AMD Kaveri APU features the Onion + bus like the PlayStation 4

[Betanumerical]

It IS read/write, on the other thread someone has been trying to spin that the having no full cache coherency thus it means no coherent memory, or as someone like to keep on repeating, coherent read only.

Let's try this one more time, coherent memory means that a pointer can be passed from the CPU to the GPU directly. Whether you need flush the cache or not (cache coherency) is an implementation detail, in X1's case, the code decides whether GPU wants snoop into CPU cache or not, and that the CPU can't snoop other components' cache.

I think you are still spinning, there's no evidence suggesting that PS4 CPU can snoop GPU cache, in fact, it's quite the contrary.

That is not required for cache coherency although it is a good way to achieve it, the PS4 on the other hand does have a coherent write bus. Whilst we know the XBONE can only coherently read.

Also you are 100% incorrect about coherency meaning that you can pass pointers between the CPU and GPU, all that requires is that you have the same style paging and virtual memory / MMU setup on both devices, cache coherency and coherent read/write between devices requires you to either be able to read from a device cache or to bypass your own cache or in the worst case flush it although that is a bad idea and you dont want to do that.

 

[taisui]

That is not required for cache coherency although it is a good way to achieve it, the PS4 on the other hand does have a coherent write bus. Whilst we know the XBONE can only coherently read.

Also you are 100% incorrect about coherency meaning that you can pass pointers between the CPU and GPU, all that requires is that you have the same style paging and virtual memory / MMU setup on both devices, cache coherency and coherent read/write between devices requires you to either be able to read from a device cache or to bypass your own cache or in the worst case flush it although the last is not really coherency.

What does coherent write even mean?

Keep on spinning, the PS4 needs to flush the GPU cache if it uses it just like the X1, before the CPU can read the data. It can also bypass the cache completely, but direct memory read/write is slower that way, and you can't claim it's cache coherent because NO CACHE is used.

 

[Betanumerical]

What does coherent write even mean?

Keep on spinning, the PS4 needs to flush the GPU cache if it uses it just like the X1, before the CPU can read the data. It can also bypass the cache completely, but direct memory read/write is slower that way, and you can't claim it's cache coherent because NO CACHE is used.

The PS4's cache flushing in the GPU is much more efficient, it only flushes the dirty compute lines, the XBONEs flushes the entire L2 cache and RO data. So your already wrong for 'just like the XONE'.

Coherent write means that the CPU can write to the memory in such a way that its cached values also go there. This allows the CPU to see any changes the GPU had cached. Theres multiple ways to do this the worst of which is a cache flush.

 

[taisui]

The PS4's cache flushing in the GPU is much more efficient, it only flushes the dirty compute lines, the XBONEs flushes the entire L2 cache and RO data. So your already wrong for 'just like the XONE'.

Coherent write means that the CPU can write to the memory in such a way that its cached values also go there. This allows the CPU to see any changes the GPU had cached. Theres multiple ways to do this the worst of which is a cache flush.

Straw man, I even quoted Cerny on the top, READ.

What the hell does this even mean.....? PS4 GPU already can snoop the CPU cache via Onion bus.

 

[Betanumerical]

Straw man, I even quoted Cerny on the top, READ.

So how exactly is PS4 coherent "write" if the GPU needs to flush, come again?

It can bypass the cache as well or selectively flush the dirty cache compute lines, this is a much better strategy then just flushing everything.

Straw man, I even quoted Cerny on the top, READ.

What the hell does this even mean.....? PS4 GPU already can snoop the CPU cache via Onion bus.

It means that just because the GPU can read the CPUs L2 cache doesn't mean the CPU can see any values cached by the GPU...

 

[taisui]

It can bypass the cache as well or selectively flush the dirty cache compute lines, this is a much better strategy then just flushing everything.

...and where did you see me saying that's it's not a good thing?
I said it ALREADY.

It means that just because the GPU can read the CPUs L2 cache doesn't mean the CPU can see any values cached by the GPU...

Ok...you are not making sense now, see below

Coherent write means that the CPU can write to the memory in such a way that its cached values also go there. This allows the CPU to see any changes the GPU had cached. Theres multiple ways to do this the worst of which is a cache flush.

 

[Betanumerical]

...and where did you see me saying that's it's not a good thing?
I said it ALREADY.



Ok...you are not making sense now, see below

I don't know how to make it any clearer, you need to move the cached values to main memory.

 

[taisui]

I don't know how to make it any clearer, you need to move the cached values to main memory.

Well so tell me what's the difference that makes X1 coherent read-only and PS4 fully coherent?
Am I missing something here? can someone else enlighten me.

 

[Betanumerical]

Well so tell me what's the difference that makes X1 coherent read-only and PS4 fully coherent?
Am I missing something here? can someone else enlighten me.

well technically it probably does have it, I just found it interesting that in the interview with the architects that they never mentioned there coherent write advantage when it should be the same as the read.

aside from the technicalities its unlikely to get any use due to the large penalty in involved in flushing the L2 cache (it would seem to me it effectively stalls the GPU until the data is read back).

So from a technical standpoint? its 90% likely its both, but from a practical standpoint its only read.

 

[adev]

well technically it probably does have it, I just found it interesting that in the interview with the architects that they never mentioned there coherent write advantage when it should be the same as the read.

aside from the technicalities its unlikely to get any use due to the large penalty in involved in flushing the L2 cache (it would seem to me it effectively stalls the GPU until the data is read back).

So from a technical standpoint? its 90% likely its both, but from a practical standpoint its only read.

The Xbox One claims that the GPU can write directly to CPU L2, that's one method for coherent write. I'm not sure exactly how these writes are implemented. To suggest it doesn't have any coherent write capability would then be false.

The Onion bus on both is the same I presume so I imagine the PS4 GPU can also write to the CPU's L2 but given that the two platforms claim different bandwidth figures I don't think we can take that for granted. If the GPU writes to CPU L2 and doesn't just invalidate lines then presumably the CPU is freed from having to stall waiting on the GPU flush. The GPU values would still have to be flushed to memory at some point.

Onion+ on the PS4 can bypass the GPU cache and write directly to memory at 10Gbps.

 

[Betanumerical]

The Xbox One claims that the GPU can write directly to CPU L2, that's one method for coherent write. I'm not sure exactly how these writes are implemented. To suggest it doesn't have any coherent write capability would then be false.

The Onion bus on both is the same I presume so I imagine the PS4 GPU can also write to the CPU's L2 but given that the two platforms claim different bandwidth figures I don't think we can take that for granted. If the GPU writes to CPU L2 and doesn't just invalidate lines then presumably the CPU is freed from having to stall waiting on the GPU flush. The GPU values would still have to be flushed to memory at some point.

Onion+ on the PS4 can bypass the GPU cache and write directly to memory at 10Gbps.

I haven't see the first claim anywhere can you show me a link, thanks. The stall im talking about involves the GPU, if you flush the entire L2 cache of the GPU it can't even exactly do anything until all the data is read back in right?.

 

[taisui]

I haven't see the first claim anywhere can you show me a link, thanks. The stall im talking about involves the GPU, if you flush the entire L2 cache of the GPU it can't even exactly do anything until all the data is read back in right?.

He didn't provide a source for PS4 GPU writing to CPU L2 either, you didn't have a problem with that...

 

[Betanumerical]

He didn't provide a source for PS4 GPU writing to CPU L2 either, you didn't have a problem with that...

If he could provide both it would be nice, and if you could start posting constructively that would also be nice.

 

[adev]

I haven't see the first claim anywhere can you show me a link, thanks. The stall im talking about involves the GPU, if you flush the entire L2 cache of the GPU it can't even exactly do anything until all the data is read back in right?.

I don't have a link or implementation details, it's a single line reference that the GPU can write to CPU L2.

The GPU flush is under programmer control.

 

[Betanumerical]

I don't have a link or implementation details, it's a single line reference that the GPU can write to CPU L2.

The GPU flush is under programmer control.

Interesting it seems works differently to how I suspected. Or maybe, depending on the implementation details I guess.

 

[Rockster]

From vgleaks re: XB1... http://www.vgleaks.com/durango-memory-system-overview/

The GPU also has a coherent read/write path to the CPU’s L2 caches and to DRAM.

For each read and write request from the GPU, the request uses one path depending on whether the accessed resource is located in “coherent” or “non-coherent” memory.
...
As the GPU is I/O coherent, data in the GPU caches must be flushed before that data is visible to other components of the system

I think Beta's question is whether or not that means that the entire GPU cache must be flushed or if it can selectively flush lines ala the PS4.

 

[temesgen]

He didn't provide a source for PS4 GPU writing to CPU L2 either, you didn't have a problem with that...

Not to be a jerk but why are you so hostile, Beta hasn't shown any attitude towards you but for some reason you and another poster in another seem to take any comment questioning anything related to a certain platform quite personally. We all love tech and as such ask questions, tear down answers and should be able to do so without being defensive or overly aggressive.

 

[Betanumerical]

From vgleaks re: XB1... http://www.vgleaks.com/durango-memory-system-overview/


I think Beta's question is whether or not that means that the entire GPU cache must be flushed or if it can selectively flush lines ala the PS4.

Yeah thats what im wondering it was originally if it could write coherently and its obvious now that it can. The question lies on the details about how it achieves it now.

 

[adev]

Yeah thats what im wondering it was originally if it could write coherently and its obvious now that it can. The question lies on the details about how it achieves it now.

I wouldn't say that it can yet .

There's a single line in a thing which says it can but I haven't seen any implementation details to say how it achieves that and there hasn't been any mention of it on VGLeaks which looks entirely accurate.

We also can't 100% say that it can't because there's a single line which says it can .

 

[taisui]

Not to be a jerk but why are you so hostile, Beta hasn't shown any attitude towards you but for some reason you and another poster in another seem to take any comment questioning anything related to a certain platform quite personally. We all love tech and as such ask questions, tear down answers and should be able to do so without being defensive or overly aggressive.

It just annoys me when wrong information are being presented as facts repeatedly even after being corrected. I could've been nicer, but it's difficult me when the other person is not making sense.