Xbox One (Durango) Technical hardware investigation

[Xenio]

doesn't excuse rudeness

if you want to talk about topic that's ok, if not please quit or use messages.

 

[dobwal]

Has anybody come across this paper. I'm not sure about the relevance of the fpga component (GCN may make the configuration relevant.)

CoRAM: An In-Fabric Memory Architecture
for FPGA-based Computing

http://research.microsoft.com/en-us/people/erchung/coram.pdf

We propose a new FPGA memory architecture called Connected RAM (CoRAM) to serve as a portable bridge between the distributed computation kernels and the external memory interfaces. In addition to improving performance and efficiency, the CoRAM architecture provides a virtualized memory environment as seen by the hardware kernels to simplify development and to improve an application’s portability and scalability.

The CoRAM architecture assumes that reconfigurable logic resources will exist either as stand-alone FPGAs on a multiprocessor memory bus or integrated as fabric into a single-chip heterogeneous multicore. Regardless of the configuration, it is assumed that memory interfaces for loading from and storing to a linear address space will exist at the boundaries of the reconfigurable logic (referred to as edge memory in this paper). These implementation-specific edge memory interfaces could be realized as dedicated memory/bus controllers or even coherent cache interfaces...

Much like current FPGA memory architectures, CoRAMs preserve the desirable characteristics of conventional fabric-embedded SRAM [16]—they present a simple, wire-level SRAM interface to the user logic with local address spaces and deterministic access times (see Figure 2b), are spatially distributed, and provide high aggregate on-chip bandwidth. They can be further composed and configured with flexible aspect ratios. CoRAMs, however, deviate drastically from conventional embedded SRAMs in the sense that the data contents of individual CoRAMs are actively managed by finite state machines called “control threads”.

Control threads form a distributed collection of logical, asynchronous finite state machines for mediating data transfers between CoRAMs and the edge memory interface. Each CoRAM is managed by at most a single control thread, although a control thread could manage multiple CoRAMs. Under the CoRAM architectural paradigm, user logic relies solely on control threads to access external main memory over the course of computation. Control threads and user logic are peer entities that interact over predefined, two-way asynchronous channels.

The most closely related work to CoRAM is LEAP [2], which shares the objective of providing a standard abstraction. LEAP abstracts away the details of memory management by exporting a set of timing-insensitive, request-response interfaces to local client address spaces. Underlying details such as multi-level caching and data movement are hidden from the user. The CoRAM architecture differs from LEAP by allowing explicit user control over the lower-level details of data movement between global memory interfaces and the on-die embedded SRAMs; the CoRAM architecture could itself be used to support the data movement operations required in a LEAP abstraction.

 

[pjbliverpool]

he says "There's nothing that will magically make a GPU magically perform 50% more calculations a second.."

not console GPU, and I don't know why a console gpu can't have an high clock

It can have whatever clock it's designed to have. If you "overclock" it beyond that design then it's going to be unreliable.

the GPU on SOC in my Galaxy nexus, from 1.2 GHz to 1.9 GHz
+58%

That's a Galaxy Nexus designed for extremely low power. Not a GCN architecture. Show me a GCN architecture than can be overclocked by 50% on standard cooling.

overclocking CU units it's easier because you have not to OC all the logic, soc, board, only math units.

If you overclock the CU's there's not much left anyway. And as far as I know the ROPS and CU's are in the same clock domain so you pretty much have to overclock the entire GPU.

or you can obtain the same results overclocking by 30% and using more cache on CU (check) and some low latency memory (check)

Have I missed something? I wasn't aware that Durango had more GPU cache than Orbis.

 

[eastmen]

I think there is a huge diffrenece between secret sauce and ms throeing out the durango design last minute and what myself and some others have been saying.

Leaks from early in 2012 on soecs could simpky have been based on yields.

A 8 core 1.6ghz jaguar cpu with 12 cus running at 800mhz with 32megs of esram and 8 gigs of ddr3 ram could easily become a 8 core 1 8ghz jaguar cpu with 17 or 18 cu gou and 64megs of esram with faster ddr3 if that is what the chips originaly were but with hardware disabled do to yields.

In those 9 plus months ms can have rolled out new spins fixing priblems or yields from test runs coukd have come back the question is if hardware is disabled for yields and if yields have improved enough or ms is willing to move ahead with worse yields to enable that disabled hardware.

Who knows at this point. None of us that I can tell .

 

[pjbliverpool]

A 8 core 1.6ghz jaguar cpu with 12 cus running at 800mhz with 32megs of esram and 8 gigs of ddr3 ram could easily become a 8 core 1 8ghz jaguar cpu with 17 or 18 cu gou and 64megs of esram with faster ddr3 if that is what the chips originaly were but with hardware disabled do to yields.

I'd be pretty amazed if Microsoft had commissioned an 18CU chip with 64MB esram and then planned to disable 6CU's and half the esram for redundancy

 

[eastmen]

I'd be pretty amazed if Microsoft had commissioned an 18CU chip with 64MB esram and then planned to disable 6CU's and half the esram for redundancy

I was just throwinng out numbers but I remember semiaccurate was talking about the ms console chip being huge and them having problems producing it last fall. So there can be redunance there but who knows how much.

 

[torbor]

Nothing to really to worry about for me, but that "slightly less powerful " is now turned into "the end of the world" by MS fanboys.

50 percent more flops(while using Gpus from the same vendor) according to the leaks is significantly more powerfull, not only "slighty".

So what is true: The leaks or that Orbis is slighty more powerfull?

 

[Averagejoe]

That phrasing (secret sauce, wizard jizz) was invented by ppl with an agenda in order to mock the idea that hardware accelerators existed and were designed to improve how adequately Durango could perform in real world applications. They were clearly wrong, as there really does seem to be extra kit that sounds like it could help Durango punch above its weight. How much so is open for speculation and debate.

It is completely reasonable and fair for onlookers to cry foul when everyone else wants to try drawing comparisons between stuff like flops and bandwidth on an apples to apples basis. It sounds like the Durango setup works to reduce the number of operations you'd need to do in order to get the same end result in the first place and the setup of the DME's and eSRAM sound like they do a lot to mix up the bandwidth comparisons.



I don't see anyone acting like you describe here. It's the opposite in fact. I see tons of ppl all over the internet asserting that the PS4 has a gigantic advantage in terms of power across the board pretty much. These ppl look at GPU specs only, ignore the accelerators in Durango (i.e. dismiss them completely out of hand as 'lol! secret sauce durrrrr'), ignore bottlenecks in the PS4 design (ex: you can't feed 32 ROPS with 176GB/s), ignore dev comments on the subject saying the two are MUCH closer than presumed, and then tell us how Kinect ruined Durango's GPU somehow. Ppl pushing ideological pseudo-dogma and/or memes as arguments while cherry picking information and dismissing reasonable tech when it is easy to mock with a magic phrase like that don't deserve to be taken seriously.



I think you may be confusing ppl openly mocking the concept of 'secret sauce' all over the internet with ppl actually expecting something radically new to show up. The eSRAM, DME's, display planes and the possibility of being designed to leverage virtual assets in a fundamental way is the 'secret sauce' in the sense that those items work together to make highly efficient use of the architecture as a whole to help it punch above its weight.



Yup. The fact ppl are eager to say PS4 is dramatically more powerful based on specs they got from a rumor telling them devs say both are roughly on par should tell you something about how these folks aim to cherry pick the information they allow to influence their worldviews here. Every actual insider and/or dev that I have seen to date commenting with vague hints has suggested the two are about the same.

I do and allot many here are actually supporting the secret sauce theory,and every time one is shut down another comes from no where.

Here was that the whole blitter argument was born,and was fast embrace in many other places on the internet because if it comes from Beyond3D it most be true,i can tell you this because i post on other forums and many people take post from here like god spell.

Is hard to know which hardware will be more powerful,Orbis appear to have better hardware and faster ram as well as a most straight forward design.

You say that Orbis can't feed 32ROP's at 176/GBs but what about feeding 24 or 26,sure it can't feed them all but that doesn't mean it can't feed more than 16,which most people think is perfect for Durango 12CU,have you look at it from that point.?

Orbis can have a horde of bottle necks,so can Durango,and here most of what i have see try always to paint a more optimistic route for Durango than for orbis.

Were ever this be more efficient without actually know,having no problems with latency while Orbis do again without any proof,what about the whole 68/GBs+104/GBs vs Orbis more straight forward 176/GBs that has been endlessly try to been pass as equal and in some cases even better.?

Most developers will always be polite even if one hardware is at disadvantage not saying this is the case but that how it is,most developer will refuse to answer that and tell you both are about on par.

And for them rightly so i remember the barrage of games on xbox that were just PS2 games with better aliasing when the xbox was a more capable machine,so even if we get a difference in this case in favor of Orbis it probably will not show on multiplatform games.

 

[warb]

50 percent more flops(while using Gpus from the same vendor) according to the leaks is significantly more powerfull, not only "slighty".

So what is true: The leaks or that Orbis is slighty more powerfull?

I think the leaks have it slightly more powerful, the GPU is not the only component. We'll perhaps see slightly better image quality with orbis/slightly more aggressive dynamic resolution on Durango.

 

[Averagejoe]

do you know "overclock"?

Do you know Over heat.?

 

[Shifty Geezer]

if he's not the Papa or amd/nvidia chief, he can be wrong as all the others.

'Overclocking' is not an option, otherwise nVidia would make available at 1.5 GHz instead of 1 GHz. 'Overclocking' means clocking a part above what it's designed to cope with, which adds enormous thermal stress to the component and will burn it out - heat increases exponentially with power, so a 50% overclock will generate insanely more heat. There's also no collection of transistors arranged in a trapezoid or with memory knick-knacking or bandwidth de-co-pollutinating that can add 50% performance. People who have worked closely on GPU architecture for 2 decades and looked for every opportunity to get better performance haven't found anything that can achieve that. I'm not sure even Unified Shaders achieved that level of performance improvement.

 

[astrograd]

MS are not doing this. The same way Sony needs to avoid $599 this gen, MS needs to avoid 'unreliable'.

To be fair, that was due to lead free solder, no? I don't think actual engineers would need to worry about that happening again unless they are incredibly incompetent in the first place. That's not to say it is realistic to think they'd do this and want to add more cooling to the setup though.

Well that how the Internet works, look at the early days of this gen.
But I think there is a difference this time around I would say that Sony 'fans 'are happy and are not in 'war" mode.
On the other hand the noise surrounding Durango either the deception at the specs or the other way around minimizing the difference between the system (though based on possibly incomplete, inaccurate information, but what else do we have?) is coming from MSFT "fans".

Not sure what forums you read, but the 3 I read (B3d, TXB, GAF) have all shown ppl ditching direct discussion and pushing back against efforts to pool together ALL reliable info. I also see some fringe MS fans trying to bring unreliable info into the discussion where it doesn't belong, but that is extremely rare and not the sort of inappropriate dismissal of ideas I am referring to. I think Sony fans are happy since hearing 8GB the other day. But the mentality I am noting established itself long before that. It was firmly in place the minute we heard info about both GPU's (but prior to hearing details about eSRAM, display planes, DME's).

Also, where is this deception you are referring to? I wouldn't say those of us wondering out loud if Durango might be built to leverage a tile based deferred renderer for handling virtualized assets in unqiue ways would be doing so deceptively.

I fully intend to pick up the new xbox on day 1 pretty much and have no intention of ever getting a PS4 so I certainly have no agenda against the xbox but I still disagree with what you're saying here.

You post doesn't read like you disagree with my comment at all. More like you are willing to discuss the 'ok, but how much and does it really matter?'. That's not the same as just dismissing the extra hardware out of hand and ignoring it while you mock those wishing to discuss it as 'lol! secret sauce...durrrrr'. THOSE ppl clearly have agendas and don't want to talk about details that they are concerned may threaten their view of Durango.

I don't believe it's been firmly established at all that the esram has any significant benefit for graphics rendering. The developer comments on here have ranged from "very little if any" to "some but probably not a huge amount" (paraphrasing of course). It sounds like it will be beneficial in some measure for compute workloads though which may also translate to gaming performance in the console world.

If nothing else it helps bandwidth and thus the amount of data available to a scene. Also, having 2 display planes for application instead of just 1 is likewise helpful for the end result of the rendering looking nice, even if we ignore what could potentially be done in extending the HUD plane to include some actual foreground objects, etc.

I still thing the primary driver behind the use of esram is to allow the console to contain 8GB of main RAM cheaply. And thus the DME's are there to minimize as much as possible the disadvantages of having your graphics memory split into two separate pools as opposed to the more efficient unified pool of the PS4.

You are casually assuming a particular perspective here. That may very well be completely reasonable and justifiable, but I want to take the opportunity to note that it's there. I think making assumptions about WHY this kit is there colors our position as to conceptualizing what it can be expected to do. Perhaps examining the arrangement of the pieces can help ppl get a better idea as to what it could be used for instead of relying on what they assume is the design agenda at the outset?

For the sake of argument I assume something different. I assume that they used 8GB because devs demanded it for a while now and they went with DDR3 to complement a customized CPU that is reportedly twice as powerful as we had assumed until recently. Perhaps they wanted to try building an architecture designed to leverage virtualized assets in a unique way? That would save on some bandwidth and processing on the GPU end iirc. Maybe the 2 application side display planes are there to help devs manage fill rates to avoid losing noticeable detail in foregrounds/HUDs? Perhaps they wanted to move some of the types of computations that are moving to the GPU in recent yrs back to the CPU for some reason? Hence the beefed up CPU and DDR3 instead of a GDDR variant?

Either way, I think it may be plausible that leveraging tile-based virtualized assets uniquely via hardware and pushing some GPU operations back to the CPU side could potentially result in a significantly lessened load on the GPU in games. I've no doubt you know more on the tech side of PC hardware than I do though, but I also know others that are well informed have made similar conjectures in the past.

 

[Shifty Geezer]

To be fair, that was due to lead free solder, no?

That's why 360 failed. expletive is (rightly) saying a GPU overclocked 50% will generate far more heat than it can cope with and fail, unless it had the most extreme and expensive cooling system, which'd cost more than just using a larger, faster GPU.

 

[fehu]

I was just throwinng out numbers but I remember semiaccurate was talking about the ms console chip being huge and them having problems producing it last fall. So there can be redunance there but who knows how much.

Has been hypotized that at some point info got so confused that have been attributed to the wrong console, so maybe was the ps4 to be having problems.

 

[astrograd]

I do and allot many here are actually supporting the secret sauce theory,and every time one is shut down another comes from no where.

Right now I think some ppl are inappropriately assuming Durango is plausibly going to get a spec change since PS4 got one. I can agree with being highly skeptical of that line of thinking. But the 'secret sauce' concept was always couched in the extra kit surrounding the standalone GPU in Durango (which we have since discovered are the eSRAM, display planes, and DME's). We can discuss their effects without ppl dismissing them and pretending the GPU specs themselves tell the whole story when obviously that's not the case at all.

Here was that the whole blitter argument was born,and was fast embrace in many other places on the internet because if it comes from Beyond3D it most be true,i can tell you this because i post on other forums and many people take post from here like god spell.

I know precisely what you mean here. It is obnoxious, I agree.

You say that Orbis can't feed 32ROP's at 176/GBs but what about feeding 24 or 26,sure it can't feed them all but that doesn't mean it can't feed more than 16,which most people think is perfect for Durango 12CU,have you look at it from that point.?

My point was the vast majority of ppl out there are looking at 32, saying it is twice as big as 16, and then pretending there is a 100% difference when you can't actually use all 32. I never suggested they can't feed more than 16. They wouldn't have 32 there to begin with if that were the case.

...what about the whole 68/GBs+104/GBs vs Orbis more straight forward 176/GBs that has been endlessly try to been pass as equal and in some cases even better.?

I don't see it being 'endlessly passed off as true'. I also don't see ppl diving into discussion about how the eSRAM, DME's and RAM in Durango work as a whole either unfortunately. I see far more dismissals of those parts entirely than I do discussion on their arrangement into a coherent design. That's part of why I came here recently.

As for efficiency, I think the discussions about the DME's here and the insider info from ppl like Arthur on GAF suggested that ppl felt Durango was being designed to be extremely efficient. The quantitative treatment of that is almost impossible atm though since anytime anyone asks for info on how PC GPU setups work with latency the comparisons get messy and fall apart pretty quickly.

Most developers will always be polite even if one hardware is at disadvantage not saying this is the case but that how it is,most developer will refuse to answer that and tell you both are about on par.

But anonymous devs? I dunno about that. Does this mean we should ignore them entirely? I'm not so sure. One dev on GAF even alluded to a PS3/360 type difference here...which is a lot less than anything ppl are asserting based on GPU specs alone.

Based on what sweetvar said I am wondering if MS wants to push certain tasks over to the CPU that have in recent yrs moved toward the GPU end of things? DDR3 instead of a GDDR variant on top of a beefed up custom CPU and his comments about how AMD's CPU ppl felt Durango was a monster and a 'supercomputer' might bear that out, no? In that case, maybe they don't need as many CU's? Maybe other dedicated peripheral hardware can help with the same operations as the CU's? I'd think the display planes could help devs manage fill rates for the ROP end of things effectively. Plus aren't CPU operations much more sensitive to latencies?

 

[expletive]

To be fair, that was due to lead free solder, no? I don't think actual engineers would need to worry about that happening again unless they are incredibly incompetent in the first place. That's not to say it is realistic to think they'd do this and want to add more cooling to the setup though.

I would hope not they had indeed learned their lesson from last time!

All that said, if MS even feel like they need to make up for some compute disparity, they don't need to make up the whole thing by burning up their chips. All they would need to do is understand what the inflection point between meaningless/meaningful in terms of power disparity is and hit that. They have plenty of development studios to which they could call a round table and ascertain where that might be and just target that in some fashion.

Based on any rumors we have right now that would be roughly

1900 GFLOPS (PS4 GPU 1800 + 100 CPU)
- 1400 GFLOPS (1200 GPU + 200 CPU)
= 500 GFLOPS difference

They don't need to find 500 GFLOPS to make the performance comparisons meaningless for compute resources, they'd probably only need to have half that and then 250GFLOPs becomes a rounding error. 2 re-activated CUs and a 10% meager OC ought to do it.

 

[Shifty Geezer]

...and a 10% meager OC ought to do it.

If a 10% overclock is easy to pull off, how come RSX got reduced from 550 MHz to 500? Why was Xenos 500 MHz and not bumped up to 550? Overclocking, as I've said elsewhere if not here, is literally clocking a part beyond it's design spec, hence the name. This generates heat. You either have to cool it more, or risk failure. Some chips will be made more able to accommodate increases in clocks just by chance, which is where PC enthusiasts can push their clocks as high as possible and get different results. You need to deal with the added heat or your chip will burn out, and clock too high beyond specs and you get too many parts that just can't reach that and you need to chuck out more components and increase costs exponentially.

There's no such thing as 'just overclocking' a consumer device. You need to hit a reliable power point and stick to it. If techie consumers want to overclock their device and post on the internet "2 GHz Durango! Lolz" with a YouTube video of their console immersed in refrigerated oil, that's their personal risk and not one customer support needs to worry about, unlike 100% of consoles edged beyond their reliable limits.

 

[Prophecy2k]

To be fair, that was due to lead free solder, no?

Actually I think that we do need to start being truthful and calling a spade a spade.

360s (and early PS3's to a lesser extent) died in droves early this generation because we had consoles designed to push the limits of TDP, die size, manufacturability and thermals yet without an even remotely adequate cooling solution that was fit for purpose. I know we've all for a while now been content to blame the EU regulation that forced lead-free solder on Sony and MS, but in actual fact the switch to lead-free only proved to exacerbate an issue that was already inherent.

RROD happened because the console's cooling solution wasn't adequate. It wasn't sufficient to remove the heat from the chips, and so the excess simply dissabpated into the board causing it to flex and the chips come loose and/or become damaged. The poor cooling solution allowed, through extended play sessions, temperatures to reach levels they should never have been able to do so in a closed box with a cooling system designed specifically for said purpose. Had the cooling solution been properly designed, the switch to lead-free solder shouldn't even have mattered, as there should have been suffient margin inherent in the cooling system design to allow it to keep internals temperatures well below the levels that would end up melting solder and flexing the PCB.

So no it wasn't lead-free solder that caused RROD (or YLOD). It was the console vendor not designing a box with a cooling solution fit for purpose.

 

[AlphaWolf]

If a 10% overclock is easy to pull off, how come RSX got reduced from 550 MHz to 500? Why was Xenos 500 MHz and not bumped up to 550? Overclocking, as I've said elsewhere if not here, is literally clocking a part beyond it's design spec, hence the name. This generates heat. You either have to cool it more, or risk failure. Some chips will be made more able to accommodate increases in clocks just by chance, which is where PC enthusiasts can push their clocks as high as possible and get different results. You need to deal with the added heat or your chip will burn out, and clock too high beyond specs and you get too many parts that just can't reach that and you need to chuck out more components and increase costs exponentially.

There's no such thing as 'just overclocking' a consumer device. You need to hit a reliable power point and stick to it. If techie consumers want to overclock their device and post on the internet "2 GHz Durango! Lolz" with a YouTube video of their console immersed in refrigerated oil, that's their personal risk and not one customer support needs to worry about, unlike 100% of consoles edged beyond their reliable limits.

I think the counter argument is there are many GCN parts reliably clocked above 880Mhz, The question is how much thermal increase their box can handle.

 

[Shifty Geezer]

Actually I think that we do need to start being truthful and calling a spade a spade.

360s (and early PS3's to a lesser extent) died in droves early this generation because we had consoles designed to push the limits of TDP, die size, manufacturability and thermals yet without an even remotely adequate cooling solution that was fit for purpose.

That's not true. PS3's cooling solution was superb, well implemented (maybe not the thermal pad), and certainly Sony's engineers aren't stupid enough to fail to go through their thermal management maths and cope with the heat. PS3's fan had several speeds - if it was overheating that caused the issues, the fan would spend time on the fastest, very loud, setting, before the console went belly up and the chips burnt out. You should also have seen overheating errors like corrupted visuals before the final death, similar to overclocked, burnt out PCs. That wasn't the way with YLOD. PS3 just wouldn't switch on, caused by fatigue in the solder points, proven by temporary heat gun fixes that wouldn't do jack if the component had fried. We also have independent repair companies telling us it was the solder.

MS's situation is maybe more debatable, but had the solder been leaded, there's reasonable chance MS's hot chips would have survived just fine even with board buckling.