tonymarraffa said:
Tabular said: So... does 50% less GPU performance equal being aligned? |
Where was I when the Specs were confirmed?
|
Interesting you should mention that. Another poster actually posted something very interesting that about the leaked specs and if this data is correct puts a huge change to how that data was interpredit. I am going to re-post what he posted changing a few things as his English is not so great. I will say take this with a grain of salt but he do seem to back up his information
Ok so here is the beak down
This is the VGLeaks diagram for Durango
http://www.vgleaks.com/dura...
Now under the Compute paragraph here is what it states
"Each of the 12 Durango SCs has its own L1 cache,LSM(Local Shared Memory), and scheduler, and four simd units" .
Ok we are talking about FOUR simds - I will not go into detail on what they do just know they perform operations or work to get you the nice graphics, physics, AI etc.
this is the PS4 leaked document which turned out to be true when Sony announce the PS4
http://www.vgleaks.com/worl...
go to the update part and you will see this line
"Each CU contains dedicated:
- ALU (32 64-bit operations per cycle)"
So what this is saying is that the PS4 GPU can execute 32 64bit threads per CPU cycle.
Now go back to the Durango leaks and you will read:
" A SIMD executes a vector instruction on 64 threads at once in lockstep "
So within the Durango leak its stating that its CU unit will execute 64 threads at once. Within the ps4 we have 32 threads.
Within each SC (this is the name MS gives to their compute cores some call it shader cores some called it shared cores) there are 4 simd in every SC
12 Shader cores each on 64 threads = 768 operations per CPU cycle on durango against
14 CU each 32 threads = 448 If you add the other 4CU it's = 557
If we are looking at comparative hardware out now and coming up there are some very interesting parallels between Durango and AMD 7970 GPU.
again on the 7970 we have 32 CU divided in groups and amd smartly call them 8 CU ARRAY...same thing happen in GTX 680 they call 8SMX a group of 32 CU.
For comparison here is a link to the 7970 diagram
http://i.imgur.com/g61Dtdy....
(amd 7970)
compared to the Nextbox (rumored) gpu
http://i.imgur.com/sqkJUmO....
So here is the punch line. Within the VGleak.com document, they call the SC one CU and this is where they are getting the difference in TFlops between the PS4 and Durango. From the document though, it appears that the SC within Durango is actually an array of 4 CU units within one SC. The reason I say this because one CU has 4 SIMD that do 32 threads per clock, while the Durango has 4 SIMD that do 64 threads per clock
If this is correct than instead of 12 CU cores averaging out to 1.2 TFlops, the Durango actually is 48CU which averages around 3.8TFlops