posted by me as a starting point of PS4 chip size discussion.

Ok, I see - as a reference point. I think before we get to the point of guessing/estimating the PS4 SoC size, we should figure out what "cooling solution" that they are using and go from there. Sony has been great (damn awesome) with coming up with unique solutions... the first model PS2s had a nice thermal heat dissipation design (EE/GS are some heat monsters).

 

[Deleted member 13524]

Don't trust wikipedia on such things. It's simply bullshit and probably edited in by some stupid fanboy.

Or someone just looking for some good laughs.

 

[BoardBonobo]

That's quite a lot higher than the competition...they only manage about 75%. I was wondering if Orbis would encounter more bus contention because of the cpu needing access ram down the same bus as the GPU? It is a factor that isn't present in the example you give.

I thought the CPU has a separate bus to access memory? One with 20GB/s bandwidth.

 

[onQ]

Or something more plausible like final dev kits including Curacao instead of Pitcairn, being Curacao a 10% more efficient.

Maybe.

I do find it funny that Sony hasn't showed off the PS4 chip yet.

 

[oldschoolnerd]

I thought the CPU has a separate bus to access memory? One with 20GB/s bandwidth.

I think it's limited to 20GB/s, but that is shared with the main 176GB/s one.

 

[McHuj]

I think it's limited to 20GB/s, but that is shared with the main 176GB/s one.

The memory itself can only physically supply 176 GB/sec. 256-bit bus @ 5.5 Gbps. Anything inside the memory controller and beyond can only share that bandwidth.

 

[oldschoolnerd]

The memory itself can only physically supply 176 GB/sec. 256-bit bus @ 5.5 Gbps. Anything inside the memory controller and beyond can only share that bandwidth.

So there is going to be bus contention then. That's going to reduce the 91% figure mentioned earlier, as that ram was dedicated to the GPU. Any advances on 91%?

 

[Pixel]

The memory itself can only physically supply 176 GB/sec. 256-bit bus @ 5.5 Gbps. Anything inside the memory controller and beyond can only share that bandwidth.

I think it's limited to 20GB/s, but that is shared with the main 176GB/s one.

This isn't an issue with ddr3 though. So are you sure its an issue with gddr5?
As we found out in X1 vs DF article, according to the engineer as long as ddr3 has a seperate memory controller for each bus it can still simultaneously hit each busses peak bandwidth on each bus.

 

[3dilettante]

The GPU can handle the combined bandwidth of the two memory systems.
The CPU block, as far as we can tell from the leaks, can't max either of them.

 

[oldschoolnerd]

This isn't an issue with ddr3 though. So are you sure its an issue with gddr5?
As we found out in X1 vs DF article, according to the engineer as long as ddr3 has a seperate memory controller for each bus it can still simultaneously hit peak bandwidth on each bus.

I don't think it's a ddr3 vs gddr5 issue, it's more the fact that on the competing console the vast majority of the GPU memory requests are serviced out of esram, so less contention for system memory. The ps4 only has the gddr5 ram, so all the GPU ram requests have to be serviced out of there. Hence the increased contention.

 

[bkilian]

This isn't an issue with ddr3 though. So are you sure its an issue with gddr5?
As we found out in X1 vs DF article, according to the engineer as long as ddr3 has a seperate memory controller for each bus it can still simultaneously hit each busses peak bandwidth on each bus.

What? No it can't. And the DF article didn't say anything of the sort.

 

[oldschoolnerd]

What? No it can't. And the DF article didn't say anything of the sort.

I think he's talking about the competing console, where the GPU has the potential to get 150GB/s out of esram plus 50GB/s out of the system ram.

 

[itsmydamnation]

I don't think it's a ddr3 vs gddr5 issue, it's more the fact that on the competing console the vast majority of the GPU memory requests are serviced out of esram, so less contention for system memory. The ps4 only has the gddr5 ram, so all the GPU ram requests have to be serviced out of there. Hence the increased contention.

complete load of crap.

1. its only 32 mb
2. its only 32 mb
3. ITS ONLY 32 MB

the point of it is to write intermediate steps to it and maybe in the future help with compute via lower latency.

You still will be reading and writing memory between the esram and dram, I would be worried far more about Xb1 memory contention then PS4.

 

[oldschoolnerd]

complete load of crap.

1. its only 32 mb
2. its only 32 mb
3. ITS ONLY 32 MB

the point of it is to write intermediate steps to it and maybe in the future help with compute via lower latency.

You still will be reading and writing memory between the esram and dram, I would be worried far more about Xb1 memory contention then PS4.

Wow. I didn't mean to make anyone angry...

I thought that writing intermediate steps was most of what the GPU did with it's memory bandwidth...it being a pipeline and all.

Still, any advances on 91% for the real world vs peak bandwidth question?

 

[Esrever]

complete load of crap.

1. its only 32 mb
2. its only 32 mb
3. ITS ONLY 32 MB

the point of it is to write intermediate steps to it and maybe in the future help with compute via lower latency.

You still will be reading and writing memory between the esram and dram, I would be worried far more about Xb1 memory contention then PS4.

You will be reading and writing it continuously to both ram pools. I really doubt it will just be a simple read/write buffer. 32MB isn't a lot but it can store enough for the most needed bits of memory, and the large texture can be stored in main memory. Writing between esram and main memory can be done by the move engines.