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

[Gipsel]

Who is talking about special sauce?

the charges sony made have been detailed.

The three "major modifications" Sony did to the architecture to support this vision are as follows, in Cerny's words:
...

And all those three points can be found in already in Temash/Kabini as well as the Bonaire discrete GPU (HD7790). Okay, just 32 compute queues, but Kaveri will have 64, too.
And the first point about bypassing the GPU caches, this can be controlled already in the original GCN architecture through the SLC (system level coherent) bit in the encoding for each individual memory instruction (it bypasses all caches, the also existing GLC bit causes the bypass of just the L1). If the original GCN would have been integrated in an APU with unified memory (as PS4), this would (and in Kabini it does) cause exactly the behaviour Cerny described.

We had this discussion before, Sony may have influenced some details of the solutions, but for the mentioned things AMD would have been integrated them in a largely similar way in any case in future APUs and GPUs. There are actually more modifications than that (like the additional graphics command processor of the PS4), but that's a bit harder to explain in a largely PR driven reveal event.

 

[grndzro]

?
Where did you hear that Sony helped design the memory controller? I don't recall hearing anything like that.

It's the only piece of the hardware that is significantly different from the other bits that are not being used in the PC sector.

And Sony's contract with AMD states there are proprietary bits inside the APU that belong to sony. I can't think of anything else inside the custom Jaguar APU that Sony would be keen on making sure no one else has.

It is a guess but I'd bet that it's spot on the money since 3rd party development started on PS4 titles before MS got the ball rolling.

MS had to face the prospect of either forking out boatloads of cash to get titles ported to a new architecture or go with the flow and find a way to make the Jaguar work. The result is the 32MB eSRAM block attached to the GPU side, which is probably responsible for the slightly weaker Xbox one GPU due to yield managment.

Bandwidth wise both consoles are just fine because once textures and such are loaded onto the eSRAM it will take a huge load off of the DDR3. The Playstation's GDDR5 will be a bit better off bandwidth wise though.

The thought that AMD, which had already been making GPUs using GDDR5 and already had UMA technology in the works, was incapable of moving forward without Sony is just inconceivable. Thanks for the laughs.

How long has AMD had UMA in the works? around 4 years right? That's how long AMD had been working with Sony on the PS4.

The GDDR5 bit might just be a contractual obligation. Sony has some very good engineers that could have helped AMD out a bunch.

 

[ERP]

Your drawing a lot of conclusions from your own speculation.

 

[grndzro]

Your drawing a lot of conclusions from your own speculation.

Doesn't everyone draw conclusions about their own speculation?

 

[Gipsel]

The GDDR5 bit might just be a contractual obligation. Sony has some very good engineers that could have helped AMD out a bunch.

Hasn't AMD co-developed the GDDR5 standard? When did AMD's first chip with a 256Bit GDDR5 interface (HD4870) appear?

 

[grndzro]

Hasn't AMD co-developed the GDDR5 standard? When did AMD's first chip with a 256Bit GDDR5 interface (HD4870) appear?

1950xtx

 

[Gipsel]

1950xtx

The X1950XTX had GDDR4 (and was the first one to use that as AMD also co-developed GDDR4 as well as they are co-developing GDDR6 in collaboration with the memory industry).

 

[patsu]

Will wait for the xrays

'Not intentionally 100% round'. That means it's bumpy, with ALU's jutting out like a mountain on a globe.

And if these ALU mountains are high enough, we should see clouds(TM) engulfing their peaks.

Seriously the best way to illustrate all these customization is to use some code case studies.

 

[bagofsuck]

i wonder why DigitalFoundry didn't ask or (why there wasn't any information )about system resources,OS footprint?..might be they are avoiding those questions?

 

[Strange]

A while back Sony had stated that OS footprint is not open information, so I don't think DF is going to waste time asking a question that everybody already knows the answer to: we're not telling.
Better to ask new questions and get more interesting responses.

 

[Tap In]

A while back Sony had stated that OS footprint is not open information, so I don't think DF is going to waste time asking a question that everybody already knows the answer to: we're not telling.
Better to ask new questions and get more interesting responses.

if it is as low as people are speculating (1GB) why not say it?

sounds to me it's more like 2-3

 

[Kaarlisk]

There is one thing that I as a lamer do not understand:

Latency in GDDR5 isn't particularly higher than the latency in DDR3. On the GPU side... Of course, GPUs are designed to be extraordinarily latency tolerant so I can't imagine that being much of a factor.

I thought that GDDR5's latency, if it is a problem, may be a problem for the CPU cores, not the GPU?

 

[Tap In]

There is one thing that I as a lamer do not understand:

I thought that GDDR5's latency, if it is a problem, may be a problem for the CPU cores, not the GPU?

it is allegedly, sounds like he was just shifting the argument

 

[bkilian]

There is one thing that I as a lamer do not understand:

I thought that GDDR5's latency, if it is a problem, may be a problem for the CPU cores, not the GPU?

From what I've heard, gddr5 latency is about the same in nanoseconds as ddr3 latency. It is a lot more memory cycles, but the memory is clocked a lot higher.

 

[cal_guy]

1950xtx

AMD's first GDDR5 card was the HD 4870 which was released in 2008, have one myself. Given that AMD essentially designed the GDDR5 spec I doubt that they needed help on the memory controller.

 

[Hecatoncheires]

How likely did Sony choose 8 ACEs because of the 8 Jaguars? Maybe this is the origin of the 14+4 rumor since there were also rumors about 4 big Steamroller cores. Maybe it was just a 14CUs + 4 ACEs iGP a long time ago and now they have a 18CUs + 8 ACEs iGP with smaller Jaguars cores.

...implies...

Well, he also explained a couple of times that he wanted x86 for PS4 ab ovo. And if you think back a little there were many rumors about stacking, wide-i/o, TSVs and interposers.

 

[TrungGap]

if it is as low as people are speculating (1GB) why not say it?

sounds to me it's more like 2-3

It's definitely better NOT to say thing...because it's basic marketing 101.

Saying you have 1 (low) OS memory reservation, you're going to give an impression to some that this is a gaming machine only. Whether it's true or not, is not the point. People going to form an impression, if there's partial data...and it might be a wrong or right impression, but you might alienate some of your potential buyers. Saying you have 2-3 (higher) OS memory reservation, going to give a different impression.

So what you really want to do is to control the narrative to something that will sell your product.

 

[Cjail]

^^^

 

[Brad Grenz]

And all those three points can be found in already in Temash/Kabini as well as the Bonaire discrete GPU (HD7790). Okay, just 32 compute queues, but Kaveri will have 64, too.
And the first point about bypassing the GPU caches, this can be controlled already in the original GCN architecture through the SLC (system level coherent) bit in the encoding for each individual memory instruction (it bypasses all caches, the also existing GLC bit causes the bypass of just the L1). If the original GCN would have been integrated in an APU with unified memory (as PS4), this would (and in Kabini it does) cause exactly the behaviour Cerny described.

We had this discussion before, Sony may have influenced some details of the solutions, but for the mentioned things AMD would have been integrated them in a largely similar way in any case in future APUs and GPUs. There are actually more modifications than that (like the additional graphics command processor of the PS4), but that's a bit harder to explain in a largely PR driven reveal event.

We've had AMD employees on this forum confirm that technologies and enhancements made for these consoles have worked their way into the PC product line. Think of it this way, Sony has very talented hardware designers and the most technologically accomplished group of developers at WWS. It's all but inevitable that their input beginning back in 2008 or 2009 would have helped to shape the way forward for both GCN and HSA.

 

[Deleted member 11852]

And if these ALU mountains are high enough, we should see clouds(TM) engulfing their peaks.

Cloud power all round then I think it's fairly clear, in my mind at least, how this 14+4 rumour started. VGLeaks originally reported the following key GPU information:

• 18 Compute Units (CUs)
• Hardware balanced at 14 CUs
• 1.843 Tflops, 922 GigaOps/s

The key part is hardware balanced at 14 units CUs. Now let's assume for a moment that Microsoft and Sony don't simply roll dice to determine the correct proportion of hardware required for general computation (CPU) and graphics (GPU). Each will have a target performance level for both. In terms of graphics Microsoft and Sony end up in the same ballpark; 12 CUs for Xbox, 14 CUs for PS4 but Sony are gambling that compute will be big in a couple of years. The Jaguar architecture already accommodates GPGPU compute but using compute can deprive graphics so Sony decide to throw in and extra four CUs, above and beyond what it already thinks is a sensible balance for CPU/GPU, to accommodate future compute use but don't establish arbitrary restrictions on the CUs division of labour, this control is put in the hands of the developer.

Now during their developer briefings and their dev documentation, this vision of split between CUs in graphics and compute is outlined, and suddenly we have the 14+4 rumour. Cerny's response to Eurogamer suggests this is what happened:

That comes from a leak and is not any form of formal evangelisation. The point is the hardware is intentionally not 100 per cent round. It has a little bit more ALU in it than it would if you were thinking strictly about graphics. As a result of that you have an opportunity, you could say an incentivisation, to use that ALU for GPGPU.

Sony weren't evangelising, that is trying to convert developers to adopt their prophesied 14/4 graphics/compute balance, but it leaked that way then the additional ALU stuff began from the references of more ALU than needed for graphics.