quoting myself but im guessing memory controller
the one thing i'm not sure about is if this core is compute capable. has "compute commands" on the side, but these are still listed as shader cores
-
Hey, guest user. Hope you're enjoying NeoGAF! Have you considered registering for an account? Come join us and add your take to the daily discourse.
VGLeaks Durango specs: x64 8-core CPU @1.6GHz, 8GB DDR3 + 32MB ESRAM, 50GB 6x BD...
- Thread starter Shikoro
- Start date
quoting myself but im guessing memory controller
the one thing i'm not sure about is if this core is compute capable. has "compute commands" on the side, but these are still listed as shader cores
Can Crusher
Banned
Can Crusher
Banned
phosphor112
Banned
Probably, because it doesn't make sense..
"O"
"C
B"
Wtf does that even mean?
What are the move engines next to the video encoders and decoders?
They haven't published the article yet, there is nothing else to see
bigtroyjon
Member
Have they released a diagram of PS4?
it kinda does look like GCN, actually. sorta.
kinda screwed around with GCN. More texturing units as opposed to GCN? At least that's what I'm sorta seeing.
Like... they texture off the SIMDs? instead of through the compute unit as a whole?
*- note my knowledge is limited
bigtroyjon
Member
thanks, I remember that one, didn't know if there had been anymore.
phosphor112
Banned
Looking at it, the "SC" is similar to the GCN.
Both have 16kb L1
Both have 64kb LS
Both have 4 SIMD
Scheduler...
Looks exactly like GCN.
EDIT: It seems the GCN has 8 SIMD units, the durango example only has 4.
Durango diagram also doesn't list Texture units or fill rate units.
phosphor112
Banned
except they have 4 texture filter units per SIMD on this GPU?
View-Space Positioning = texture filter?
edit:^^^^ isn't it only 4 SIMDs?
You're right. That diagram I posted above has "2" layered, making it look like 8.
maybe.. I know nothing about GCN. Never really paid attention to AMD's new line of cards...
phosphor112
Banned
So which one is right? They are both slightly different.
Top has 4 vector (SIMD) units.
Bottom has 8.
Top has 16 texture fetch/store units.
Bottom has 8.
phosphor112
Banned
They are extremely vague...
phosphor112
Banned
Four Os per VSP. Four VSPs per SIMD. Four SIMDs per SC. Twelve SCs in total.
That's 768 Os. That's right: Real-time blast processing. None of that pre-computed shit like back in the Genesis days.
So which one is right? They are both slightly different.
Top has 4 vector (SIMD) units.
Bottom has 8.
Top has 16 texture fetch/store units.
Bottom has 8.
i dunno about the texture units
but there's def only 4 SIMD units
phosphor112
Banned
:lol
Durango king of Spaghetti O's confirmed.
i dunno about the texture units
but there's def only 4 SIMD units
Sigh. I wish I knew.
Proelite
Member
So which one is right? They are both slightly different.
Top has 4 vector (SIMD) units.
Bottom has 8.
Top has 16 texture fetch/store units.
Bottom has 8.
Bottom has 4 SIMD units, they're illustrated in that way to show that they're composed of 16 ALUs each.
phosphor112
Banned
But it says it's 16 wide. Why would they show it like that? Also the Texture fetch/store units are displayed the same way. I'm confused.
EDIT:... so we assume they are the same figures... it's just a GCN then.. right? They wouldn't be any different.
phosphor112
Banned
Mk. so 16 in each CU?
So GCN confirmed pretty much. I don't really see anything difference between a "SC" and a "GCN" CU.
Brad Grenz
Member
So which one is right? They are both slightly different.
Top has 4 vector (SIMD) units.
Bottom has 8.
Top has 16 texture fetch/store units.
Bottom has 8.
I think the layered squares are just supposed to abstractly illustrate the fact that there are multiples of those units. 16 Vector SIMDs, or 4 Texture Fetch units. Don't read too much into it.
Looks like GCN, though. L1 cache is too big for VLIW4 or 5.
phosphor112
Banned
Definitely, but that's good.
Not GCN2 though as some people thought. I think it might be just different wording in the documentation.
Let's wait and see what they post. Apparently MS don't refer to durango stuff as CU but shader cores.
Brad Grenz
Member
Most interesting part: Distinct Compute and Graphics sheduling hardware. Not unlike Orbis, perhaps? Could be half and half, could be 4/8, could be fully dynamic, but that may solve our Orbis question as well...
EDIT: GCN already has Asynchronous Compute Engines that handle the GPGPU stuff, so even that is standard.
Ahh so the colour block is finally back! Still wish I knew what it, and the DME's did lol.
Most interesting part: Distinct Compute and Graphics sheduling hardware. Not unlike Orbis, perhaps? Could be half and half, could be 4/8, could be fully dynamic, but that may solve our Orbis question as well...
Oh I see. So you are saying that if this is similar to Orbos then maybe the those 4 dedicated compute CUs don't have a graphics command hardware but just the compute command.
phosphor112
Banned
Depth block and color block sounds like units to help with kinect processing.
Something has to process the infrared data, and something has to processes the RGB data on the connect. To keep lag at a minimum, this might be it.
Or maybe not. But I'm sure they wouldn't want 200+ms latency for kinect responses like this gen. And assuming they are using cameras with resolutions and frame rates that aren't ass, these might be it.
Brad Grenz
Member
Probably not, actually. It was pointed out to me GCN already has dedicated Compute Command units called ACEs that take care of that. It could be segmented specifically like Orbis in some fashion, but not neccessarily. You can see them in this image:
Depth and Color buffers are a standard part of a GPU's ROPs, but if we assume each block shown represents 4 Render Out Processor it looks like Durango may have only 16 compared to the rumored 32 in Orbis. That would mean half the fill rate.
phosphor112
Banned
So you think that's what they are? VG leaks needs to explain their stuff soon.
If it's half the ROPS.. that... I don't think that's true. =S
Brad Grenz
Member
Yes, the depth buffer and the color buffer would be part of the same render back end. In GCN terms a RBE has 4 ROPs. So you take the four depicted in the diagram and multiply that by 4 and you'd get 16.
Frankly it's a bit weird for MS documentation to break them out individually like that, but it might explain aegies' confusion about color and depth compression being some kind of "special sauce", but the fact is that's a pretty old idea. the old Radeon 9700 had like 6:1 lossless color compression and 32:1 Z-compression in hardware. GCN rates are better than that, but it's not new or remarkable, and it's nothing Orbis won't have as well.
bigtroyjon
Member
Kinect data is processed in the Kinect itself, that's what the primesense chip does.
yellow submarine
Member
I await the reaction from everyone riding the hype train when they find out that Kinect Sports XVII for Xbox Durango(tm) still renders in 400p native/30fps (but can now figure out if you're showing a middle finger during your victory celebration instead of just a stubby hand).
The depth/color block it is IMHO. It gives you free 4xMSAA + FP16.
Brad Grenz
Member
That's literally impossible.
