x86-64 a viable choice for Xbox2?

[Vince]

The variant of CELL going into PS3 is ment to have 500 million transistors, using hundreds of Cell processors and thousands of sub-processors. (speculative, but that is what has been reported by different scources

Why do you repeat such BS? This simply isn't true.

Even using common sence and your numbers - you'd realise that even with 1 'thousand' subprocessors on a 500Mtransistor chip, each of these subprocessors would have 1/2 a transistor apiece - and thats ONLY if it's ONE thousand... you imply more. I'm guessing you're not up on the whole tranistor/logic gate stuff....

 

[zidane1strife]

I would expect that GS3 will be roughly equally outclassed by the NV5X as the GS2 is outclassed by the NV2A.

Nv2a doesn't even beat Gs one in all areas let alone the GS2!!!!!

Hmmm.. as previously mentioned there is no gs2... as for the announced performance for the gs2 project .... If it had been released(in early-mid 2002 as planned... hmmm, just a few months after nv2a....) it would eclipse the NV2A, and even the Geforce FX(even thought Geforce fx is a cough 2003 cough cinemati cough gpu...) by an order of magnitude in many areas... since it was to be 100 fold ps2 performance and with good amounts of memory, and targetted to hollywood realtime stuff... It would eclipse even the mighty 1B poly+ GS cube(which still outperforms almost anything outhere...)....

The GS has 7M tranistors dedicated to rasterization only - what do you think they could have done if the lithography of the time (0.25um) wasn't constraining them?

EXACTLY with the manufacturing kinks now ironed out, and with ibm tech, they won't be transistor budget restrained, and neither will they be as handicapped as nvidia is with their outsourced manufacturer... hmmm, geforce.... fx..... hmmm... it could be delayed till april or beyond... or maybe even CANCELLED ALL TOGETHER!!!!

PS: let's talk about some rich delicious nougat!!!!

FALL COMDEX 2002, LAS VEGAS, Nov. 18 /PRNewswire-FirstCall/ -- NVIDIA Corporation (NASDAQ: NVDA) , the worldwide leader in visual processing solutions, today announced that Sony Online Entertainment (SOE), a worldwide leader in massively multiplayer online gaming, has selected NVIDIA graphics processing units (GPU) for the development of EverQuest(R) II, the highly-anticipated sequel to the wildly successful online role-playing phenomenon, EverQuest(R).

Currently in development at SOE studios, EverQuest II uses a groundbreaking 3D engine designed to take full advantage of the new NVIDIA(R) GeForce(TM) FX GPUs to achieve cinematic-caliber visual effects that bring EverQuest II's 3D world to life in stunning detail, and deliver an unprecedented interactive entertainment experience.

"With EverQuest II we are pushing the boundaries of next-generation role-playing games, to deliver the first cinematic quality interactive entertainment experience," said Scott McDaniel, vice president of marketing, Sony Online Entertainment. "The investment we've made in NVIDIA, and they in us, embodies our shared dedication of better technologies that open the door to new and exciting gaming experiences."

The 3D artists and technologists on the EverQuest II development team rely on the programmability of NVIDIA hardware and software to create rich 3D effects and fluid character animations. The result will be a gorgeously dynamic 3D world inhabited by unique, stunningly real characters.

"EverQuest II is going to deliver an amazing experience," said Bill Rehbock, director of developer relations at NVIDIA. "The talented design team at SOE has a clear vision of what they want to achieve with EverQuest II, from both a technology and gameplay perspective, and we have provided them with the perfect platform for unleashing their creative visions."

As part of this joint effort for the development of EverQuest II, NVIDIA is providing SOE with early versions of new hardware, in addition to software driver updates, training support and testing resources. EverQuest II is due to ship Winter 2003.

Heh, heh seems like nvidia's next gen gpu's in a sony dev.'s hand.... well we can all guarantee sony will just ignore all the specs and nv gphx tech from their own dev teams... and just go on some crazy path to nowhere land..... yeah, right....

If anything i'd expect sony to take notes and use their massive r&d capacity and time, to have many teams focus independently on many of these features... evolving them far beyond what a time constrained team ever could. So that features far more advanced are present in a healthy soup of deliciously BIG transistor budget for the GS3.... remember the professional mode GS3 was planned to be capable of 4000X2000 120fps, with nighphotoreal detail... heck sony might even get some 256bit areas in there...

 

[zurich]

Is there some law of physics that I don't know about that's going to make this a forgon conclusion (and is such necessarily bad)?

I didn't say it was bad at all. I said that if Sony is PRing its CPU like mad, then there has to be a reason, reason being that they're following a programmable EE-like design again. I've actually always thought that was much cooler

Hmmm... The NV2A is grossly parrallel, must to a total bitch to program for...

Don't you at me! You've said many times that one of the downsides to the PS2's architecture was that all of its resources were exposed and put into the hands of the developers. SCE backs the HLSL initiative for graphics, but have they made any promises about restricting data flow management? If not, and in a chip as complex as Cell is supposed to be, wouldn't this be very very bad?

 

[zidane1strife]

I said that if Sony is PRing its CPU like mad, then there has to be a reason, reason being that they're following a programmable EE-like design again. I've actually always thought that was much cooler

Yup, and if the GS3 is feature rich(pixel, texture, lighting, etc...), there's nothing to worry about... they just need some nasty busses to connect EE3/Cell based cpu/etc and GS3... and u have a gphx rendering Monster... combined we could be talking 1B transistors or BEYOND.... thanks to such a large transistor budget, and the massive power and programmability of the cpu... If done right, a half@ssed cpu(intel/amd) paired with a supah Gpu will HAVE AN INSANELY TOUGH TIME competing...

 

[archie4oz]

I didn't say it was bad at all. I said that if Sony is PRing its CPU like mad, then there has to be a reason, reason being that they're following a programmable EE-like design again. I've actually always thought that was much cooler

Oh I won't complain about the programmability of the EE. Sony's pimping CPU capabilities because games are still heavily CPU bound...

Don't you at me! You've said many times that one of the downsides to the PS2's architecture was that all of its resources were exposed and put into the hands of the developers. SCE backs the HLSL initiative for graphics, but have they made any promises about restricting data flow management? If not, and in a chip as complex as Cell is supposed to be, wouldn't this be very very bad?

Hehehe... (consider chain pulled)

I state that it can be a problem for some developers, and offers interesting approaches to others... I really don't consider it a downside however since it allows a system with relatively meager transistor logic deliver pretty impressive performance (in comparison to it's competitors and it's age)...

As for a HLSLs, they've been around quite a long time, and for real-time practices, they aren't worth the paper the spec is written on without a good profile target. Whether or not the PS3 has it's resources as exposed as the PS2 matters little. It's the state and performance of the developer tools that play the critical role. Xbox benefits in that aspect since Intel and MS have been developing x86 compilers what seems like forever. The only compilers I'd put on the same pedestal would be SGI's MIPSpro and IBM's Visual Age, neither of which the PS2 benefits from... That's just a small example. It's kind've difficult to say what state the PS3's devkit will be when the time comes. (I *do* know that Sony CSL has been working on similar problems that Cell proposes from the software aspect for quite some time, so they may have something up their sleeves).

 

[PiNkY]

zurich:

But half the strength of the EE is its programmability, so that's not really fair

point is though that you still have to do transform ops, which will normally occupy most of the processing time of (at least) vu1, so in that regard having a fixed unit isn't exactly a bad thing for this generation (lots of static geometry).

You forget that the PS2 needed that leg up in CPU power because the EE handled all the transformation & lighting, while the GS was just a rasterizer. If Sony is beating its chest about the CPU power in the PS3, one can only assume that again, the T&L logic will be decoupled from the graphics chip.

I agree, Sony will probably put most of the work on the cell chip. PC GFX hardware (the stuff that will more or less show up in next xbox & gamecube) also seem to evolve in the direction of large arrays of vector & scalar processors (see P10 & GFFX Vertex Shader and the diminishing differences between Pixel & Vertex Processors in DX10)

 

[megadrive0088]

Yup, and if the GS3 is feature rich(pixel, texture, lighting, etc...), there's nothing to worry about... they just need some nasty busses to connect EE3/Cell based cpu/etc and GS3... and u have a gphx rendering Monster... combined we could be talking 1B transistors or BEYOND.... thanks to such a large transistor budget, and the massive power and programmability of the cpu... If done right, a half@ssed cpu(intel/amd) paired with a supah Gpu will HAVE AN INSANELY TOUGH TIME competing...

I totally agree with this, and other comments like it.

Also, going back a few posts, indeed, the GS2 was ment to be out in 2002
and, with EE2, it would have 100x the graphics power of PS2. EE2+GS2 would likely slaughter the GeForceFX and even the future NV35 in terms of poly performance, fillrate and bandwidth.

The GS3 + EE3 was slated for 2006 or there about, and feature 1000x PS2's graphical output. I could see GS3 alone being 1B transistors, or beyond, not to mention 1/2 billion for EE3/PS3 CPU.

If it's true that GS in PS2 only has 7M transistors dedicated to rasterizing, with the other 36M (or so) taken up by the 4MB of eDRAM--Can you imagine what SCEA could do with a few hundred million? depending on how much eDRAM is in GS3, it could easily have 1B transistors for eDRAM alone. the other few hundred million would be for actual rendering.

One more possibility, aside from EE3/CELL CPU, there could be another CELL dedicated to T&L for GS3, perhaps bolted on to it. Then the EE3 could be solely used for CPU duties, not having the added strain of geometry and lighting.

You forget that the PS2 needed that leg up in CPU power because the EE handled all the transformation & lighting, while the GS was just a rasterizer. If Sony is beating its chest about the CPU power in the PS3, one can only assume that again, the T&L logic will be decoupled from the graphics chip.

I didn't forget that PS2's graphics chip, the Graphics Synthesizer, is just a rasterizer. I know it's composed of 16 extremely simple (Voodo1 like) pixel rendering pipelines coupled with 4MB eDRAM and a massive bus for 48 GB/sec bandwidth (pretty much unparalleled) yet it lacks T&L... meaning a geometry & lighting processor that GameCube's Flipper has. or that NV2A has (the twin Vertex Shader array) Therefore the EE has to provide the GS with all the polygon and lighting transforms. Sony should have bolted a few VUs onto the GS to unburden the EE. Perhaps Sony will use an additional CELL to act as the T&L unit for the GS3, if GS3 is just another rasterizer.

PS2 still has more programmable CPU resources than XBox. I think Gekko is an exellent design though, and it has the advantage of not only not having to provide Flipper with T&L, but also it has 256K cache, more than EE and XBox CPU put together, AFAIK.

 

[Gubbi]

The PS2's r5900 is a dual issue in-order execution architecture with small lvl 1 caches and NO lvl 2 cache. As a CPU core it is clearly inferior to both the PPC in GC and the XCPU. The only reason it is able to perform at all is thanks to black magic exercised by PS2 developers, pushing as much heavy duty math work to the VUs (in micromode), using alot of effort maintaining caches etc.

People here tend to count the VUs both as CPU resources and as T&L resources. True, that they can be viewed as both, but not at the same time.

It is fairly clear what direction the XBox 2 will take as it *does* leverage PC architecture: A fairly fast host CPU to do I/O, control (physics etc.) with subsystems for all the computationally heavy tasks. The graphics system of the XBox 2 will probably be DX 10 style hardware ie. an array of VLIW SIMD processors (shaders). These shaders will be more or less general purpose (ie. almost Turing complete) with conditional branches (but with no effort spent doing branch prediction) and they will be stream oriented/optimized.

I'll bet that Sony will go down the same path (fast general purpose host CPU and arrays of more specialised processors).

Since they will have similar cost constraints, transistor budgets will be more or less the same and hence performance will be on par. Technologically it will come down to the tool chain, here M$ has the edge. Commercially it will come down to marketing and here SONY has the edge,

All IMO.

Cheers
Gubbi

 

[megadrive0088]

Gubbi ~ interesting take on XBox 2 and PS3.

 

[zidane1strife]

Since they will have similar cost constraints, transistor budgets will be more or less the same and hence performance will be on par.

That is true, but u have to remember that intel cpus don't use all their transistor budget, and as with the xbox MS could throw out a processor that is outdated by the time xbox2 comes out, and on top of that downgraded... so u have only one chip on the MS platform competing against a 2 chip combo... and we've all seen the massive amounts of transistors sony can cram, and how they're willing to take a HUGE loss... ps2 was selling initially at 380$ in japan, and at a HUGE loss...

The nvgpu for 2005 i've heard is to be 300M transistor... having to do all the work, a good portion of that goes to many things the EE3/Cell/etc will do(with 500M transistors), this allows far more space for the gs to have massive pixel, bandwith, fillrate, and embedded memory, etc... That is why i see it being difficult for nvidia to compete, they'll be going against a GPU which basically has it's resources split in two processors... and after two consecutive successes, sony might even take a 300$ loss...

EDIT:

Let's not forget that part of the loss sony had was due to the cost of building two new fabs... at a hefty tags... having low yield, due to manufacturing probs... and in the US having to pay for air delivery due to said manufacturing woes....

This time sony is more confident(after two consecutive gen.s being number 1#)... There won't likely be any new costs due to new fabs/air delivery/bad manufacturing tech, neither will sony be hindered by non-state of the art manufacturing tech.... In other words, all of the expenses that were part of their previous(and first real R&D supported console), can now go to PURE R&D or launching at a loss... and after two gen.s at number one, and with ms launching with them... sony is likely to take some nasty losses to guarantee their edge... and in the h/w side the h/w won't be handicapped by the manufacturing techniques as ps2 h/w was....

 

[Saem]

Sony's transistor count advantage is dubious.

x86 has R&D like you wouldn't believe (Why do you think other architectures struggle against x86?) and not just in hardware, but in the software end as well.

Intel's new version of Hyper Threading will be available in the Prescott. Chances are the FPU will be improved, especially the SSE2 units throughput (I suspect it'll be 1 instruction per cycle rather than every other). The CPU will boast a lot of bandwidth and cache. Not to mention Intel's process technology is is about as good as it gets.

While, AMD will likely be offering a very nice Hammer solution.

I believe most people are seriously deluded with CELLs promises for disgusting amounts of computing power. I think it'll definately be advantagous, just not to what's being claimed. IBM's process technology is kinda poor, consider the cost to results ratio (see transmeta).

What it should work out to be is the following, Sony will have more computing power but it will likely be starved in other areas due to cost. Say RAM and most likely their graphics chip. While MS will likely have the software and cost advantage at the least.

Finally, note that MS is no longer in a rush, they'll produce much faster, thus devs will be far better prepared - expect their launch titles to be very good.

 

[BenSkywalker]

Vince-

You do realise that historically, with equal lithography processes - SCE is outpacing nVidia and Intel by a factor of over 4. Thus, if you take away the temporal advantage that MS had designing the XBox, the picture doesn't look so rosey.

Over a factor of four, where exactly do you get this figure from? In late '99 nV had a 23Million transistor consumer part being mass produced, what was Sony's 92Million+ transistor part? Or are you saying that if we completely ignore Sony being behind the curve in fabrication technology then they are four times ahead of the chip that nVidia released in prior years....? The TNT was built on .35u, the GeForce .22u.

Which is due to the extra R&D time between when PS2's design was finished as the XBox's was - according to MS, their won't be such a period this time around. While I know nVidia's part will be a beast, I wouldn't count SCE and their R&D buddies ouyt just yet.

The GS doesn't stack up all that great against the TNT2 in much outside of raw fill. The GS hit in the GeForce time frame.

The GS has 7M tranistors dedicated to rasterization only - what do you think they could have done if the lithography of the time (0.25um) wasn't constraining them?

Would the relatively miniscule amount of transistors it would have taken to add S3TC been worth losing a couple KBs of eDRAM? What about adding in some decent useable texture filtering? How about using a 8x1 or 4x4 pixel pipe arrangement instead of a 16x0? What would they have done is a good question, maybe 6GTexels raw fill while still using CLUT and bilinear?

 

[Gubbi]

Since they will have similar cost constraints, transistor budgets will be more or less the same and hence performance will be on par.

That is true, but u have to remember that intel cpus don't use all their transistor budget, and as with the xbox MS could throw out a processor that is outdated by the time xbox2 comes out, and on top of that downgraded... so u have only one chip on the MS platform competing against a 2 chip combo... and we've all seen the massive amounts of transistors sony can cram, and how they're willing to take a HUGE loss... ps2 was selling initially at 380$ in japan, and at a HUGE loss...

I never said XB 2 would be a single chip design, in fact I argued that it will NOT be. Does Sony make their own chips? I thought Toshiba and NEC took care of that. Anyway, Intel and AMD are king of the hill in process technology, which is also reflected in performance figures:

SpecFP
SpecINT

Athlon XP/MP showing it's age here, but Hammer is supposed to really up the ante, even in ia32 mode.

On top of that they are mass market items, hence: cheap.

So it boils down to the price/performance of the GPU and I don't see the Japanese foundries being much ahead of the Taiwanese.

And last: While Sony is a big company I seriously doubt they have deeper pockets than Microsoft.

Cheers
Gubbi

 

[Tagrineth]

Uhm... to all the people talking about 1-billion transistor chips and such...

WTF? Do you have ANY IDEA how abysmal the yields will be with that many transistors? Forget about it! The potential losses Sony would incur are awful.

Even on 90nm, 1 billion transistors is INSANE. Look at NV30, that's "only" 125 Million (one eighth of one billion), and it's already pretty big at 130nm.

I actually doubt even Intel's highly advanced processes could pull it off with a remotely reasonable yield.

Seriously, people, Sony may be willing to incur losses, but they aren't stupid...

To be honest, I think that first off, all this extra power will bring in somewhat diminishing returns compared to last gen's hyper-charge.

Second, how do any of you come to the conclusion that nothing can compete with the impossibly incomparable UBER-might of five hundred CELL processors? There are a few obvious drawbacks, POWER CONSUMPTION being one of them ("Hey, what the hell? Our power bill just tripled..."). Then there's the serious fact that CELL is fully unproven technology... have any of you seen a sample CELL running anything? Do they even EXIST yet?

And finally, what do you think people were saying about the PS2 before it came out?! Everyone was cheering 'omg nothing can compare to that' and 'nintendorks are screwed hahaha' and stuff like that, and oh shock horror! PS2 is the relative WEAKEST console this gen.

I say we wait until more information seeps through the cracks before we start making such ludicrous predictions.

 

[PC-Engine]

On the console sector, NEC only makes chips for Nintendo and SEGA. You might find little ICs from NEC on many electronic devices though including SONY consoles.

You won't be seeing any difference on the television screen between PS3 and Xbox2 anyway so why argue over paper specs?

 

[Fafalada]

Sony should have bolted a few VUs onto the GS to unburden the EE.

EE's VU1 function is to perform T&L. "unburdening" it would be like adding another T&L processor to NV2a to "unburden" vertex shaders.
Not that I would complain if we had more then 2 VU's on the machine

People here tend to count the VUs both as CPU resources and as T&L resources. True, that they can be viewed as both, but not at the same time.

Actually I was more under impression people view EE as a single resource around the net. Apparently if VU1 executes a vector transform, all other units in EE have to wait for it - after all it's a single CPU right?
On a more serious note, the VUs ARE both CPU and T&L resource at the same time - that's why there's two of them, and why they don't share identical memory paths and what not.

A fairly fast host CPU to do I/O, control (physics etc.) with subsystems for all the computationally heavy tasks.

I would rather see something else then host CPU perform physics if possible, unless it was REALLY fast. It very much falls under the computationally very heavy tasks, and come next gen I would like to see much less need for cutting shortcuts all over the place to fit a decent dynamics model into a game. (yeah ok, so I'm bitter about performance issues thanks to our physics programmer :\ )


Edit:

The GS doesn't stack up all that great against the TNT2 in much outside of raw fill. The GS hit in the GeForce time frame.

In what way? Outside 32bit color, TNT2 was a step back compared even to V2 in some ways. (featurewise)

How about using a 8x1 or 4x4 pixel pipe arrangement instead of a 16x0?

16x0? You lost me here...

 

[Gubbi]

People here tend to count the VUs both as CPU resources and as T&L resources. True, that they can be viewed as both, but not at the same time.

Actually I was more under impression people view EE as a single resource around the net. Apparently if VU1 executes a vector transform, all other units in EE have to wait for it - after all it's a single CPU right?
On a more serious note, the VUs ARE both CPU and T&L resource at the same time - that's why there's two of them, and why they don't share identical memory paths and what not.

I should have been more precise then. I made the statement in objection to the statement that the CPU in PS 2 is the most powerful. What is the CPU in a PS 2 anyway? the r5900 core, the r5900 + VU0 in macromode or r5900+ VU0/1 ?

It all depends on the breakdown of your game, right ? Trouble is comparing individual components in a system, where in one, GC or XBox, host CPU and GPU are 2 clearly distinct entities, while in PS 2 the VUs are shared between the host cpu and the graphics subsystem. If one count the FP power of the EE (one or two VUs) as a CPU resource one really shouldn't count them when stating the (peak) T&L performance, right ?

A fairly fast host CPU to do I/O, control (physics etc.) with subsystems for all the computationally heavy tasks.

I would rather see something else then host CPU perform physics if possible, unless it was REALLY fast. It very much falls under the computationally very heavy tasks, and come next gen I would like to see much less need for cutting shortcuts all over the place to fit a decent dynamics model into a game. (yeah ok, so I'm bitter about performance issues thanks to our physics programmer :\ )

Agreed, but in 2005 a 3GHz PC mpu will be close to a commodity, with SSE that's 12GFLOPS available for physics and other stuff. Chances are that you'll be bandwidth limited.

Cheers
Gubbi

 

[Fafalada]

If one count the FP power of the EE (one or two VUs) as a CPU resource one really shouldn't count them when stating the (peak) T&L performance, right ?

Granted, although I would note that VU1 is pretty much a fixed graphic subsytem resource. Letting the CPU use it would require you to use debug data paths for writeback, which strikes me as rather selfdefeating for purposes of performance

 

[ERP]

In what way? Outside 32bit color, TNT2 was a step back compared even to V2 in some ways. (featurewise)

OK I give in what feature did V2 have that TNT2 didn't?
Neither are particular feature complete GPU's, and I'm finding it tricky to come up with a differentiating feature in voodoo's favor, other than the voodoo rasterizer had more bits of subpixel accuracy 8 vs 4 I believe.

 

[archie4oz]

Does Sony make their own chips? I thought Toshiba and NEC took care of that.

Yes Sony fabs their own chips, the EE at the Oita fab (a Sony/Toshiba joint op, run by Sony Semiconductor), and the GS at Nagasaki Fabs 1 and 2 (SCEI fabs).

Anyway, Intel and AMD are king of the hill in process technology, which is also reflected in performance figures:

Process has little to do with SPEC performance. Cache/memory subsystem, and compiler/math libs play a *far* larger role...

On top of that they are mass market items, hence: cheap.

It also means they're cost reduced parts, hence: it's a bitch to reduce in price in significant quantities after the initial contract. It also negates pretty much any benefit of any VLSI integration to cut system part count.

So it boils down to the price/performance of the GPU and I don't see the Japanese foundries being much ahead of the Taiwanese.

Actually they are. In fact most of the process advancements migrate down to Taiwan from Japan...

And last: While Sony is a big company I seriously doubt they have deeper pockets than Microsoft.

Actually both are publicly traded companies, so it's really not their money to burn, but their shareholders... Shareholders usually don't spending wars (especially since Microsoft stock doesn't pay dividends)...

Even on 90nm, 1 billion transistors is INSANE. Look at NV30, that's "only" 125 Million (one eighth of one billion), and it's already pretty big at 130nm.

Is it really that insane? The I-32 was over 280million transistors at .18um... A Power4 in the same ballpark at .13um (and that's 2 cores, 3 L2s, and chip-interconnect logic). An IC with a billion transistors at 90nm really doesn't seem all that far-fetched... Plus you have to keep in mind that 90nm is right around the corner (Intel, IBM and Toshiba will be sampling devices at various stages on that design rule early next year), you have to start thinking about 70nm and 65-55nm...

Agreed, but in 2005 a 3GHz PC mpu will be close to a commodity, with SSE that's 12GFLOPS available for physics and other stuff. Chances are that you'll be bandwidth limited.

Also, blasting operands through SSE also means you're not doing anything else on the chip...