NGGP: NextGen Garbage Pile (aka: No one reads the topics or stays on topic) *spawn*
- Thread starter Barso
- Start date
- Status
(((interference)))
Veteran
It seems the 4 GPGPU compute units are just 4 normal CU units with additional scalar ALUs that make them better suited for GPGPU, but they can still work as normal compute units for graphics with no penalty.
Yep my bad, That was another feature that Xenos listed from R6xx - the setup rate was one triangle per clock rather than half like the R5xx and G7x series.
I've edited my post.
anexanhume
Veteran
Assuming the concept isn't preposterous, 12 CU's would be a lot for a "casual" box.
Where did you got those info?
The part about the HD 7750 is incorrect (so could be the rest I haven't check but remembered the hd 7750 being pretty impressive wrt power consumption so I knew your statement were incorrect).
It is a 45 Watts card.
So I would bet your estimate is wrong though 120Watts (peak figure) for the whole system doesn't sound too crazy (with RAM chips, BRD drive, HDD, etc.).
Last edited by a moderator:
anexanhume
Veteran
And how does this render his argument invalid? They are dedicating HALF (15ms) of GPU budget for simple blur. If thats not wasting resources on very unoptimized technique then I don't know what is.
(((interference)))
Veteran
Nope, they're for Durango alright.
Unless you can find some other fantastically corroborated bunch of rumoured specs out there for a more powerful Xbox console.
Hey Chefo, we actually don't KNOW anything for certain.
But taking the VG leak at face value we know it is not PVR but GCN. Just look at the cache design and the configuration.
Btw, the real wild card is... price.
All bets are off if MS can fit this all into a $249 box.
The die space allocated to the CPU and GPU in Durango is a lot less than the Xbox 360 (which from a HW perspective was not overly complex or expensive once supply channels flowed). Xenos was 2 chips (about 182mm^2 and 80mm^2) at 90nm at launch and Xenon was about 176mm^2 (each CPU was about 30-40mm^2 and the remainder of die space was the 1MB of shared L2 cache and bus). That is about 438mm^2, give or take, of total die space.
Cape Verde (7770) is a 123mm^2 GPU at 28nm. Durango is fairly close to Cape Verde units + 32MB of ESRAM (which sounds like a variant of 1T-SRAM in which case it will be in the 20-50mm^2 range?). So the total GPU budget is shrinking about about 100mm^2.
The CPU side is similar. Jaguar cores are about 3mm^2 each and the L2 cache is 3-5mm^2 (give or take) so the CPU is going to fall under 60mm^2. Again, about 100mm^2 less than the Xbox 360 budget for the CPU.
Durango is looking to cut out 200mm^2 of silicon footprint--about 40% reduction in budgets.
The main memory this go around is also going on the cheap side; the Xbox 360 had for the time relatively fast GDDR3 memory; Durango is going with bog standard DDR3 (8GB of DDR3 can be scored easily for $50 at retail).
Depending on the cost of Kinect 2, packed in software, any HW BC, and the cost of the controller it looks like MS will be in a much better position on cost than last go around.
And then there is the crazy stupid Cell Phone pricing agreement plan--expect to see $99 Xbox 3's with 2-3 year contracts of XBL for $15/mo.
Which will position MS where they want to be: a small, affordable console that is a full blown set top box/media all in one experience.
Lucid_Dreamer
Veteran
I thought the slide was quite clear. The whole DoF process took 15ms on the GPU only. When using a SPU WITH the GPU, the entire process is 5ms (1.5 - 2ms on SPU + 3ms on GPU).
IDK, that's what lherre posted on NG:
http://www.neogaf.com/forum/showpost.php?p=47291466&postcount=678
People should beware of SuperDae. He tends to give out misleading information / string people along.
Although I believe VGLeaks uses info from the Durango Summit held in Feburary 2012.
The most critical Cell work is to overcome RSX's vertex setup bottleneck. e.g., triangle culling reduces its vertex workload. Then they moved on to augment other vertex work (AI-driven animation, skinning). Once they have done that, they tackled high quality post processing. After that, they enhanced dynamic lighting. Ideally, the SPU jobs are done while RSX is busy with something else.
Most of the time, the objective is not to use brute force. The GPU has more cores. The challenge is to work smarter (because CPU is more flexible, and free) and also differentiate in quality (because CPU + GPU can afford to spend "more" compute time on the same work).
For that depth of field example, I think they picked a compute heavy case (Slide 9). To lighten RSX's load, they let the SPUs categorize the frame(s) into different "classes" of small tiles first. When RSX gets to the frame, it only needs to process a fraction of the data. Depending on the scene, some tiles may not require GPU work at all; some require less computation; while others require full attention.
To save time, they may also do multiple categorizations per SPU pass (for several follow-up GPU tasks later).
In such scenario, the performance gain would depend on the complexity of the scenes and amount of work saved.
Most of the time, the objective is not to use brute force. The GPU has more cores. The challenge is to work smarter (because CPU is more flexible, and free) and also differentiate in quality (because CPU + GPU can afford to spend "more" compute time on the same work).
For that depth of field example, I think they picked a compute heavy case (Slide 9). To lighten RSX's load, they let the SPUs categorize the frame(s) into different "classes" of small tiles first. When RSX gets to the frame, it only needs to process a fraction of the data. Depending on the scene, some tiles may not require GPU work at all; some require less computation; while others require full attention.
To save time, they may also do multiple categorizations per SPU pass (for several follow-up GPU tasks later).
In such scenario, the performance gain would depend on the complexity of the scenes and amount of work saved.
I was working out the power usage earlier today by looking at the 7770, Jaguar cores, memory, DVD, HDD, etc and came in the range of 120W. Maybe they can pull off 100W of actual load with a 130W or so power brick but I would expect more in the 120W range and a 150W brick.
Anyone expect in the 60W range needs to go look at the WiiU and contrast these Durango specs.
The first time anyone heard of him he was saying the Durango had an nVidia GPU. He is not reliable, you could interpret his tweet in a number of ways (if you can even parse it) and we know very little about his personal biases.
(((interference)))
Veteran
Ha, let's not forget that DaE was the guy who thought there was an Intel CPU and nVidia GPU in the kits.
He's also a big tease and a month ago was claiming to reveal the Durango specs in some Google Hangout that never eventuated.
The vgleaks specs are unlikely to vary by any large degree from what's in the final Xbox - regardless of what the fanboys may hope for.
- Status
Similar threads
