NGGP: NextGen Garbage Pile (aka: No one reads the topics or stays on topic) *spawn*

[Pugger]

That sounds reasonable, but why would he think it would have the upper hand with regards to geometry processing? The data move engines and ESRAM wouldn't really help with that, would it, certainly not while Orbis is packing more CUs.

Unless there are as yet unrevealed customisations to Durango's GPU that give it a vertex/tri advantage.

Your guess is as good as mine, I suspect thuway knows more about Orbis than he does Durango and he's passing on information from another poster who also posted geometry speculation.

 

[Iamshadwolark]

... aaaaaaand we're back to PS4 being twice as powerful as Xbox 3!

Some math you got there...

 

[Love_In_Rio]

Alright, I didn't realize this!, or didn't read it through the noise.
I had no idea why you talked about APU+GPU. This makes better sense.

What could be argued : is a 18 CUs APU with context switch better than 4 CUs + 14 CUs without context switch. I'll feel like saying yes. But that would be a theoretical argument with no relevance to what's in Durango or Orbis.

So, 12 CUs in Durango vs 4 + 14 in Orbis.

14+4 without context switching > 12 with context switching

Context switching only wouldn´t be enough. So, latency will be the clue to know if that situation reverse, as then you would have 12 CUs in Durango processing way more percentage of cycles.

On other hand... wouldn´t have sense in an architecture more GPGPU friendly to have more geometry engines?. In Orbis it looks like they are still two. Maybe once of the changes made to Durango´s GPU is to add more, something similar to Fermi-Kepler architecture.

 

[Shifty Geezer]

An APU that only deals with GPGPU algorithms and a GPU that only deals with graphics rendering will deliver a much better performance. In the case of Orbis we have the same thing (if the rumors are correct), but integrated into a SoC: The APU consists of 8 Jaguars and 4 GCN CUs, the GPU consists of 14 GCN CUs.

That's not at all how the VGLeaks article reads. It says 8 core CPU and 18 CU GPU. The diagram shows CPU and GPU connected via memory controller. There's no APU+GPU or SOC described, or separation of the 4 CU computer hardware from the 14 CU GPU.

 

[function]

Some math you got there...

My math? You might want to check the working on that.

 

[Hecatoncheires]

That's not at all how the VGLeaks article reads. It says 8 core CPU and 18 CU GPU. The diagram shows CPU and GPU connected via memory controller. There's no APU+GPU or SOC described, or separation of the 4 CU computer hardware from the 14 CU GPU.

In fact there is a SoC described in the VGLeaks article: It says "Liverpool SOC" at the very beginning.

I'm not talking about a physical seperation of elements that are dedicated to graphics rendering and elements that are dedicated to computing. I'm just abstracting how I understand this Liverpool SoC works. If you have 4 GCN CUs dedicated to computational works (Leadbetter described the same phenomenon), then it's basically nothing else as if you had an APU consisting of 8 Jaguars and 4 GCN CUs. It's all on the same die anyway, so physically it is one big processor, a single APU, but according to the rumors the elements are dedicated to special tasks (4CUs doing GPGPU, 14 CUs doing graphics rendering), so the concept behind this SoC is having an APU and a GPU combined in a single SoC - in the abstract.

 

[upnorthsox]

In fact there is a SoC described in the VGLeaks article: It says "Liverpool SOC" at the very beginning.

I'm not talking about a physical seperation of elements that are dedicated to graphics rendering and elements that are dedicated to computing. I'm just abstracting how I understand this Liverpool SoC works. If you have 4 GCN CUs dedicated to computational works (Leadbetter described the same phenomenon), then it's basically nothing else as if you had an APU consisting of 8 Jaguars and 4 GCN CUs. It's all on the same die anyway, so physically it is one big processor, a single APU, but according to the rumors the elements are dedicated to special tasks (4CUs doing GPGPU, 14 CUs doing graphics rendering), so the concept behind this SoC is having an APU and a GPU combined in a single SoC - in the abstract.

Then your understanding is wrong

 

[Shifty Geezer]

In fact there is a SoC described in the VGLeaks article: It says "Liverpool SOC" at the very beginning.

They might call it that, but it's not an SOC if it doesn't include the system aspect. A combined CPU+GPU does not an SOC make. Of course, if the RAM is on there, some may feel that warrants the SOC nomenclature.

I'm not talking about a physical seperation of elements that are dedicated to graphics rendering and elements that are dedicated to computing. I'm just abstracting how I understand this Liverpool SoC works...

Okay.

 

[mrcorbo]

They might call it that, but it's not an SOC if it doesn't include the system aspect. A combined CPU+GPU does not an SOC make. Of course, if the RAM is on there, some may feel that warrants the SOC nomenclature.

I always thought SoC implied CPU, GPU, North Bridge and South Bridge integration.

 

[arijoytunir]

I always thought SoC implied CPU, GPU, North Bridge and South Bridge integration.

so no stacked ram

 

[Hecatoncheires]

Then your understanding is wrong

To what extent?

VGLeaks says: "4 additional CUs (410 Gflops) “extra” ALU as resource for compute"

Leadbetter says: Paired up with the eight AMD cores, we find a bespoke GPU-like "Compute" module, designed to ease the burden on certain operations - physics calculations are a good example of traditional CPU work that are often hived off to GPU cores. We're assured that this is bespoke hardware that is not a part of the main graphics pipeline but we remain rather mystified by its standalone inclusion, bearing in mind Compute functions could be run off the main graphics cores and that devs could have the option to utilise that power for additional graphical grunt, if they so chose.

AMD's HSA is all about GPGPU. Going with AMD for next gen without utilizing GPGPU algorithms makes no sense. AMD itself recommends dedicated hardware for GPGPU algorithms in video games.

 

[mrcorbo]

so no stacked ram

It can include it, I just don't think it has to to qualify as a SoC.

I don't expect stacked RAM for Durango and Orbis, though. For the reasons that have already been specified. Maybe in future iterations of the hardware, but in that case it will be for cost-reduction not for performance.

 

[msia2k75]

It can include it, I just don't think it has to to qualify as a SoC.

I don't expect stacked RAM for Durango and Orbis, though. For the reasons that have already been specified. Maybe in future iterations of the hardware, but in that case it will be for cost-reduction not for performance.

Sure bet imo.

 

[adis]

somebody mentioned next gen consoles amd jaguar cpu 102.4 glops.. but I can only calculate 51.2 glops.. at least 4x less then 360/ps3.. scarry? partly explains Wii U CPU? even I still can't understand why Wii U CPU isn't at least quad-core

*edit: even then still can't

 

[MTd2]

There is another alternative.

Durango = 0 + 12 CU = X number of shaders
Orbis = 12 + 6 CU = same X number of shaders.

Consider that Durango uses only the special computer units and Orbis mixes them.

 

[Hornet]

somebody mentioned next gen consoles amd jaguar cpu 102.4 glops.. but I can only calculate 51.2 glops.. at least 4x less then 360/ps3.. scarry? partly explains Wii U CPU? even I still can't understand why Wii U CPU isn't at least quad-core

(1 MUL + 1 ADD) * (4 operands) * (8 cores) * (1.6 GHz) = 102.4 GFLOPS

 

[adis]

specs only mention 128 bit FPU/SIMD..

 

[Hornet]

specs only mention 128 bit FPU/SIMD..

It's single precision, hence 4 32-bit opeands.

 

[dumbo11]

There is another alternative.

Durango = 0 + 12 CU = X number of shaders
Orbis = 12 + 6 CU = same X number of shaders.

The weird bit is that Kinect would be appear a 'great' match for running on CUs. (in the same way that the Move algorithm can be run on a CU, the early processing of the Kinect data is apparently well-suited to the same thing).

I am wondering if the numbers for Durango are simplified or incorrect - i.e. is Durango 12+(N reserved for Kinect)?

 

[adis]

It's single precision, hence 4 32-bit opeands.

yes but from where you got (1 MUL + 1 ADD) ... fma??