NGGP: NextGen Garbage Pile (aka: No one reads the topics or stays on topic) *spawn*

[rockaman]

I wonder how a single AMD employee can know so much about both projects. Especially the part "Him and the team at AMD feels the Xbox is going to be more 'powerful'." sounds like complete nonsense to me. I doubt that AMD is so heavily unprofessional.

To be fair, the two guys I know in engineering at AMD are goofballs, and completely dorky. One of them I know is working on Orbis.

These statements can potentially come from young guys with engineering backgrounds working on the thing, bit at a lower level and have no problem throwing out there conjecture for fun to friends, some of who might frequent forums like these.

 

[bkilian]

Whoah folks, calm it down a little. As to Brad's assertion that there are aspects of the console's design compensating for the speed of Main Memory, that is obviously true from the leak documents.

Other than that, you guys keep putting words in each other's mouths and then arguing based on what you assumed they meant. Makes for entertaining posts... the first couple of times... after that it's just tiresome.

Brad never said that any of the design was in response to some other console maker, his statement is simply that there are elements of the design obviously related to addressing issues introduced by other elements of the design.

That's not to say, Brad, that there aren't other reasons for including those elements, irrespective of their role in addressing issues. ESRAM, for instance, has benefits that exceed simply adding bandwidth to the design, as very clearly called out in the leaked docs. The design is not a series of patches.

 

[bkilian]

I wasn't putting words in your mouth so much as making an inference from your post, which I assume is limited somewhat in what it can convey directly. If MS has never referred to the CPU as related to Jaguar directly it suggests there are some differences in how they operate and/or perform.

Or it was still in negotiations with suppliers and didn't want to be put in a position of giving them leverage. It may sound strange, but that has actually happened to MS before and cost them millions of dollars.

 

[ROG27]

I see little (if any) evidence to support this. To wit, I see Sony trying to follow MS's lead in this area by bundling motion cameras in every box and further trying to update their media feature set. They played catch up in this area most of this current cycle after all. You can even see specific rumors out there talking about how they will push the platform services as the selling point instead of exclusive gaming features. I see their approach here as slightly more conservative, but only because I think MS is in a much more fortified position to be ambitious in that space than Sony is atm. They have the money, the connections, and the inroads with the consumer base to go further with this stuff than Sony does atm.



I also disagree here. I don't see anything in Durango as a platform that suggests gamers are being left out in the cold. That viewpoint is largely founded on ignorance related to MS's publishing strategy of mitigating risk late cycle and the presumption that the specs for next gen GPU's (and GPU's only) are to be compared in a direct, straight forward way. There is a middleground, whether you'll admit it or not, which is that MS wants to dominate all demographics and has the positioning and cash to forge ambitions aiming to achieve as much, even if that means carefully engineering an unorthodox platform that punches above its weight.



When the GPU specs were first leaked before ppl talked about DME's or display planes or eSRAM or the CPU performance boost we just heard about...before all of that ever popped up a narrative emerged that painted Durango as a weak console and the second bkillian mentioned extra hardware blocks ppl desperate to cling to the initial sentiment started mocking the idea that additional hardware could actually make a difference. And today the more we find out the smaller that initial performance gap dwindles. Don't mock the ppl who were open minded enough to suggest that might happen.



This isn't necessarily true. I mentioned this to Brad. You can leverage very high end, relatively new rendering tech (virtualized assets) that are known to play big roles in future game engines in a way that strongly prefers more RAM to faster RAM. At least that is what I recall reading. If I am wrong there explain why but spare me a reply aimed at being condescending.

I think the contention that one is significantly more powerful than the other flies right in the face of what insiders and actual devs have said on the subject. I think if you gather all the semi-reliable info together you start seeing a picture of MS going all in on a design to leverage virtualized assets in ways Orbis isn't. We will see. Hopefully soon-ish.

I do think there is a small but significant flops and bandwidth deficit on the Durango end, but there are advantages over Sony's console in terms of processing utilization efficiencies and larger RAM volumes. At the end of the day, we are still talking about fundamentally similar core CPU and GPU architectures dressed up with smallish customizations that suit the direction these companies are aiming for in terms of key marketing differentiators (i.e., Kinect and services for MS which will help transition users into the Windows 8 ecosystem versus 3D gaming / 4K gaming for Sony which helps them cross-sell their complimentary high-end tv hardware, etc.). Most likely, these systems will be so close power-wise as to be a wash with most customers being hard pressed to see any difference in terms of 3rd party games. First party games will leverage each system's inherent strengths, strengths which owe themselves to those aformentioned key marketing differentiators.

Every company ideally wants to do everything the best. But in pragmatic terms, there will always be tradeoffs made in design decisions based on marketing bets being placed at the time.

Both companies are in it for the money. They are counting on similar (but not the same) marketing strategies in leveraging their existing products and ecosystems in order to maximize their profit potential.

 

[itsmydamnation]

Well having FMA instructions in Steamroller automatically makes it better. Durango will be a compute monster with 8 fmas and access to ESRAM.

you realise that its

8 SP ADD AND 8 SP MUL in an 8 core jaguar
vs
a total of 8 SP ADD/MUL/FMA in 8 core BD/PD/SR

assuming they can both sustain throughput to the execuation units i dont see how this makes Durango a "monster" it could very well loose. I also dont see how access to the ESRAM makes a major difference look how little difference the L3 makes in FX4300 vs A10-5800k.

 

[Silent_Buddha]

16 ROPs of Durango will max out at 102GB/s (memory could have allowed more)
32 ROPs of Orbis will max out at 176GB/s (limited by memory)
So the performance of Orbis ROPs will be 72% higher than durango.

None of the AMD cards max out their ROPs, there's always more than enough. Durango is weird, it doesn't have enough ROPs to max out it's 170GB/s. How's that supposed to be called an efficient design?

Durango's GPU will have sole access to the 102 GB/s from eSRAM. Which theoretically is about as much as those 16 ROPs needs. That 102 GB/s pool is also serviced by the DMEs but not by the CPU.

The other 68 GB/s is shared between the GPU and CPU. That is also served by the DMEs. That 68 GB/s second by the way is pretty close to the 72 GB/s available to the 16 ROP 77xx PC video cards with the caveat being that it's shared by the CPU.

So the GPU won't likely ever need to or have to address all 170 GB/s (both pools combined). In theory while it is accessing Data in one pool or another the DMEs can use excess bandwidth to shuffle data around in an attempt to prevent the GPU with associated ROPs from stalling due to waiting for data from memory. It will happen at some point, just like it'll happen on Orbis. But the whole point is to try to reduce the number of times/cases where it will happen.

I don't even think that is accurate. It's too narrow a view. I guess technically if Durango is constantly saturating its ROPs and Orbis never can, we might say that constitutes higher "efficiency". But we should be looking at the entire design. If Orbis can never be fill limited because it always has more ROPs than it needs, but Durango CAN become fill limited, that implies greater efficiency overall for Orbis. Maybe Durango strikes a better balance in terms of ratio, or even die space committed to RBEs, but that's not the same as being more efficient when running actual games.

You still seemed to be getting confused. Hence, why I keep stressing utilization. In that case that you proposed, Orbis still isn't likely to be getting full utilization of GPU resources, while Durango will be much closer to full utilization of its resources. Hell, in the case where Orbis somehow managed 100% GPU utilization it's likely that Durango's GPU would also be at 100% utilization. The difference? Orbis would be faster obviously, but utilization (efficient use of GPU resources) would be exactly the same.

And again, let me stress something that you miss every single time when replying to me. Higher utilization (efficient use) of hardware resources does not necessarily mean higher performance. Hence, lower efficiency on Orbis will still likely result in higher performance in real world rendering.

So, why keep banging on about increased utilization of available hardware resources? Because it means the performance differential is likely going to be smaller than the pure hardware specs would imply.

And that's ONLY going by the rumored leaks, which may not even contain the correct details. And certainly doesn't contain all the relevant details in order to make a truly informed decision about where absolute or relative performance for each console will be.

Additionally before you get all uptight that I'm claiming that your Orbis is inefficient... Please note that I fully expect Orbis to make more efficient use of its hardware than is currently the case on PC. And I've never gone around claiming that the GPUs in PCs are inefficient.

Just because speculating on what was added and why to otherwise standard hardware would suggest that the entire focus of the hardware choices in Durango was to maximize efficient use of available hardware while Orbis isn't quite as aggressive by going with a more "traditional" layout and data flow, doesn't mean that Orbis is some incredibly inefficient design. It's just that with what is released thus far, Durango's design appears to be putting much more effort into increasing the efficient utilization of available hardware resources above and beyond what you get with a relatively more standard configuration.

Will it bear out that way in the final console designs? Maybe, maybe not. This is speculation after all, based on incredibly incomplete information.

Regards,
SB

 

[Xenio]

the next-generation console experience will be similar to what Epic can do on PC right now.
What we’re doing on high-end PCs is going to be representative of the future consumer gaming experience and it’s going to be awesome.

http://www.edge-online.com/news/epic-on-the-industrys-giant-leap-to-next-gen/


this doesn't match what we know about durango and orbis, not with those specs, so what's missing here?
Surely Epic know very well what we don't

 

[pjbliverpool]

http://www.edge-online.com/news/epic-on-the-industrys-giant-leap-to-next-gen/


this doesn't match what we know about durango and orbis, not with those specs, so what's missing here?
Surely Epic know very well what we don't

You're confusing the capabilties and theoretical potential of the most powerful PC's with "what is currently being done on high end PC's".

Take Crysis 3 or Battlefield 3 as good examples of what is currently being done on high end PC's. High end PC's in this context can mean roughly anything from Pitcairn upwards. All GPU's fitting into that segment can max out the most demanding games on the PC today even if they are only doing to at 30fps and/or at less than 1080p.

So all epic are saying here is that the next gen consoles could run any current (or upcoming) PC game at it's maximum settings. That doesn't mean they are actually as powerful as the top end PC's or can equal their potential. They can't. Top end PC's (single GPU) are upwards of double the potential performance of Orbis and that performance is likely to go up by another 50% before the new consoles launch.

 

[Acert93]

Xenio: Orbis is essentially a "high end" PC GPU. It is not a flagship model / enthusiast GPU, but it is a gamer GPU. Looking at AMD's pecking order for non-crossfire boards Orbis has essentially a 7850 in a closed box:

7970 Tahiti (32CUs @ 1050MHz, 4.3TFLOPs)
7950 Tahiti (28 CUs @ 925MHz, 3.7TFLOPs)
7870 Pitcairn (20 CUs @ 1000MHz, 2.5TGFLOPs)
7850 Pitcairn (16CUs @ 860MHz, 1.7TFLOPs)

Orbis: 18 CUs @ 800MHz, 1.8TFLOPs

Take Epics comments into perspective: Epic is NOT aiming to push a full blown Tahiti 7970 to its max at 720p. If Epic is aiming for 2560x1600/1440 at 30Hz or 1920x1080 at 60Hz on Tahiti or the 680GTX put that into comparison for a second:

Orbis should be game for 1080p30Hz, especially if resolution can dynamically drop down slightly.

All that to say: Yes, Orbis appears to have enough horses under the hood to pull of what high end PC GPUs will be doing at higher resolutions/refresh rates.

 

[astrograd]

Whoah folks, calm it down a little. As to Brad's assertion that there are aspects of the console's design compensating for the speed of Main Memory, that is obviously true from the leak documents.

Nobody is arguing that those things don't improve things like bandwidth. But it sounds like it can do other things too, like play an important role in leveraging virtualized textures for example.

Brad never said that any of the design was in response to some other console maker, his statement is simply that there are elements of the design obviously related to addressing issues introduced by other elements of the design.

Here is what he said:

All evidence suggest they have been added to make sure it isn't far LESS efficient than Orbis.

That sure sounds like he is directly suggesting as much. Even if he simply misspoke, the premise that MS chose a certain design as a means to compensate and patch holes itself is still a pov assuming a certain context.

That's not to say, Brad, that there aren't other reasons for including those elements, irrespective of their role in addressing issues. ESRAM, for instance, has benefits that exceed simply adding bandwidth to the design, as very clearly called out in the leaked docs. The design is not a series of patches.

This is what I was trying to convey. Assuming that the motivation was aimed at patching a weak setup as opposed to playing crucial roles in a more thoughtful/ambitious design is just that, an assumption (unless there is evidence to support that). I'd argue that there seems to be growing evidence supporting the alternative moreso than what Brad is assuming.

Or it was still in negotiations with suppliers and didn't want to be put in a position of giving them leverage. It may sound strange, but that has actually happened to MS before and cost them millions of dollars.

If that's the case one would expect leaked documents from 2012/2013 to still be updated enough to reference Jaguar, no? I don't think it really matters either way, if the performance is dramatically better that matters more than what it's named.

 

[Lucid_Dreamer]

Well, if 16 rops can drive 102 GB/s, then 32 would easily exceed the rumored bandwidth of the PS4.

Shouldn't it be the other way around? Bandwidth drives ROPs, right? 200 GB/s could drive 16 ROPs, as well.

What provides a continuous 102 GB/s to the ROPs? According to the leaked documents, Durango can provide 68 GB/s of continuous bandwidth it's 16 ROPs. So, wouldn't you just double 68 GB/s to have an answer for 32 ROPs?

BIgger hard drives, when the hard drives are user replaceable and bitstreaming, which has questionable usefulness when you can pass the decoded audio in full fidelity over HDMI don't (to me) equal what was removed. I stand by my statement that the fat PS3 was the best model Sony produced.

You moved from facts to the realm of subjectivity. I showed the statements were false. Those were, indeed, hardware additions and not removals. The end.

 

[Xenio]

What provides a continuous 102 GB/s to the ROPs? According to the leaked documents, Durango can provide 68 GB/s of continuous bandwidth it's 16 ROPs.

ROP's points to esram, 102 GB/s
not to ddr
According to the leaked documents, Durango can provide 170 GB/s of combinated bandwidth, but 102 for its ROP's

 

[Silent_Buddha]

Xenio: Orbis is essentially a "high end" PC GPU. It is not a flagship model / enthusiast GPU, but it is a gamer GPU. Looking at AMD's pecking order for non-crossfire boards Orbis has essentially a 7850 in a closed box:

7970 Tahiti (32CUs @ 1050MHz, 4.3TFLOPs)
7950 Tahiti (28 CUs @ 925MHz, 3.7TFLOPs)
7870 Pitcairn (20 CUs @ 1000MHz, 2.5TGFLOPs)
7850 Pitcairn (16CUs @ 860MHz, 1.7TFLOPs)

Orbis: 18 CUs @ 800MHz, 1.8TFLOPs

Take Epics comments into perspective: Epic is NOT aiming to push a full blown Tahiti 7970 to its max at 720p. If Epic is aiming for 2560x1600/1440 at 30Hz or 1920x1080 at 60Hz on Tahiti or the 680GTX put that into comparison for a second:

Orbis should be game for 1080p30Hz, especially if resolution can dynamically drop down slightly.

All that to say: Yes, Orbis appears to have enough horses under the hood to pull of what high end PC GPUs will be doing at higher resolutions/refresh rates.

It's more likely that Epic is targeting 1920x1200 or 1920x1080 on a 7970 or 680 class GPU. 1920x1080 being the most popular desktop resolution.

So I'd imagine that at 720p or sub 720p, Orbis should be able to do something similar. And since I'd say most developers targeting advanced rendering techniques are likely targeting 720p on next gen consoles, then that fits right in.

Regards,
SB

 

[Xenio]

most developers targeting advanced rendering techniques are likely targeting 720p on next gen consoles

I hope you're joking: if the target is 720p this means that some will go sub720

Xenio: Orbis is essentially a "high end" PC GPU. It is not a flagship model / enthusiast GPU, but it is a gamer GPU. Looking at AMD's pecking order for non-crossfire boards Orbis has essentially a 7850 in a closed box:

they are talking about both consoles, not only orbis

 

[pjbliverpool]

you realise that its

8 SP ADD AND 8 SP MUL in an 8 core jaguar
vs
a total of 8 SP ADD/MUL/FMA in 8 core BD/PD/SR

assuming they can both sustain throughput to the execuation units i dont see how this makes Durango a "monster" it could very well loose. I also dont see how access to the ESRAM makes a major difference look how little difference the L3 makes in FX4300 vs A10-5800k.

If Durango was packing an 8 core / 4 module Steamroller (which I don't expect it will be) then it would also likely be running at at least twice the clock speed of the Jaguars in Orbis meaning that in the worse case scenario it would match it in SIMD throughput and in best case it would double it. With reality being somewhere inbetween.

Plus both single and multithreaded scalar performance would be well over double.

 

[MrFox]

I'm having a hard time understanding what's the big deal with ROP "efficiency", it's never been talked about for graphics cards. It's always been balanced on a need basis, and the 78xx has 32ROP for 153GB/s. That was supposed to be a good balance. Is Durango expected to perform like a 77xx card? It has a LOT more bandwidth available than a 77xx, unless the DME are wasting half of it.

Do the shaders have their own path to memory or are they limited by the ROPs bandwidth?
I was sure they had their own path, but I'm looking at GCN block schematic and it looks like the ROPs are the access path. Maybe it's too simplified.

Spoiler

 

[pjbliverpool]

It's more likely that Epic is targeting 1920x1200 or 1920x1080 on a 7970 or 680 class GPU. 1920x1080 being the most popular desktop resolution.

So I'd imagine that at 720p or sub 720p, Orbis should be able to do something similar. And since I'd say most developers targeting advanced rendering techniques are likely targeting 720p on next gen consoles, then that fits right in.

Regards,
SB

I agree with the principle but I'd say developers target 1080p/60fps on 7970 class PC GPU's were they would likely target 1080p/30fps on the consoles.

 

[Bagel seed]

From Anandtech on ROPS efficiency and bandwidth

As it turns out, there’s a very good reason that AMD went this route. ROP operations are extremely bandwidth intensive, so much so that even when pairing up ROPs with memory controllers, the ROPs are often still starved of memory bandwidth. With Cayman AMD was not able to reach their peak theoretical ROP throughput even in synthetic tests, never mind in real-world usage. With Tahiti AMD would need to improve their ROP throughput one way or another to keep pace with future games, but because of the low efficiency of their existing ROPs they didn’t need to add any more ROP hard ware, they merely needed to improve the efficiency of what they already had.

The solution to that was rather counter-intuitive: decouple the ROPs from the memory controllers. By servicing the ROPs through a crossbar AMD can hold the number of ROPs constant at 32 while increasing the width of the memory bus by 50%. The end result is that the same number of ROPs perform better by having access to the additional bandwidth they need.

http://www.anandtech.com/show/5261/amd-radeon-hd-7970-review/4

 

[Lucid_Dreamer]

Sweetvar also mentions his friend's lack of interest/knowledge in gaming, so his perspective about the hardware is going to be different than most of us would see it. They probably are looking at the overall picture to make that assessment. We OTOH tend to hone in on the CPU/GPU/memory to make our assessment.

Can someone please explain, to me, why a guy that develops performance gaming hardware would "lack knowledge in gaming"? How do they test the performance of their GPUs? Wouldn't he have to "hone in" on the CPU/GPU/memory, to make assessments? They haven't design entire systems, until now, right? I mean it's not what drives their business.

 

[3dilettante]

If your job is verifying the electrical signalling of the IO, what difference does it make what the signals are for? Lots of things go into making a complex system, and most are way below the level where the games are relevant.