The Technology of PS3

I artiklen , Keith R.
Williams skrev:

In article ,
says...
I artiklen , Keith R.
Williams skrev:

In article ,
says...
I artiklen ,
(Hans de Vries) skrev:

snipping much

Just imagine what it could mean, with Microsoft now included, If
Windows XP 64 runs not only on Xbox, but also on Playstation 3 !?
or even on Nintendo game consoles?

That would be quite an historical turning point.

You got carried away. Certainly, if the Xbox2 chip is a 440 core plus lots
of SIMD, you can forget about emulating Xbox1 games.

Why? Emulation of x86 on PPC is fairly well know to work.

....s...l...o...w...l...y....

Now wait just a minute. You stated that:

Equating the 440 core with the POWER4 core means that he is
talking in *very* broad terms.

Are you stating that you're talking about PPC here? Or are you
shifting to the specific core?

My statement about emulation was conditional on a 440 core scenario.

Hans de Vries was seeing the same core everywhere, and BlueGene uses a 440.

Emulating a ~700 MHz PIII with a 440, so that the difference wouldn't be too
painfully obvious, would be very difficult. Having copious SIMD resources is
no help.

Since (at least the mill I've been listening to) the 440 is
DEADBEEF, I don't think that's what is being proposed.

Uhm, unallocated memory? Are you saying in a roundabout way that speculation
leading to a 440 core in the Xbox2 took a wrong turn somewhere?

I agree that Microsoft is likely to go for a desktop like programming
environment, where they can reuse technology from their hometurf.

--
Mvh./Regards, Niels Jørgen Kruse, Vanløse, Denmark

I artiklen ,
(Hans de Vries) skrev:

"Niels Jørgen Kruse" wrote in message
...
I artiklen ,
(Hans de Vries) skrev:

(lmurata) wrote in message
. com...
Could Microsoft license Cell technology from IBM for Xbox2? Some have
suggested that Sony's lawyers would never allow it. Others have said
that IBM does not own the technology.


It seems so, It seems that Nintendo will use it as well.
See this press-release he

http://www.forbes.com/home_asia/news...tr1147955.html


" IBM vice president of technology and strategy Irving
Wladawsky-Berger said that the supercomputer used 1,000
microprocessors that are based on PowerPC microchip
technology. The PowerPC chip is currently used in Apple
Computer Inc. computers.

Equating the 440 core with the POWER4 core means that he is talking in
*very* broad terms.

It is also the technology that will be the foundation
of the next generation of gaming consoles from Nintendo
Co. and Sony Corp., which IBM is working on, he said.

This just means that some sort of PPC are in all of these.

He said the chips were less expensive and consumed less
power than traditional microprocessors, making it possible
to pack the same amount of computing power into a smaller
space. Producing the chips in volume for gaming will help
offset the costs of building supercomputers, he said"

Using the same cores in multiple products reduce development costs. The
silicon is *not* the same.

How many different PowerPC's can there be. How much resources can you
waist... One for Apple, one for Sony, one for Nintendo, one for Microsoft,
one for their own Servers...

Why should it be so expensive to modify a working core, particularly if it
is designed to be extensible? IBM doesn't handtweak every circuit.

Which one of the game chips should carry the overhead of the super-
computer interconnect. Or will buyers have a choice between Nintendo,
Play-station and Xbox flavored supercomputers? :^)

Do you think IBM is losing money on their e-servers?

The use of a modified (32 bit) 440 core seems more like a prototyping
vehicle. The original BlueGene design used simple but multithreaded
processors (8 threads per processor) and 32 processors per chip.

Did you read the description of the final item in
http://www.llnl.gov/asci/platforms/bluegenel/pdf/software.pdf?

One can only guess what the final product will be in 2006. I do not think
that either one of the two options above could successfully get Microsoft,
Sony and Nintendo to bet their game-console business on it.

You still think there is a connection beyond the tenuous one of basic ISA.

I do think that a Power6 core could do just that. It should be less then
half the size of today's G5 based on the Power4 core when implemented in
65 nm technology. The Power6 will have a much higher clock speed (by
splitting the pipeline stages in two?) and is likely to support more (4?)
threads.

Do you have any information on the POWER6 core? The dramatic clock increase
projected in that timeframe, I suspect has to do with IBM's HOT technology
which will benefit static (synthesized) logic much more than dynamic logic.


So indeed, Blue Gene/L == Cell now.

(maybe there will be Sony specific APU's although the fact that they
use the same presentation slides seems to suggest otherwise)

I didn't see any PS2/PS3 like 128 bit (4x32) bit SIMD in Blue Gene/L.
Rather 2x64 with two independent 64 bit Floating Point units. However,
it is relatively simple to implement dual 32 bit on those units by
re-using much of the hardware (like the multiplier Wallace trees).
Ideally would be something which is also compatible with Apple's
Altivec. IBM's realizes what mass-production can do I guess.

Blue Gene/L: http://sc-2002.org/paperpdfs/pap.pap207.pdf

Blue Gene/L released right now:


http://groups.google.com/groups?dq=&...f&selm=bp2s7u%
24k3m%241
%40news.rchland.ibm.com

Just imagine what it could mean, with Microsoft now included, If
Windows XP 64 runs not only on Xbox, but also on Playstation 3 !?
or even on Nintendo game consoles?

That would be quite an historical turning point.

You got carried away. Certainly, if the Xbox2 chip is a 440 core plus lots
of SIMD, you can forget about emulating Xbox1 games.

Games spend 90%+ of their time in directX-like libraries. The libraries don't
need emulation and can run in native PPC code.

If you happen to run into a massive lump of 10% code, you will notice. I
don't know how important emulation is, but the current market leader is
backwards compatible. (It is not clear if the PS3 will be.)

--
Mvh./Regards, Niels Jørgen Kruse, Vanløse, Denmark

"Niels Jørgen Kruse" wrote in message
...
I artiklen ,
(Hans de Vries) skrev:

[SNIP - hopefully I got this snip right :/ ]

Games spend 90%+ of their time in directX-like libraries. The libraries
don't
need emulation and can run in native PPC code.

If you happen to run into a massive lump of 10% code, you will notice. I
don't know how important emulation is, but the current market leader is
backwards compatible. (It is not clear if the PS3 will be.)

Storage & memory are cheap(ish)... Surely you could translate the
code statically on the first time the game is played, stash it
away in a persistant cache and then just rely on the translation &
native libraries to yield the required performance.

Cheers,
Rupert

Keith R. Williams wrote:
Since (at least the mill I've been listening to) the 440 is
DEADBEEF, I don't think that's what is being proposed.

Until IBM offers a better replacement embedded CPU option,
qualified on their ASIC fab processes as well as available
as a standalone CPU, the 440 isn't going anywhere.

It probably won't even if a 64-bit version (640? ;-)
intended for embedded applications shows up; sometimes the
extra 1-2 mm^2 matter, and 64-bittedness doesn't.


-george william herbert

Hans de Vries wrote:
How many different PowerPC's can there be.

A lot, though the field seems to be shrinking over time
to fewer models.


-george william herbert

"Niels Jørgen Kruse" wrote in message ...
I artiklen ,
(Hans de Vries) skrev:

How many different PowerPC's can there be. How much resources can you
waist... One for Apple, one for Sony, one for Nintendo, one for Microsoft,
one for their own Servers...

Why should it be so expensive to modify a working core, particularly if it
is designed to be extensible? IBM doesn't handtweak every circuit.


The reason to use a single die for a range of products is almost
always logistics. So in the next three months your customors may need:

Sony 5.5 million processors
Microsoft 2.5 million processors
Nintendo 1.5 million processors
Supercomputers 0.5 million processors

or

Microsoft 4.5 million processors
Nintendo 3.5 million processors
Sony 3.0 million processors
Supercomputers 2.5 million processors

But how do you know? Worse, you had to start production months ago.
A single die that you can use for every product even if the products
are not exactly the same allows you to react faster to the needs of
your customers. It avoids that you're left with inventory you have
to dump. Less inventory means better cash flow et-cetera.

Which one of the game chips should carry the overhead of the super-
computer interconnect. Or will buyers have a choice between Nintendo,
Play-station and Xbox flavored supercomputers? :^)

Do you think IBM is losing money on their e-servers?


I don't know any exact numbers but did see people talking about IBM
micro-electronics losing something like $ 1 Billion last year. More
than enough reason to pay attention to the economics...

The use of a modified (32 bit) 440 core seems more like a prototyping
vehicle. The original BlueGene design used simple but multithreaded
processors (8 threads per processor) and 32 processors per chip.

Did you read the description of the final item in
http://www.llnl.gov/asci/platforms/bluegenel/pdf/software.pdf?


This one describes the same hardwa
http://sc-2002.org/paperpdfs/pap.pap207.pdf

One can only guess what the final product will be in 2006. I do not think
that either one of the two options above could successfully get Microsoft,
Sony and Nintendo to bet their game-console business on it.

You still think there is a connection beyond the tenuous one of basic ISA.


The connection would rather be in the ISA than in the underlying
micro-architecture. So Blue-Gene/L does introduce something 128 bit
SSE2-like to the PowerPC's ISA. A single 32 bit instruction word
executes either a single 64 bit FP op or a dual 2x64 bit SIMD op.

Most x86 SSE2 implementations get the second 64 bit FP for "free".
A highly pipelined 64 bit FP unit generally has a lot of unused
timeslots that can be used to handle the 2nd 64 bit part. That's
actually what makes it tempting to add these strange 2x64 bit
instructions to the ISA. The 440 based version demonstrates that
it's a prototype ISA implementation by actually adding a real
second 64 bit FP unit...

The final game computer version may have a single higher pipelined
64 bit FP which can handle the 4x32 stuff too by re-using parts of
the 64 bit circuits for the dual 32 FP operations (Think Wallace
trees.)


I do think that a Power6 core could do just that. It should be less then
half the size of today's G5 based on the Power4 core when implemented in
65 nm technology. The Power6 will have a much higher clock speed (by
splitting the pipeline stages in two?) and is likely to support more (4?)
threads.

Do you have any information on the POWER6 core? The dramatic clock increase
projected in that timeframe, I suspect has to do with IBM's HOT technology
which will benefit static (synthesized) logic much more than dynamic logic.


I guess everybody is using "ping-pong dynamic logic" now:
Two dynamic circuits for every circuit. One evaluates while the
other is recharged and visa versa in the next cycle.

See for instance Paul DeMone's articles: "The Stuff Dreams Are Made
Of". Dynamic logic is handled at the end of the second article.

http://www.realworldtech.com/page.cf...WT050802020022
http://www.realworldtech.com/page.cf...WT090402005224

These circuits can be synthesized like static logic.


Games spend 90%+ of their time in directX-like libraries. The libraries don't
need emulation and can run in native PPC code.

If you happen to run into a massive lump of 10% code, you will notice. I
don't know how important emulation is, but the current market leader is
backwards compatible. (It is not clear if the PS3 will be.)

It's one of the reasons why it makes sense to use an up to date
high performance core rather than a modified embedded 440.

Regards, Hans.

Hans I guess everybody is using "ping-pong dynamic logic" now

Was that overheard information or are there actually references?

I would be kind of surprised to see dynamic logic circuits used
in an arrangement where the inputs of two gates running on
alternate clock cycles were tied together, and their outputs run
into a static NAND. Is that what you were describing?

I _can_ imagine unrolling a tight pipeline by a factor of two,
in order to cut the clock frequency by a factor of two and thus
give the gates time to recharge. I did that on a DES loop in
1999.

I artiklen ,
(Hans de Vries) skrev:

"Niels Jørgen Kruse" wrote in message
...
I artiklen ,
(Hans de Vries) skrev:

How many different PowerPC's can there be. How much resources can you
waist... One for Apple, one for Sony, one for Nintendo, one for Microsoft,
one for their own Servers...

Why should it be so expensive to modify a working core, particularly if it
is designed to be extensible? IBM doesn't handtweak every circuit.


The reason to use a single die for a range of products is almost
always logistics. So in the next three months your customors may need:

Sony 5.5 million processors
Microsoft 2.5 million processors
Nintendo 1.5 million processors
Supercomputers 0.5 million processors

or

Microsoft 4.5 million processors
Nintendo 3.5 million processors
Sony 3.0 million processors
Supercomputers 2.5 million processors

But how do you know?

Because they tell you? Do you think IBM speculatively manufacture costum
silicon without the intended recipient being committed to buying?

Consoles are manufactured unchanged for years and development paid up front,
so I don't see why a bit of inventory of costum chips is a problem.

--
Mvh./Regards, Niels Jørgen Kruse, Vanløse, Denmark

On Wed, 19 Nov 2003, Niels J[ISO-8859-1] =F8rgen Kruse wrote:

I artiklen ,
(Hans de Vries) skrev:

The reason to use a single die for a range of products is almost
always logistics. So in the next three months your customors may need:

Sony 5.5 million processors
Microsoft 2.5 million processors
Nintendo 1.5 million processors
Supercomputers 0.5 million processors

or

Microsoft 4.5 million processors
Nintendo 3.5 million processors
Sony 3.0 million processors
Supercomputers 2.5 million processors

But how do you know?

Because they tell you? Do you think IBM speculatively manufacture costum
silicon without the intended recipient being committed to buying?

Consoles are manufactured unchanged for years and development paid up fro=
nt,
so I don't see why a bit of inventory of costum chips is a problem.

Console hardware can change over its life time. Take a look at the
original PSX and compare it to the PSone. They play the same games and
the hardware specs are the same but what's under the hood is different.
Sony announced awhile back that they are aiming to merge the Emotion
Engine CPU and the Graphics Synthesiser GPU in the PS2 into one huge
custom chip. I'm not sure if they are actually shipping the EE/GS chip
much less using them in the PS2's currently.

Keith R. Williams wrote:

[snip]

Why? Emulation of x86 on PPC is fairly well know to work.

It's possible to emulate the functionality OK, but very difficult to
emulate the timing. For game code, this can be a big problem.

- Derek