Difference between XB1 and PS4 GPU architectures *spawn

Status
Not open for further replies.

Commenter

Newcomer
I just noticed that Xbone's GPU (Radeon 7790) has opencl 2.0 support vs 1.2 (Radeon 7870) in the PS4. Is that right? I think one of the main additions is support for dynamic parallelism.
 
I just noticed that Xbone's GPU (Radeon 7790) has opencl 2.0 support vs 1.2 (Radeon 7870) in the PS4. Is that right? I think one of the main additions is support for dynamic parallelism.
Where are you getting that from?? What makes you think Sony's GPU is a year older architecture than MS's?
 
I just noticed that Xbone's GPU (Radeon 7790) has opencl 2.0 support vs 1.2 (Radeon 7870) in the PS4. Is that right? I think one of the main additions is support for dynamic parallelism.
AMD OpenCL 2.0 driver essentially compatible with all GCN gpu. Some of the listed compatible GPU even belong in GCN 1.0.
 
There are mainly articles popping up around the web with regards to this, the main points that I can gather are to do with L1 cache on each CU/SC compared to the PS4 sharing L1 between 3 CU's and the Durango leaks that show 4 SMIDS each SMID has 4 VSP's with each VSP able to do 4 threads which is more akin to the later GFX cards (the SMID details are in the leaked SDK which basically justifies alot of the original Durango leak, esp. with regards to the CU/SC.
I have not seen anything to do with opencl support for either console, but the whole dual pipe and 2 graphics command processors is leading many to believe you will be able to saturate all CU's instead of waiting for them to finish (like on PS4) even if the current task is only using 2 CU's

PS. Don't shoot the messenger
 
There are mainly articles popping up around the web with regards to this, the main points that I can gather are to do with L1 cache on each CU/SC compared to the PS4 sharing L1 between 3 CU's and the Durango leaks that show 4 SMIDS each SMID has 4 VSP's with each VSP able to do 4 threads which is more akin to the later GFX cards (the SMID details are in the leaked SDK which basically justifies alot of the original Durango leak, esp. with regards to the CU/SC.
I have not seen anything to do with opencl support for either console, but the whole dual pipe and 2 graphics command processors is leading many to believe you will be able to saturate all CU's instead of waiting for them to finish (like on PS4) even if the current task is only using 2 CU's

PS. Don't shoot the messenger

What are these articles and what evidence do they have? Because the 2CU thing itself is false, a compute task can use as many CUs as it wants, it just defaults to 4.
 
The evidence is based off Durango and the leaked SDK, I should point out I am only writing what conclusions others have come up with, it's not necessarily my own :)

Sorry I am not explaing it properly :(
I was trying to say that if the graphics command processor issues a task then it would use X amount of CU's to complete the task. Without 2 graphics command processors (which the PS4 does not have) you would have to wait for the single one to finish which would be wasting CU's or cycles would it not ? I am still trying to understand, apologies.

I was saying 2 as an extreme example a more accurate one would be 10 CU's used on a command issued leaving 4 idle which could be used if a second command could be issues in the same cycle.
 
No. The GCP will not stall waiting for compute to finish. The rendering pipeline has many different aspects to it and compute only AFAIK only leverages the CUs. All GCN will handle dispatch jobs the same IIRC, as I understand you just fire them off to ACE and if it's available they go in. Ps4 has 64 threads available for compute Xbox has 7.

The reason why compute shaders are a big deal is because you can have your graphics pipeline working on something else like shadows while you use compute shaders to perform some lighting and particle effects. This has been around since dx11 cards have been launched. The inclusion of ACE enabled them to fit additional compute jobs when the CUs are waiting between work AFAIK.

Many of these things are answered, you need to do some research and at the very least if you are going to cite an article post the article here. If you can't at least provide that much than you should have a pretty strong idea that the claim is false. Anything real and anecdotal would have made it to the larger sites.
 
The evidence is based off Durango and the leaked SDK, I should point out I am only writing what conclusions others have come up with, it's not necessarily my own :)

Sorry I am not explaing it properly :(
I was trying to say that if the graphics command processor issues a task then it would use X amount of CU's to complete the task. Without 2 graphics command processors (which the PS4 does not have) you would have to wait for the single one to finish which would be wasting CU's or cycles would it not ? I am still trying to understand, apologies.
http://www.amd.com/Documents/GCN_Architecture_whitepaper.pdf

You have a number of resources on GPU. The command processor(s) dispatch instructions to these resources. The GCP handles streams for different shader types (vertex, hull, domain, geometry and pixel shaders) and fixed function operations. So the GCP queues work and shouldn't ever see idle hardware as long as work is provided, while the ACE's will slot compute work in between any lulls (graphics work dependent on other input). The introduction of high priority compute enables compute to take precedent which should be important in future.

So seeing as the CCP and GCP already provide load balancing with fine granularity and multiple threads (as I understand it, a stream is a thread), it's not easy to see the advantage in having a second GCP. It'll have queues for different jobs wanting to use the same resources as the other GCP. Both GCP have a load of vertex and pixel and geometry shaders wanting to use the same resources. One could split the hardware up to provide half for one GCP and half for the other, allowing both to function at once, but the actual time to completion will still be the same. ie. take 10 ms to complete both GCP's streams in parallel, or take 5 ms to complete the first job and then 5 for the second job.

TBH the discussion of the functionality of a second GCP in an AMD GPU should be carried out in the 3D Hardware forum. You'll get truly knowledgeable folk contributing then!
 
AMD states the following cards as supporting Open Cl 2.0:

AMD Radeon™ R9 200 Series AMD Radeon™ HD 8950
AMD Radeon™ R7 200 Series AMD Radeon™ HD 8600 Series
AMD Radeon™ HD 7700 Series AMD Radeon™ HD 8500 Series

The Xbox 7790 is supported because it is a 7700 series card.
The PS4 7870 is supported because all 7800 and 7900 series cards were rebranded as R9 200.

Is your source MIsterXmedia? :)
 
I was pointing out other peoples comments not my own, I did make that clear. You have seen them on reddit when you posted in a similar thread.

I just feel that there is a point to having 2 GCP's for games not one for UI/OS and one for games.

I am going to bow out this place clearly is not for me.
 
ItCHWSz.jpg
 
Status
Not open for further replies.
Back
Top