1up claims Kojima says MGS4 demo could be done on 360

[Deepak]

also: I wouldn't hold my breath for any mindblowing, photo-realistic textured games running at 1900x1080 at 60fps on the ps3 with 512mg ram.

Why not, GT4 was 1080i with 32MB RAM this gen.

 

[Laa-Yosh]

Why not, GT4 was 1080i with 32MB RAM this gen.

Yeah, what, 640 * 540 scaled up to 1920 * 1080i?

 

[aaronspink]

CPU: CELL has maximum of 9 usable hardware threads (2 PPE, 7 SPE) with 256k "cache"/thread but Xenon has only 4 with 256k cache/thread. Use of additional available hardware threads or blocked cache method on Xenon for streaming compressed data for GPU = even less usable cache for sharing by remaining threads and higher instability and perhaps lower overall performance.

Incorrect, Xenon has 6 hardware threads.

Bandwidth: CPU and GPU bandwidth is higher and XDR has very low latency to reduce cost of cache miss.

XDR isn't very low latency, the dram core is effectively the same as any other dram core.

Shaders: RSX (maybe) has more pixel shader ALU/second than entire Xenos GPU as well as additional Vertex Shader ALUs. ALU implementation for vertex shader use by Xenos = even fewer ALU for pixel shader use so even larger gap is developed.

Shader power is most likely in Xenos's favor. You are making a lot of assumptions about the RSX which may or may not be true.

Fill-rate: RSX has 16 ROP at 550mhz but Xenos only has 8 ROP at 500mhz so PS3 fill-rate is greater than double so more than double quantiy of generated pixels.

The fill-rate of RSX will be severly limited by the memory bandwidth available. At least half the time the ROPs will be idle waiting on memory if RSX indeed has 16 ROPs. Realistically, Nvidia will cut it down to 8 ROPs.

Aaron Spink
speaking for myself inc.

 

[TurnDragoZeroV2G]

Incorrect, Xenon has 6 hardware threads.

I'm fairly certain he was trying to equate things with how much cache, or local store in SPE's case, each thread has access to. Not that I think it's a very good comparison, however.

 

[Entropy]

Incorrect, Xenon has 6 hardware threads.

Those "hardware threads" are a lot weaker than those of the Cell though. Using Intel parlance we are talking Xenon "hyperthreading" vs complete separate processors with their own memory pools. It's a world of difference, even if I trust IBM to have done a somewhat better job than Intel did, judging from Power5.

XDR isn't very low latency, the dram core is effectively the same as any other dram core.

Latency is dependent on a lot of factors. The combined latency of memory controller, XDR protocol and dram core is only the start, and even that start seems to be obscure (or I've just looked in the wrong places. Been busy.) And there is bus turn around, multi-processing effects et cetera. I'm not making any claims either way, quite the contrary, if you or anyone else has solid data, I'd be most interested.

Shader power is most likely in Xenos's favor. You are making a lot of assumptions about the RSX which may or may not be true.

To be blunt, nobody not under an NDA knows, and even for them the answer may not be clear cut.
We'll see when (if) solid data gets out into the public domain. Up until then it is all just prejudice talking either way.

 

[SubD]

Looks like the 360 == Dreamcast continue to hold true.

Welcome to 2000, again.

 

[ihamoitc2005]

Assumptions.

Incorrect, Xenon has 6 hardware threads.

If you read my post with care you will find I said 4 threads with same cache per thread as CELL of 256k/thread. Question is not theoretical total threads but practical threads. More hardware thread use of Xenon will result in less than 256k cache per hardware thread average. You must consider what is purpose of cache and why it is important to have enough. It is not wise to have to have cache-miss when frame-rate is important, no?

Can Xenon do 6 fully independent hardware threads or must resources in each core be shared by two threads of each core? No. It is 3 primary threads and 3 secondary threads.

CELL has 8 primary threads and 1 secondary thread. So you can say in simplistic form it is 4.5 threads for Xenon and 8.5 threads for CELL.

XDR isn't very low latency, the dram core is effectively the same as any other dram core.

Maybe you forgot difference in clock-speed. There is reason XDR is clocked at 3.2Ghz, same as CPU of PS3.

Shader power is most likely in Xenos's favor. You are making a lot of assumptions about the RSX which may or may not be true.

Do you mean theoretical or benchmarks? If you mean theoretical, you will have to demonstrate your calculations. If you mean benchmarks, you will have to provide links.

Theoretical calculations I have seen in this forum show opposite of what you said. I have not seen benchmarks for Xenos or RSX. In fact, we do not know what RSX is, only guess.

Also, although you have said I make too many assumptions about RSX yet I feel you have made a large assumption for claim that shader power is most likely in Xenos's favor despite much lower published specs of Xenon. You also make following assumption about memory bandwidth.

The fill-rate of RSX will be severly limited by the memory bandwidth available. At least half the time the ROPs will be idle waiting on memory if RSX indeed has 16 ROPs. Realistically, Nvidia will cut it down to 8 ROPs.

Maybe maybe not. If you have evidence of this I would like to see. Developer on this forum has said 1080P at 30fps is possible for Heavenly Sword. Are there any Xbox360 games with true 720P at 60fps with no upscaling? This is something we need to discover. I can only say that "you are making alot of assumptions about RSX which may or may not be true."

 

[Bill]

Looks like the 360 == Dreamcast continue to hold true.

Welcome to 2000, again.

Except for the fact that in this case, X360 is more powerful than the Vaporwarestation 3

 

[weaksauce]

Except for the fact that in this case, X360 is more powerful than the vaporware station 3

Well people say the dreamcast was more powerful. :smile:

 

[Bill]

Well people say the dreamcast was more powerful. :smile:

And it was

 

[weaksauce]

And it was

So it is 2000 all over again?

 

[Bill]

Yep.

PS3=Dreamcast. Except in this case less powerful than the competition.

 

[mckmas8808]

Before stupid fan***s get their panties all in a bunch over the Kojima said PS3=360 read this.

IGN: OK, but the other part of my question was, does Mr. Kojima see some really great possibilities for making future games on the Xbox? I realize that this is a rather difficult transition to make going from the PS2 to the Xbox, but I'm interested in whether Mr. Kojima sees any fantastic possibilities making great games on the Xbox?

Kojima: Ahhh, the Xbox, the graphics are probably a little better, but the systems are pretty much the same, the Xbox and the PS2. They're not that different. If I spent a lot of time creating a game specific for the Xbox, I could come up with something great. But I feel that they are basically the same, you know, memory wise, etc.

http://xbox.ign.com/articles/360/360488p1.html

Kojima said the PS2 and Xbox were the same also. So lets stop the stupidness. Okay?

 

[TurnDragoZeroV2G]

Do you mean theoretical or benchmarks? If you mean theoretical, you will have to demonstrate your calculations. If you mean benchmarks, you will have to provide links.

Theoretical calculations I have seen in this forum show opposite of what you said. I have not seen benchmarks for Xenos or RSX. In fact, we do not know what RSX is, only guess.

Would those theoretical specs account for texturing as well? And we have plenty that says what RSX is. E3. Interviews. Etc. Just because everything isn't known doesn't mean nothing is known.

You also make following assumption about memory bandwidth.

16 pixel writes a clock would require a significant amount of bandwidth. Xenos needs 256GB/s (uncompressed) to support its fillrate w/ 4xAA. You're asking RSX to handle more with less than a fifth of that bandwidth (which would require using XDR Ram and such too, which probably isn't realistic at all for the framebuffer) or less than a tenth, if using only GDDR3. Compression will make up for alot, but it can't take care of everything. And of course, there has to be bandwidth in each pool used for textures, geometry, etc. otherwise megabytes of memory go to waste.

The half the time comment may be overexaggerated or not, but the more important part is that it will be bandwidth limited. Which is entirely true.

Are there any Xbox360 games with true 720P at 60fps with no upscaling? This is something we need to discover.

Which isn't related to RSX's fillrate being bottlenecked or not, since 360's limit is not its bandwidth, which you should know considering you're on B3D, with the Xenos article and these forums. Meanwhile, being able to support or not support the theoretical max fillrate doesn't necessarily mean being able to support 1080p at 30fps or not.

 

[ihamoitc2005]

Evidence

Would those theoretical specs account for texturing as well? And we have plenty that says what RSX is. E3. Interviews. Etc. Just because everything isn't known doesn't mean nothing is known.

His comment was regarding shader power, not texturing or effect & frequency of texture interference with shader power no? So that is why I referred to all those calculations on this forum.

For effect of texture interference with shader power I can only guess but if someone has approximate average proportion of this we can then compare with RSX if assumption is RSX = G70 @ 550mhz.

16 pixel writes a clock would require a significant amount of bandwidth. Xenos needs 256GB/s (uncompressed) to support its fillrate w/ 4xAA. You're asking RSX to handle more with less than a fifth of that bandwidth (which would require using XDR Ram and such too, which probably isn't realistic at all for the framebuffer) or less than a tenth, if using only GDDR3. Compression will make up for alot, but it can't take care of everything. And of course, there has to be bandwidth in each pool used for textures, geometry, etc. otherwise megabytes of memory go to waste.

The half the time comment may be overexaggerated or not, but the more important part is that it will be bandwidth limited. Which is entirely true.

If actual developer says 1080P at 30fps is possible on PS3 in a real game with textures and geometry, then we know bandwidth is enough no? Developer even said he thinks 60fps may be possible. It is not clear if 30fps is CPU limitation or GPU limitation.
So maybe compression is very effective because 30fps is accomplished and perhaps even possibility for 60fps. Also, do not forget real-time PS3 demo of MGS4 Lair also showing probably not fully optimized RSX output capability for real-time graphics.

Therefore with evidence from developers that bandwidth is sufficient for atleast 1080P at 30fps for game of exceptional Heavenly Sword fidelity, I am sorry my friend I cannot agree that it is true that bandwidth is serious limitation.

Real world performance is always better evidence than calculations since calculations for RSX are based on guesses.

Which isn't related to RSX's fillrate being bottlenecked or not, since 360's limit is not its bandwidth, which you should know considering you're on B3D, with the Xenos article and these forums.

Xbox360 is reference platform for any discussion of PS3 capability so real-world Xbox360 output capability is what is important regardless of cause. If PS3 can be equal or better output of Xbox360 we know it is enough, if PS3 has worse output then we know it is not enough and must look for cause.

Meanwhile, being able to support or not support the theoretical max fillrate doesn't necessarily mean being able to support 1080p at 30fps or not.

Fill-rate potential and pixel output performance is not neccessarily connected, but not all pixels are the same no? This is why original Xbox can output 720P with 6.4GB/s bandwidth.

 

[aaronspink]

If you read my post with care you will find I said 4 threads with same cache per thread as CELL of 256k/thread. Question is not theoretical total threads but practical threads. More hardware thread use of Xenon will result in less than 256k cache per hardware thread average. You must consider what is purpose of cache and why it is important to have enough. It is not wise to have to have cache-miss when frame-rate is important, no?

I have a pretty good idea of what cache is. Likely better than anyone else on this site. I stand by my statement, Xenon has 6 hardware contexts and Cell has 9. If anything were to be derated, it would have to the the SPE's since they don't have direct memory access and must "waste" significant portions of their local store to bring in and send out data.

Can Xenon do 6 fully independent hardware threads or must resources in each core be shared by two threads of each core? No. It is 3 primary threads and 3 secondary threads.

This is once again incorrect. Xenon contains 3 cores each of which has 2 hardware contexts. You can't say one hardware context is primary and another is secondary, they are both co-equal.

Maybe you forgot difference in clock-speed. There is reason XDR is clocked at 3.2Ghz, same as CPU of PS3.

Hmm, maybe something to due with it being a very narrow interface that is targetted at pin efficiency? They clock the XDR at the highest data rates possible in order to reduce the number of pins required to the minimum possible. this has nothing to do with the actual latency of the memory. XDR has all the same latency issue as EDO DRAM, SDRAM , RDRAM, DDR, DDR2, and DDR3, which is it is still using a cost optimized DRAM core. This mean the same CAS latencies, the same RAS latencies, the same ACT latencies.

Do you mean theoretical or benchmarks? If you mean theoretical, you will have to demonstrate your calculations. If you mean benchmarks, you will have to provide links.

Theoretical calculations I have seen in this forum show opposite of what you said. I have not seen benchmarks for Xenos or RSX. In fact, we do not know what RSX is, only guess.

Xenos has 48 VEC4+1 ALUs for 240 GFLOPS of computational power.
G70 has 2 * VEC4 + 2 ALUS for each of its 24 pipes for a total of 264 GFLOPS, but texturing blocks out one of the VEC4 ALUs which results in 158.4 GFLOPS when texturing is required. In addition, the batch sizes for banches on G70 are huge resulting in significant performance loss when branching.

Overall, Xenos will have a definite edge in shading.

Maybe maybe not. If you have evidence of this I would like to see. Developer on this forum has said 1080P at 30fps is possible for Heavenly Sword. Are there any Xbox360 games with true 720P at 60fps with no upscaling? This is something we need to discover. I can only say that "you are making alot of assumptions about RSX which may or may not be true."

This has nothing to do with fill rate or bandwidth. And no, although I like and would like to believe Deano, I don't believe that the final game with all effects will be able to do 1080P @ 30FPS. And yes, there are games that do 720P @ 60 FPS with no upscaling.

Regardless, the number of ROPs won't be the issue of them making it or not. 8 ROPs with the available memory bandwidth, won't be an issue.

Aaron Spink
speaking for myself inc.

 

[ihamoitc2005]

Heavenly Sword

I have a pretty good idea of what cache is. Likely better than anyone else on this site. I stand by my statement, Xenon has 6 hardware contexts and Cell has 9. If anything were to be derated, it would have to the the SPE's since they don't have direct memory access and must "waste" significant portions of their local store to bring in and send out data.

Hardware context of which you speak is very difference from full independent hardware threads, which is what CELL has of 7 count. Xenon has no fully independent hardware threads all have competition for resources. CELL has only 2 non-independent threads.

As for"waste" what do you mean by this? Why would you want processing core reading directly from RAM? Entire purpose of SPE design is to prevent this. You are criticizing SPE for not having a large flaw of normal processors.

This is once again incorrect. Xenon contains 3 cores each of which has 2 hardware contexts. You can't say one hardware context is primary and another is secondary, they are both co-equal.

They are equally not independent so but both must share resources because they are only thread contexts, not true full independent hardware threads. You can have one thread with very good performance but if you add second thread you will not get double performance. That is why at any one time it is more like 1.5 threads per core so that is why one can say Xenon is not 6 threads but 4.5. CELL PPE has same problem so that is why one can say CELL is not 9 threads but only 8.5.

Hmm, maybe something to due with it being a very narrow interface that is targetted at pin efficiency? They clock the XDR at the highest data rates possible in order to reduce the number of pins required to the minimum possible. this has nothing to do with the actual latency of the memory. XDR has all the same latency issue as EDO DRAM, SDRAM , RDRAM, DDR, DDR2, and DDR3, which is it is still using a cost optimized DRAM core. This mean the same CAS latencies, the same RAS latencies, the same ACT latencies

What is relevent is proportion of non-latency cycles. Because of high clock speed, this is very high for XDR therefore proportion of latency cycles is low compared to GDDR3.

Xenos has 48 VEC4+1 ALUs for 240 GFLOPS of computational power.
G70 has 2 * VEC4 + 2 ALUS for each of its 24 pipes for a total of 264 GFLOPS, but texturing blocks out one of the VEC4 ALUs which results in 158.4 GFLOPS when texturing is required. In addition, the batch sizes for banches on G70 are huge resulting in significant performance loss when branching.

You have left out reduction of available unified shaders for vertex-shading in you number for Xenos. Since 3D games use vertex-shading, with full vertex shading load of 550Mhz G70 with 8 vertex shaders, Xenos will require minimum of 9 unified shaders and maybe much more for vertex shading. So remaining unified shaders for pixel shading is only 48-9=39, with assumption of no vertex shading from CELL. In this situation Xenos has about 196 Gflops average. If CELL also provides large portion of vertex-shading, then available average Gflops for pixel-shading in Xenos falls even more.

You also fail to provide what proportion of operations are texture fetch in your number for G70. If it is 50%, then 550mhz G70 has 211.2 Gflops and if it is 10%, then it is 253.4 Gflops. What is important is not peak or lowest Gflops as you provided but average available Gflops.

As for branching again you have not provided any evidence or even guessed numbers of relevance of this issue to both GPUs or effect on average available Gflops.

Overall, Xenos will have a definite edge in shading.

Again you contradict developers but more important is that you are making large assumptions with no real data and only "vague" statements that do not account for real world 3D application with vertex shading.

This has nothing to do with fill rate or bandwidth. And no, although I like and would like to believe Deano, I don't believe that the final game with all effects will be able to do 1080P @ 30FPS. And yes, there are games that do 720P @ 60 FPS with no upscaling.

I feel Deano is a good source but this is what is called judgement call no?

Regardless, the number of ROPs won't be the issue of them making it or not. 8 ROPs with the available memory bandwidth, won't be an issue.

Are you referring to Heavenly Sword? Then you may or may not be right depending on how much bandwidth is limitation on output quality. At 1080P at 30fps, it seems bandwidth is very sufficient. For other games extra pixel fill-rate might be useful and not limited by bandwidth.

 

[TurnDragoZeroV2G]

His comment was regarding shader power, not texturing or effect & frequency of texture interference with shader power no? So that is why I referred to all those calculations on this forum.

Many shaders use textures. Shading power is useless if you have nothing to apply it to/use it for.

If actual developer says 1080P at 30fps is possible on PS3 in a real game with textures and geometry, then we know bandwidth is enough no? Developer even said he thinks 60fps may be possible. It is not clear if 30fps is CPU limitation or GPU limitation.
So maybe compression is very effective because 30fps is accomplished and perhaps even possibility for 60fps. Also, do not forget real-time PS3 demo of MGS4 Lair also showing probably not fully optimized RSX output capability for real-time graphics.

Therefore with evidence from developers that bandwidth is sufficient for atleast 1080P at 30fps for game of exceptional Heavenly Sword fidelity, I am sorry my friend I cannot agree that it is true that bandwidth is serious limitation.

Real world performance is always better evidence than calculations since calculations for RSX are based on guesses.

We weren't talking about whether RSX will output those graphics. We were talking about whether bandwidth would be a bottleneck for fillrate, and if those 16 ROPs were even necessary (if you look back to your response to aaron). You can do the calculations yourself. how will it support 8.8Gpixels/s in a realistic situation (just like: how will RSX sustain its peak theoretical shader operations all the time, even if it's 100% efficient, due to the way the nV pipelines are organized?)

BTW, what guesses do you speak of? Bandwidth has been confirmed, with no mention of changes. shader instructions/clock have been stated, and match up with G70. It has been stated that it follows from the same path as G70 (all of which implies plenty about its pipe structure, despite what some rough dot product calculations seem to lead people to believe). And didn't nV even try to release a chart showing the capabilities of G70, RSX (and even incorrect Xenos info)?

Xbox360 is reference platform for any discussion of PS3 capability so real-world Xbox360 output capability is what is important regardless of cause. If PS3 can be equal or better output of Xbox360 we know it is enough, if PS3 has worse output then we know it is not enough and must look for cause.

Real-world 360 output is bottlenecked by development time at the moment, not the issue we are discussing regarding PS3 (bandwidth and fillrate support in light of bandwidth restrictions)

Fill-rate potential and pixel output performance is not neccessarily connected, but not all pixels are the same no? This is why original Xbox can output 720P with 6.4GB/s bandwidth.

See above. (and yes, not all final pixels that are output have had the same number of pixel read/modify/writes applied to them)

 

[scooby_dooby]

Maybe you missed the memo, but Deano already said he's aiming for 720p, hopefully 60fps. He said no-one has 1080p TV's so what's the point of wasting resources targeting it?

I think he later clarified it would "support" 1080p, but it seems they are aiming for 720p internally.

 

[one]

Why is this part ignored? It looks like very loose categorization. Heck there's PC too. He's probably talking about the restrictions in the previous generation such as memory size which doesn't matter that much in this generation.

http://www.1up.com/do/minisite?pager.offset=3&cId=3146425

1UP: It's interesting that you think you could do this on 360, though.

HK: Maybe some nuance or a small details here and there might be different, but I feel that hardware is no longer a matter. I'm just talking about PS3, 360 and PC. Revolution is totally different, but there are really no differences among the other three.