Xbox One (Durango) Technical hardware investigation

Again, what is your source? Don't want specifics, just a general idea of the source's relationship to the tech in question. And iirc Mejdrich's RT hardware patents are assigned to IBM (though published while he and his team were working at MS).

 

The thing is, they could have added so many more CU's for that area.

Why use ESRAM to gain some sort of performance bump, when maybe you could have added 8 or 10 CU's in the same area??? You're jumping through hoops for no reason.

Of course it's all a 3 way cost trade. you must account for DDR3 over GDDR cost savings, etc. And I imagine there are things we're not privy too here (for example, Sony recently basically said they nixed Cell based BC in PS4, not because of Cell so much, but because there a ton of other minor components surrounding Cell that can no longer be sourced. But we forumers would never guess something like that) I trust MS has run the numbers, obviously.

Also, the dual APU rumors rear again, fresh off the neogaf presses, salt trucks n all:

http://www.neogaf.com/forum/showthread.php?t=533769

 

It seems to back up that Kotaku stuff regarding the stripe patterns all over the controllers and the console. And this was before the Kotaku article, but it sounds way too unbelievable to be true.

 

Someone with firsthand knowledge of the system and a technical background, more than capable of understanding the things we're talking about.
Not a journalist like Aegies.

Dual APU rumours are completely fake, even Proelite has called it out.

 

Yea, I'm not buying it either.

 

I still like the post in the semi accurate forums that the move engines are actually PowerPC cores. /right...

Have we already thought it was a dual apu unit due to the arm being on board, and not actually dual duct taped super computers?

 

True. But Proelite has also been known to post the opposite of the truth.

Confirming this in a public forum would not be a good idea.

 

Ah, but he was actually a proponent of the dual APU rumours at one point - and even he's changed his mind

And in any case, I know myself that such rumours are bunk (as common sense would indicate).

 

32MB of eSRAM seems to be part of their strategy since Yukon, so Microsoft's engineers must have seen value in it over extra ALUs/CUs. When it became clear DDR4 wasn't going to be an option I suspect MS locked themselves to embedded memory to help alleviate DDR3 bandwidth limitations. They might have considered GDDR5, but didn't want to place bets on if/when it hit 4Gbit densities.

I don't think that would be efficient. Bandwidth to main memory is already limited with one APU, splitting it between 2 would probably starve them despite the eSRAM. Then there is coherency to think about, I wonder what a cache miss would cost you.

If MS wasn't happy with the CPU and/or GPU performance, they would be better off with a discrete design over doubling up the APUs.

 

Unfortunately for your theory, the actual VGLeaks data specifically rules it out.

 

Which is strange, the GCN articles say double precision is 1/2 with the ability to configure it for 1/4, 1/8, etc. So MS got the low end?

http://www.anandtech.com/show/4455/amds-graphics-core-next-preview-amd-architects-for-compute/6

 

What are you talking about? The only mobile TB design readily available that is commonly described as TBDR is PowerVR. Adreno isn't an anagram of Radeon by happen stance. AMD might not sell a TB based design anymore. It doesn't mean AMD suddenly forgot how to design one. If the tech held no promise, I doubt Qualcomm would have snap up the design and used it as a basis for the gpu in SnapDragon.

I don't see how tile rendering on Durango can be so easily dismissed as it was available in the 360. I am not saying Durango has to be a tile based design. But AMD building a gpu in the console space capable of tiling isn't something new.

 

It's also now confirmed to be a lie. The dual APU rumors, I mean.

 

The point is not that you can't implement a tile based renderer on Durango, I'm sure you could - it might be better suited to Durango than PS4 given the ESRAM and move engines.

The point is that Durango has not been explicitly designed for tile based rendering.
I.e. MS did not set out to build a machine that would particularly excel at tile based rendering.

....

Can someone also confirm what the difference in latencies between 1T and 6T SRAM are? Is it just 28 vs 32 cycles as was posted earlier?

 

I see your point. The 360 isn't strictly a tile based design even though it's commonly described as a tile based design. The issue i have with Durango is that its designed with eSRAM and a rather lacking amount of bandwidth, which seems to have handicap Durango in a way the 360 never was when talking IMR.

 

Perhaps SRAM and GDDR were only options for Microsoft, or best of them.

 

If I compare it with AMD discrete gpus it's lack of bandwidth is very apparent. It's provides more bandwidth than standard AMD apu parts but those are no where close to high performance parts.

Part of me wants to believe that tiling is very much a part of Durango because a low bandwidth environment is a reality in a pretty big market. So it's kind of hard for me to believe that two companies that don't have a big presence in that market found a robust alternative design to deal with low bandwidth.

That same part of me sees the Durango as a design that will readily extend downward like going from high PCs to low end laptops being service by AMD gpus. Cut away some CUs/CPUs and reduce the SRAM/clocks and you might have something that fits into tab or phone and essentially be Durango based. Basically a common arch that fits well with the type of products sold by the edd div at MS. Being tile capable makes that a lot easier.

 

Not really, the 1.2 teraflop 7770 ships with 72 GB/s. The 1.2 alleged teraflop Durango has 68 GB/s.

That's of course completely ignoring the ESRAM.

Granted there's a CPU to deal with as well.

 

Is it really a serious handicap? Both need to provide some bandwidth to their 8 core CPUs and Durango's GPU is quite conservative relative to its available bandwidth and appears to also have access to quite a bit more memory bandwidth than either a 7770GHZ Edition or Radeon 7790.

Also, the ESRAM bandwidth belongs only to the GPU and its memory clients. Even if you only look at the 68GB/s from the 8GB of DDR3, you still have 102GB/s helping because of the ESRAM. Taking the CPU into account and the vgleaks Durango memory system example, around 26GB/s of the DDR3 would go to the CPU while the other 42GB/s of it would go to the GPU before considering ESRAM. It seems like there's quite a bit there for this kind of GPU.

 

And why would you do that ?

Cheers