Xbox One (Durango) Technical hardware investigation

Status
Not open for further replies.
what modifications did amd/ms do to the jaguar cpus aside from the 4mb of cache? it does get 30gb/s of bandwidth correct?

if someone could explain one thing to me. if the cpu gets 30gb/s of bandwidth that leaves less than 38gb/s for everything else including the dmas feeding data to the gpu. will that ever pose a problem?

the 4MB of cache isn't a modification its standard jaguar.
 
what modifications did amd/ms do to the jaguar cpus aside from the 4mb of cache? it does get 30gb/s of bandwidth correct?

if someone could explain one thing to me. if the cpu gets 30gb/s of bandwidth that leaves less than 38gb/s for everything else including the dmas feeding data to the gpu. will that ever pose a problem?

The GPU and DMEs have independent access to the DDR3 memory controller.

The CPU isn't going to be accessing data from main memory constantly. When it isn't accessing it, then the GPU/DMEs will get the full 68 GB/s. Likewise the GPU won't necessarily be accessing main memory constantly.

Regards,
SB
 
The GPU and DMEs have independent access to the DDR3 memory controller.

The CPU isn't going to be accessing data from main memory constantly. When it isn't accessing it, then the GPU/DMEs will get the full 68 GB/s. Likewise the GPU won't necessarily be accessing main memory constantly.

Regards,
SB

i see, thank you for explaining it. the system sounds very thought out. is there any downside with this approach compared to a more desktop gpu approach? as in do you think we might see esram + ddr4 on high end gpus in the future or would that be more expensive than just using gddr5/6/future standards? if i understand correctly the gpu sees the esram + ddr3 as one collective pool correct?
 
i see, thank you for explaining it. the system sounds very thought out. is there any downside with this approach compared to a more desktop gpu approach? as in do you think we might see esram + ddr4 on high end gpus in the future or would that be more expensive than just using gddr5/6/future standards? if i understand correctly the gpu sees the esram + ddr3 as one collective pool correct?

Downside would be extra code to write to get around less bandwidth of the DDR3.
Probably won't see esram on desktop gpu because PC devs aren't likely to target it specifically, though I wonder if a small esram can be used as hidden TRAM at the driver level with PRT being a standard now.
I believe GPU sees them as different pool, so devs need to write code to manage them.
 
Downside would be extra code to write to get around less bandwidth of the DDR3.
Probably won't see esram on desktop gpu because PC devs aren't likely to target it specifically, though I wonder if a small esram can be used as hidden TRAM at the driver level with PRT being a standard now.
I believe GPU sees them as different pool, so devs need to write code to manage them.

Hi taisui, I see you mentioned TRAM ...

I asked a question in the below linked thread regarding TRAM and its involvement in what I believe is a new Direct2D feature "Block Compression (BC1-3)" ... What is TRAM vs eSRAM .. are they the same thing?

http://beyond3d.com/showpost.php?p=1780832&postcount=60
 
Hi taisui, I see you mentioned TRAM ...

I asked a question in the below linked thread regarding TRAM and its involvement in what I believe is a new Direct2D feature "Block Compression (BC1-3)" ... What is TRAM vs eSRAM .. are they the same thing?

http://beyond3d.com/showpost.php?p=1780832&postcount=60

BC1-3 are not really new, they are pretty old in fact.

http://msdn.microsoft.com/en-us/library/windows/desktop/hh308955(v=vs.85).aspx

They were first introduced in DX9.1 and I believe have been a hardware feature of ATI cards for a really long time.
 
Hi taisui, I see you mentioned TRAM ...

I asked a question in the below linked thread regarding TRAM and its involvement in what I believe is a new Direct2D feature "Block Compression (BC1-3)" ... What is TRAM vs eSRAM .. are they the same thing?

http://beyond3d.com/showpost.php?p=1780832&postcount=60

It would seem that TRAM is just a low latency memory that to store the temporary texture, eSRAM is embedded SRAM, typically SRAM is faster than DRAM. From what I know, the PRT is like a HW implementation of the "mega texture" (clipmap), so it would make sense to have a fast access memory to stage the texture.

Block Compression is just a generic term for image compression algorithm, in this category the image is divided into 4x4 blocks and the group of pixels are compressed. Think of it like JPEG images with high comopression ratio, the "block" artifact. I believe that D1-D3 means is just diff compression algorithm for diff image types, and in the case of D2D, it seems that they recommend just use D1-D3 since all DX9 hardware would support it.

Hope this helps, I'm quite behind on DX technology so don't quote me on this :)
 
It would seem that TRAM is just a low latency memory that to store the temporary texture, eSRAM is embedded SRAM, typically SRAM is faster than DRAM. From what I know, the PRT is like a HW implementation of the "mega texture" (clipmap), so it would make sense to have a fast access memory to stage the texture.

Block Compression is just a generic term for image compression algorithm, in this category the image is divided into 4x4 blocks and the group of pixels are compressed. Think of it like JPEG images with high comopression ratio, the "block" artifact. I believe that D1-D3 means is just diff compression algorithm for diff image types, and in the case of D2D, it seems that they recommend just use D1-D3 since all DX9 hardware would support it.

Hope this helps, I'm quite behind on DX technology so don't quote me on this :)


Oh I understand how the DirectX side of things works (DDS BC1-3) :)

I was more questioning why its called TRAM... I'm assuming that on Xbox One its the eSRAM that behaves as the "Texture Memory" they speak of in the patents where these compressed DDS sits..

The whole TRAM name is throwing me off :)
 
BC1-3 are not really new, they are pretty old in fact.

http://msdn.microsoft.com/en-us/library/windows/desktop/hh308955(v=vs.85).aspx

They were first introduced in DX9.1 and I believe have been a hardware feature of ATI cards for a really long time.

Sorry for the confusion, I understand DDS BC1-3 is old tech, and I know that they only just introduced it to Direct2D in DirectX 11.2..

My question was more that it appears to me that in the patent these DDS BC1-3 compressed textures are stored in TRAM "Texture Memory", which im assuming on Xbox One is eSRAM in the Main Soc..

I'm really confused on why its being called TRAM :)
 
Sorry for the confusion, I understand DDS BC1-3 is old tech, and I know that they only just introduced it to Direct2D in DirectX 11.2..

My question was more that it appears to me that in the patent these DDS BC1-3 compressed textures are stored in TRAM "Texture Memory", which im assuming on Xbox One is eSRAM in the Main Soc..

I'm really confused on why its being called TRAM :)

TRAM would be the texture memory , most likely the read only texture buffers in the GPU or the normal memory at higher level.
 
Oh I understand how the DirectX side of things works (DDS BC1-3) :)

I was more questioning why its called TRAM... I'm assuming that on Xbox One its the eSRAM that behaves as the "Texture Memory" they speak of in the patents where these compressed DDS sits..

The whole TRAM name is throwing me off :)

Don't have X1 access so I'm just speculating:
The eSRAM on the X1 should be pretty flexible compare to the eDRAM which can only be used as frame buffer. I was thinking that it'll be utilized as the G buffer for deferred rendering, but as a "TRAM" it would seem sensible as well.
 
Don't have X1 access so I'm just speculating:
The eSRAM on the X1 should be pretty flexible compare to the eDRAM which can only be used as frame buffer. I was thinking that it'll be utilized as the G buffer for deferred rendering, but as a "TRAM" it would seem sensible as well.

It would appear to me that you would get more benefit by writing to the eSRAM then reading back using it as a intermediate buffer.

You need to remember that copying anything from DDR3 -> eSRAM is going to cost you 2x the bandwidth of just reading it straight from the DDR3 to the GPU you'd probably be better off reading the textures straight from the DDR3 in this case.
 
It would appear to me that you would get more benefit by writing to the eSRAM then reading back using it as a intermediate buffer.

You need to remember that copying anything from DDR3 -> eSRAM is going to cost you 2x the bandwidth of just reading it straight from the DDR3 to the GPU you'd probably be better off reading the textures straight from the DDR3 in this case.

Of course, I was thinking more along the lines of deferred rendering:
geometry pass, read from DDR3 (verts/indices), write to ESRAM
read G buffer from ESRAM, read texture from DDR3, write to ESRAM.

Again, I'm speculating that you can use the ESRAM as the tile pool, a 8MB pool should be able sample a 1G texture easily.
Say the worst case is streaming 32MB each frame at 60 fps, that gives ~2Gbps, since the eSRAM BW is not the bottleneck, it seems 2Gbps would still be easy for the DDR3 BW
 
Last edited by a moderator:
Here's the original PDF for the Hot Chips presentation (includes all the Kinect stuff too that wasn't really covered by the publications):

http://www.hotchips.org/wp-content/...1-epub/HC25.26.121-fixed- XB1 20130826gnn.pdf

It wouldn't surprise me if the Time of Flight sensor ended up in a lot of other products or was licensed out to third parties.

I'd love to sit and watch a detailed presentation on how the new Kinect sensor works. I'm not particularly knowledgeable about the physics involved, but it's incredibly interesting.
 
Here's the original PDF for the Hot Chips presentation (includes all the Kinect stuff too that wasn't really covered by the publications):

http://www.hotchips.org/wp-content/...1-epub/HC25.26.121-fixed- XB1 20130826gnn.pdf

It wouldn't surprise me if the Time of Flight sensor ended up in a lot of other products or was licensed out to third parties.

I would post that in the Kinect Tech thread. Neat stuff(read: fascinating but didn't understand a lick of it LOL).

Tommy McClain
 
The meaning of total exposure time seems obvious, but what is "time to software"? That would be about 1 frame of latency at 30fps, if software is a game.
 
so MS is under so kind of NDA? that's why they couldn't speak all about the xbox one silicon at Hotchips?
Yes, allegedly so. I heard about that from quite a few people, and it has something to do with AMD waiting to unveil some new technology, whatever it is. That's how I understood it at least.

Besides that, I think that the GPU is not so powerful so the new Xbox One is going to be like the Gamecube of the 128 bits generation -one of my favourite console generations-, performance/capabilities wise. The GC wasn't the most powerful console of its generation but it was uniquely capable in some particular scenarios.
 
Status
Not open for further replies.
Back
Top