Masayuki Chatani (SCEI CTO) interview

one

Unruly Member
Veteran
http://pc.watch.impress.co.jp/docs/2005/0520/e303.htm

PC Watch hosts the interview with Masayuki "Masa" CHATANI, SCEI CTO, about PS3. Here's a translation.

honda01.gif


Q. You must have a good reason to add 2 HDMI transmitters, what usage do you expect for dual HDMI?

A. It's purely because we wanted to add them. Since multi display expands gaming like a horizontally wide dual display or an extended game screen, we've wanted to add it at any cost. You don't have to connect 2 HDTVs, a small sub-display is enough.

Q. Does HDMI contain S/PDIFx6?

A. Yes it's totally compliant to the HDMI spec and audio output is assigned to it.

Q. This is related to the OS too, when you have a communication tool (such as IM) in the firmware and play games talking with your buddies by the same tool, can you use a sub-display even in games that don't support dual display?

A. Yes. A communication tool is embedded in the firmware of course. You want to communicate while gaming, don't you? You can popup a chat window in the main screen or in the sub screen.

You know PSP has much more functions in the OS than PS2 does. Likewise, the PS3 system will be extended and evolved one, in which a communication tool via internet is included.

Q. Judging from the message that you recommend PS3 as a digital media center, it's very natural, in a sense, that it has an SD card slot. But as a default feature, in a Sony-branded product? It's unprecedent, isn't it?

A. To begin with we expect use with a digital camera as you can see from the fact that it has a CF slot too. Also, SD card is in many devices including mobile phones. Users will be annoyed if media have to be converted except for memsticks. Of course you want to put save data in all media types, don't you?

Q. I can see you need those 3 slots to support digital cameras, but is it possible to put music purchased via EMD onto an SD card by this SD slot?

A. It's dependent on DRM rather than on the slot types. If you need Magic Gate, you have to use a memstick. All media types including memstick are equivalent except for such aspects.

Q. In the spec sheet it has remote-control via IP network, what kind of usage is expected?

A. Controlling a game via IP, or controlling PS3 via an IP-based remote controller.

Q. It's supposed to play BD-ROM video format, is it implemented on sale?

A. Yes of course. The video application format is not finalized but almost done and HD video decoding is very easy for us. Interactive functions will also be implemented in software, just after it's finalized.

Q. The published PS3 diagram doesn't contain a sound chip.

A. This time it has no sound chip because Cell can generate all sounds. In the leaf demo in the event we create sound data by running a code to simulate sound of a leaf in a wind then assign it to 5.1 channels to synthesize surround sound. CELL has abundant processing power and can do sound processing with ease so it doesn't need a dedicated chip to create a sound wave in the DSP style.

Q. Gigabit Ethernet connectors are listed as In x 1 + Out x 2, what does it mean? Does it have router function and are they for WAN/LAN? Or do you use a special connection to connect multiple PS3s together?

A. It has no router function. It's supposed to be a switching hub internally, but I don't know about the meaning of In/Out frankly so will answer about it at the next opportunity. Of course you can connect PS3s together. After it's on sale, some will make a supercomputer by connecting many PS3s. Apparently Sony Picture Entertainment is considering to use PS3 in a rendering farm for movies.

Q. For the connectors or the design, it looks very AV-oriented. It's based on the PS2 design and just rounded, but when put horizontally it has the atmosphere of B&O (Bang & Olufsen) in the top-loading era. Is it strongly meant as one of AV products?

A. When TVs in home get HD, we recognize PS3 as the standard in the HD era. Thus we made it to be able to play all 12cm-disc formats.

Q. Considering power of CELL and GPU, the case size, and the internal power unit, it seems tough to manage heat. How is the noise?

A. Since quietness is very important, our hardware team is trying to make it as quiet as possible. Of course it has all power-saving efforts in the chip-level. Though it's not compared with PS2 yet, I'd like to make it more quiet in the final product.

Q. The demo in the press conference suggested the possibility that physics simulation can change games. Though it has the aspect that it's done to utilize the idiosyncrasy of CELL, I think whether PS3 can be an attractive game console or not is dependent on how it can achieve in this area.

A. In PS3, in addition to the full HD output, the rendering quality per se is very high. But the more high-def and the more real it gets, the more wierd it's felt when it animates with a strange manner that's not experienced in a real life. For example it's very uncomfortable if skin contours don't change naturally when facial expression is changed in rendering a face through which even cutaneous sensasion is conveyed.

It's only an example, but on all occasions we have to actualize natural animation proportionally to improvement in graphics. In games it backfires if graphics improves while AI/simulation stays the same.

Q. To which part of physics simulation does parallel processing by 7 SPEs contribute?

A. So far, a simulation code in games had handled an in-game character as one rigid body. From now on more soft things such as human, plants and fluid can be represented more naturally. When you represent one character, you can generate natural animation and natural facial expression by spliting it into multiple parts and calculating them separately.

Likewise in a car-race game each part can have a different behavior calculation. For example, how a car moves when left and right wheels are on different kinds of surface can be solved by calculating behavior of each part.

Q. Physics simulation includes games such as Gran Turismo in which simulation precision per se is connected to the value of the game and games in which it's just an assistant. How easy will it be for developers to use physics simulation as a tool?

A. Of course there'll be a vendor that can implement and tune simulation code. It always surprises us, entertainment software developers have very high skill. No matter how difficult some processing seems, they come up with a very fast code. One point is to supply a tuning tool to make use of their skill.

The other point is, since there are middlewares for physics simulation, it can be easily implemented with them. In the case of PS2 it had taken some time until they got usable, but this time Havok and others are already in. For example, automatic generation of natural terrain or cloth simulation code are supplied.

Q. Is the PS3 development environment different from that of PS2?

A. Basically it doesn't change very much. The target hardware, the regular development environment, and tools. The difference is Cg compiler and tools, OpenGL/ES support by the change to nVIDIA.

Q. Is it the same kind of approach in performance tuning?

A. Since we have performance analyzer tools for cache usage and bus usage that game developers have evaluated so far, we extend them further. We concentrate on tools to extract programmers power. Also, CELL has the function to emit processor usage statistics as a log. By using this, you can know detailed usage statistics of each core in CELL.

Q. In regard to PS3, because few devkits are supplied, I hear 2 kinds of opinions. Some expect it'd be hard to extract performance and others are surprised as they can extract performance very easily in an actual development. How is it in the real development scene?

A. To be honest, since it has so much margin in performance, we've heard no such feedback that it's difficult to develop on it. Naturally it's different from single-core programming, but it doesn't add to man hours required for game development that much. I repeat it but entertainment programmers have tremendous ability.

For example we showed the demo that renders London City, it's not rendered in the GPU but the CELL does lighting and texture processing then outputs it to the frame buffer. Even without GPU, only CELL can create good enough 3D graphics.

Q. A few years ago it seemed you were lost for how you use CELL in augmenting entertainment factors. Have you found the answer?

A. As I told you here, physics calculation, simulation, and AI. The better graphics gets, the more often you see unnaturalness. To extract reality that matches graphics quality, programming abstract animation is not enough already. CELL exists to represent the worldview closer to nature. In future, expression in games will evolve into the different level from until now.

(The row to wait for the PS3 booth reached over 2 hours long)
honda02.jpg
 
Last edited by a moderator:
For example we showed the demo that renders London City, it's not rendered in the GPU but the CELL does lighting and texture processing then outputs it to the frame buffer. Even without GPU, only CELL can create good enough 3D graphics.
Now THAT's impressive.
 
SedentaryJourney said:
For example we showed the demo that renders London City, it's not rendered in the GPU but the CELL does lighting and texture processing then outputs it to the frame buffer. Even without GPU, only CELL can create good enough 3D graphics.
Now THAT's impressive.


Wow indeed. This ties in pretty nicely with the commentary in the other WatchImpress article.

What exactly would the RSX have been doing there, then? Screenshot analysis, stat! ;)
 
SedentaryJourney said:
For example we showed the demo that renders London City, it's not rendered in the GPU but the CELL does lighting and texture processing then outputs it to the frame buffer. Even without GPU, only CELL can create good enough 3D graphics.
Now THAT's impressive.
Yes that's really quite something.
 
With the memory logic on the 360 on the GPu couldn't the Tri-core be used in similar way? Or is there a lot more to it.
 
Well the fact this scene was produced on Cell has already been mentioned on this forum. Regards whether it's possible on a tri-core Xenon CPU, I wouldn't have thought so based on real-world performance of similar cores. Cell is designed for this sort of data manipulation and works more akin to a GPU then a CPU. I'm kinda curious now. I haven't seen a software renderer on a PC in ages and don't know what they can manage.
 
Shifty Geezer said:
Well the fact this scene was produced on Cell has already been mentioned on this forum. Regards whether it's possible on a tri-core Xenon CPU, I wouldn't have thought so based on real-world performance of similar cores. Cell is designed for this sort of data manipulation and works more akin to a GPU then a CPU. I'm kinda curious now. I haven't seen a software renderer on a PC in ages and don't know what they can manage.

Even bilinear filtering on a PC CPU makes everything lag like hell. Last time i checked that is.
 
Shifty Geezer said:
Well the fact this scene was produced on Cell has already been mentioned on this forum. Regards whether it's possible on a tri-core Xenon CPU, I wouldn't have thought so based on real-world performance of similar cores. Cell is designed for this sort of data manipulation and works more akin to a GPU then a CPU. I'm kinda curious now. I haven't seen a software renderer on a PC in ages and don't know what they can manage.

XeCPU has 3 VMX128 cores at 3.2 Ghz, at worst it could get within 30% of Cell doing the same job, because the cores of Xenon are more flexible then SPU in all likelyhood it could achieve results much closer than that.

Cell has a FLOP advantage, XeCPU has a flexibilty advantage... I suspect a tuned advanced software engine for both would be within 70-80% of each other. I'm not even sure that if you have lots of vertex and texture data, that XeCPU would lose...
 
DeanoC said:
Shifty Geezer said:
Well the fact this scene was produced on Cell has already been mentioned on this forum. Regards whether it's possible on a tri-core Xenon CPU, I wouldn't have thought so based on real-world performance of similar cores. Cell is designed for this sort of data manipulation and works more akin to a GPU then a CPU. I'm kinda curious now. I haven't seen a software renderer on a PC in ages and don't know what they can manage.

XeCPU has 3 VMX128 cores at 3.2 Ghz, at worst it could get within 30% of Cell doing the same job, because the cores of Xenon are more flexible then SPU in all likelyhood it could achieve results much closer than that.

Cell has a FLOP advantage, XeCPU has a flexibilty advantage... I suspect a tuned advanced software engine for both would be within 70-80% of each other. I'm not even sure that if you have lots of vertex and texture data, that XeCPU would lose...
interesting.. thanx for the insight DeanoC :)
 
DeanoC said:
XeCPU has 3 VMX128 cores at 3.2 Ghz, at worst it could get within 30% of Cell doing the same job, because the cores of Xenon are more flexible then SPU in all likelyhood it could achieve results much closer than that.

Cell has a FLOP advantage, XeCPU has a flexibilty advantage... I suspect a tuned advanced software engine for both would be within 70-80% of each other. I'm not even sure that if you have lots of vertex and texture data, that XeCPU would lose...

That is interesting DeanoC, thanks for sharing.

The CELL is obviously a FLOPs monster, but your quote kinda of backups the general gist that these are different architectures with different strengths. The CELL is an awesome chip and a great platform (PS4 woot!), but it is nice to hear someone who has worked with both give a nice feedback. I liked it so much I made it my sig :D

Btw, have you been complimented on your Sweet looking PS3 game in the last hour? If not... :oops: Nice job man.
 
DeanoC said:
Shifty Geezer said:
Well the fact this scene was produced on Cell has already been mentioned on this forum. Regards whether it's possible on a tri-core Xenon CPU, I wouldn't have thought so based on real-world performance of similar cores. Cell is designed for this sort of data manipulation and works more akin to a GPU then a CPU. I'm kinda curious now. I haven't seen a software renderer on a PC in ages and don't know what they can manage.

XeCPU has 3 VMX128 cores at 3.2 Ghz, at worst it could get within 30% of Cell doing the same job, because the cores of Xenon are more flexible then SPU in all likelyhood it could achieve results much closer than that.

Cell has a FLOP advantage, XeCPU has a flexibilty advantage... I suspect a tuned advanced software engine for both would be within 70-80% of each other. I'm not even sure that if you have lots of vertex and texture data, that XeCPU would lose...

It doesn't really have that to spare, though, at least relative to Cell? I would have thought there'd be a lot more headroom to do this kind of thing on Cell on top of general game CPU processing, than there would be on X360's CPU (?)

There is a big floating point advantage there. General software rendering aside, that's gotta be useful for other more specific floating-point-heavy graphics tasks that you might like to use the CPU for?
 
DeanoC said:
Shifty Geezer said:
Well the fact this scene was produced on Cell has already been mentioned on this forum. Regards whether it's possible on a tri-core Xenon CPU, I wouldn't have thought so based on real-world performance of similar cores. Cell is designed for this sort of data manipulation and works more akin to a GPU then a CPU. I'm kinda curious now. I haven't seen a software renderer on a PC in ages and don't know what they can manage.

XeCPU has 3 VMX128 cores at 3.2 Ghz, at worst it could get within 30% of Cell doing the same job, because the cores of Xenon are more flexible then SPU in all likelyhood it could achieve results much closer than that.

Cell has a FLOP advantage, XeCPU has a flexibilty advantage... I suspect a tuned advanced software engine for both would be within 70-80% of each other. I'm not even sure that if you have lots of vertex and texture data, that XeCPU would lose...

Maybe with the xCPU they want to process IA and some "superficial" stuff :rolleyes:


No way man. Xbox 360 has been designed to have a CPU and a GPU. Both of them working at their own work.

But that use of cell-RSX is quite amazing :oops:
 
Vaan said:
Maybe with the xCPU they want to process IA and some "superficial" stuff :rolleyes:
Well hopefully Cell willing be doing some AI and "superficial" stuff as well...


Vaan said:
No way man. Xbox 360 has been designed to have a CPU and a GPU. Both of them working at their own work.

But that use of cell-RSX is quite amazing :oops:
So has PS3, if you "abuse" cell doing software rendering you can abuse XeCPU in a similar manner. Except for specific procedureal effects you aren't but we were comparing CPU vs CPU in a software renderer role :rolleyes:
 
Forget complimenting Deano about Heavenly Sword (not really :LOL: ). Thank him for his honesty and ability to stay away from hyperbole WRT the pros and cons of the new platforms, even though his publisher is SCEE. Pure class.

OT question Deano. Is Ninja theory allowed to work with different publisher concurrently?
 
Vaan said:
Maybe with the xCPU they want to process IA and some "superficial" stuff :rolleyes:

Yeah, the 3 PPC cores (of which the CELL has 1) do nothing... :rolleyes: I think DeanoC pretty much laid it out well: Flexibility vs. FLOPs. Different architectures, different strengths. Each will have game scenarios they excell at. The question is which will access more frequently, and none of us know that yet. My bet is on CELL, but but as DeanoC was saying it is not like the xCPU is a wuss.

No way man. Xbox 360 has been designed to have a CPU and a GPU. Both of them working at their own work.

But that use of cell-RSX is quite amazing :oops:

Go read the patents. They explicetly state offloading some of the GPU vertex processing to a CPU core.

So does that make the xCPU-Xenos chip amazing also :oops:

;)
 
nelg said:
Forget complimenting Deano about Heavenly Sword (not really :LOL: ). Thank him for his honesty and ability to stay away from hyperbole WRT the pros and cons of the new platforms, even though his publisher is SCEE. Pure class.
Its just tech, if you talk to the engineers from Sony or MS they are really complementary about the competation. Its only people who are blinkered with respect to a brand who have to come up with rubbish arguments.

In the example of a software rasterisor its clear to anybody who has ever written one that the memory architecture of Cell is gonna hurt. The reason is simple for rasterisation FLOP count is largely irrelevant, its lots of data movement and random access to memory. Now if we want to talk about procedural synthesis that creates procedural textures, than Cell is gonna whip XeCPU into touch in most cases...

nelg said:
OT question Deano. Is Ninja theory allowed to work with different publisher concurrently?
Don't know to be honest, but we are simple not big enough to do anything but Heavenly Sword at the moment.
 
Deano - can I what you think about Cell as an assitant to the GPU for lighting? Generating more "dynamic" data for input to a GPU shader versus precomputed data used in some lighting algorithms, perhaps? Harrison seemed to allude to the possibility with reference to the DocOc demo, though my question isn't necessarily specifically in terms of the techniques used there (i doubt those techniques are generally useful).
 
Back
Top