Larrabee, console tech edition; analysis and competing architectures

[Squilliam]

Carl, you're right! I think I got a bit ahead of myself and probably the whole world. Im sitting at home waiting so I have far too much time on my hands.

I think I should leave my speculation where people in the know don't have to see it!!

 

[liolio]

I was thinking that way for quiet sometime (see the predict next-gen thread).

But people here have managed to change my mind.

On the technological side of thing, it could make a lot of sense for ms:
Intel is likely to able to put more transistors together than anybody and to run them at a higher clock speed.
From the few thing I understand from the larrabee thread is that larrabee could be OK, if Intel is not crazy enough to don't use dedicated hardware for some tasks (texturing,etc.).
As for the cell MS could go with some nehalem like cores and a lot of larrabee cores and use two chips.

But Ms would be completely dependent on Intel pricing...
Ok they could buy two chips per system and reach quickly high volume but even if Intel is accomodating I could see MS show cold feet after the first xbox disaster.

 

[Acert93]

Speaking in the theoretical, I'm wondering given its philosophical similarities to Cell

How so?

 

[Carl B]

How so?

From the very first day Larrabee was floated - and the Tera-scale project before it - the similarities to Cell have been widely noted. Firstly the fact that it follows in the vein of a many-simple core architecture focused on running vectorized code to achieve its massive FP performance, and secondly the fact that in some of its original rumored incarnations it was heterogeneous. The heterogeneous angle seems to be left behind now - perhaps to its detriment - but regardless you could think of the chip as being very philosophically similar to the notion of a bunch of SPEs strapped together. Now, as Barbarian pointed out here the execution of the cores are very different, with Intel playing to their traditional strengths in its design to create something more approachable; shared cache, L1 instead of LS, and an x86-extended ISA (presumably)... but both chips are very much of the "do it in software" school of thought.

Granted Cell isn't targeted directly at GPUs as Larrabee will be, but remember that Larrabee, even though competing there, won't really *be* a GPU as we understand them today, but will very much be a CPU with direct lineage to the x86 ISA targeted towards graphics. Cell 2 could easily follow the same path (though I've already stated I think this unlikely), and as mentioned the original BE patent actually had it in that role. The open-ended questions on Larrabee's side will be if there are any hardware accelerated graphics considerations made in its design, such as for texturing. And right now those questions are open-ended (with different levels of opinion on how much it matters).

Anyone truly interested in the architecture by the way should read the thread linked to in the OP here. Arun's article on the chip is what's started a firestorm of discussion here on B3D recently revolving around Larrabee, and that thread is probably the best source of back-and-forth you could find anywhere on the net in terms of immersing yourself in the arguments on either side.

 

[AzBat]

Wouldn't Intel have to be able to sell an IP design instead of a final chip product? That was the whole reason they went IBM and ATI. They bought the IP for the chips so they could own them and not be dependent on the supplier for the pricing. To me I think Microsoft has a too big of an investment in the IBM and ATI designs just to scrap them and go with Intel, NVIDIA or any other supplier. I'm sure we'll see more advanced versions of the current CPU and GPU configurations. If anything major changes it will most likely be a move to a integrated CPU/GPU type design. A hardware accelerated physics engine might be nice though, but the CPU/GPU may be able to handle that without the need for dedicated silicon. It'll be interesting to see where Microsoft goes with the technology that's for sure.

Tommy McClain

 

[22psi]

I would be very surprised if the next Xbox CPU is not a beefier PowerPC with more core, and fully backwards compatible with Xbox 360.

Current trends say manufactures are starting to care less about BC.

 

[Carl B]

Wouldn't Intel have to be able to sell an IP design instead of a final chip product? That was the whole reason they went IBM and ATI. They bought the IP for the chips so they could own them and not be dependent on the supplier for the pricing. To me I think Microsoft has a too big of an investment in the IBM and ATI designs just to scrap them and go with Intel, NVIDIA or any other supplier. I'm sure we'll see more advanced versions of the current CPU and GPU configurations. If anything major changes it will most likely be a move to a integrated CPU/GPU type design. A hardware accelerated physics engine might be nice though, but the CPU/GPU may be able to handle that without the need for dedicated silicon. It'll be interesting to see where Microsoft goes with the technology that's for sure.

Tommy McClain

Well, remember that MS 'owns' the R500 and Xenon designs specifically, not the architectural family per se. So any upgraded chips along those branches would need to be licensed anew following development work performed by the respective companies. That said, I agree that Intel probably wouldn't license in that manner... but at the same time, fact is for XBox 1 Intel viewed MS as just another OEM customer; this time would be different, as Intel needs MS more than the other way around for Larrabee. So *if* MS went down that road, I would expect a very favorable sourcing contract to land on MS' desk, as opposed to the XBox 1 drama.

I'll add the caveat here that no one's saying Larrabee is MS' future with XBox, this is all just possibility exploration.

Current trends say manufactures are starting to care less about BC.

I don't agree with that... it's just that it's a feature that manufacturers will offer when they can. Sony was simply losing too much money on the hardware and had to lower the price; I have no doubt that they would rather be able to include PS2 B/C if financially trivial to do so. And in that vein, if MS does go the IBM/ATI route for 720, I expect it to be fully B/C, just as I expect the PS4 to be fully B/C if Sony sticks Cell/NVidia. Remember that NVidia wasn't choice number one in terms of GPUs, so it threw things off when they changed.

 

[liolio]

Wouldn't Intel have to be able to sell an IP design instead of a final chip product? That was the whole reason they went IBM and ATI. They bought the IP for the chips so they could own them and not be dependent on the supplier for the pricing. To me I think Microsoft has a too big of an investment in the IBM and ATI designs just to scrap them and go with Intel, NVIDIA or any other supplier. I'm sure we'll see more advanced versions of the current CPU and GPU configurations. If anything major changes it will most likely be a move to a integrated CPU/GPU type design. A hardware accelerated physics engine might be nice though, but the CPU/GPU may be able to handle that without the need for dedicated silicon. It'll be interesting to see where Microsoft goes with the technology that's for sure.

Tommy McClain

Owning the IP in fact wouldn't change anything as Intel is likely to be the only one able to do larrabee, larrabee may be less efficient by design than what other vendors provide but Intel is likely to cram more transistors and to clock thel higher than what others vendors can achieve.
So if Intel doesn't have to let Ms own the IP, but if MS goes with Intel (again) they will have to sign a really good contract!

I don't agree at all with this

To me I think Microsoft has a too big of an investment in the IBM and ATI designs just to scrap them and go with Intel, NVIDIA or any other supplier

I'm not sure Ms has spend that much in RD for the 360.
more xenon may need a lot of work to be a competitive cpu, ie I think that just adding more ppx cores would make it perf wise.

If Intel accept to sell at MS the larrabee at reasonable price (and if obviously the larrabee is a good product), MS will jump in, they will save a lot of money in R&D.

If the larrabee is good but not on par with nvidia or ati offer, it could be a good way for Intel to democratize its hardware ie devs will get theirs hand on hardware.
More Intel is likely to sell havoc to a lot editors

Anyway I can see Intel do some effort to attract MS.

 

[Squilliam]

Thanks for the input guys! Really great stuff, all of you!

I've seen a trend here with people looking at larrabee as the next Xbox GPU. I was wanting to know why they wouldn't use it as the CPU and go with a normal gpu architecture. My inititial idea was that larrabee would be the CPU, basically as competition to the cell and they would still use dedicated graphics hardware. Because my simple understanding has it's best for the job.

Larrabee would still benifit as much as the cell has from developer input. I feel it should be put where the most flexibility is needed as it is an extremely flexible architecture.

 

[Acert93]

From the very first day Larrabee was floated - and the Tera-scale project before it - the similarities to Cell have been widely noted. Firstly the fact that it follows in the vein of a many-simple core architecture focused on running vectorized code to achieve its massive FP performance

I wouldn't argue that it doesn't have similar goals (e.g. significant floating point performance through many core design) but I would content that "philosophical similarities to Cell" is a stretch.

First I think it is fair to note that the many-core high-flops approach isn't new or peculiar to Cell at all. Within just the last couple years alone we have seen products from Ageia, Clearspeed, even IBM with "Waternoose" derivatives like Xenon have gone with an approach of simpler, if not specialized cores, with beefed up vector performance. In fact GPUs have been pretty much the posterchild for this general trend. I think the goal, and the general road thereto, has been pretty clear for a long time.

"Philosophically" I think Larrabee is quite distinct from Cell. You noted the heterogenous/homogenous cores, but it goes a lot deeper than that.

First and foremost is that Larrabee is going the add-in route, instead of a new platform. It seems, as of this moment, Larrabee is going to be a PCIe card which can be utilized in already established infrastructures. (Not true of Cell). Taking it one step further with x86 legacy Intel is leveraging an existing ISA and legacy code (something SPE's didn't attempt to do). Philosophically, as a product, Larrabee is quite different from Cell.

And while the trend is toward "more minimal cores over and against fewer robust cores" Larrabee appears to be more inline with chips like Xenon than with Cell in this regards. Cell is very clean and has a high peak flops/mm2 approach. Based on the little we know about Larrabee it seems to make key concessions here. First is going with x86 and generating necessary performance from new vector units (my guess would be the cleanslate design of the SPEs is more transistor effecient) but more importantly is that Intel appears to be avoiding the scratchpad design of Cell and instead going with managed memories. x86 with managed memories, at the sacrifice of peak performance, appears philosophically quite distinct from Cell with its own SPE ISA, low legacy, and hands on approach to memory utilization. They have overlap in their general goal but are going about it quite differently in terms of architecture (new and minimalist for highest gain versus the pros/cons of legacy) as well as market approach.

And I didn't even get to the specialized graphics hardware. Cell isn't a GPU. You aren't going to toss it into a PC for this task. Larrabee, to some degree, will. It also appears to be leveraged toward workstations as an add-in product.

As Barbarian noted a number of months ago, the ability to run x86 code on Larrabee is a "big deal" and I think pivotol to what Intel's goal and philosophy: get the best performance possible while support some degree of legacy support (through x86, manged caches, etc). Running your "legacy score board update" on Larrabee may be slower than running it on a more robust core -- and definately slower than a custom app for Cell or Larrabee -- but something like this would be such low usage to begin with that you can let a low level engineer be sloppy with it or, a high level designer to gain more time to make the performance intensive code swifter.

But it all comes at a cost, in his case HW. Per mm^2 I wouldn't bet against Cell having twice the peak floating point performance when laid side by side Larrabee.

but regardless you could think of the chip as being very philosophically similar to the notion of a bunch of SPEs strapped together.

So is that how we have come to "philosophically" view Cell? If so, how does Cell even have a footprint on this approach of bundling together a lot of cores?

I was pretty sure I got quite a few lecturings, with little protest from others, how the EIB, SPE non-legacy and notably the LS design and the like were what defined Cell apart from the general theme of "Stream Processors" which is quite a large umbrella of thought.

Now, as Barbarian pointed out here the execution of the cores are very different, with Intel playing to their traditional strengths in its design to create something more approachable; shared cache, L1 instead of LS, and an x86-extended ISA (presumably)... but both chips are very much of the "do it in software" school of thought.

Exactly... they are the same... but completely different. If they are architecturally quite distinct, as you just outlined, and how they fit into the market (platform versus addon card and the like) I think talking about Larrabee being "philosophically" similar to the same as Cell a real stretch.

They have a similar goal: high FP performance.

Generalized (many cores) they are similar, but the similarities stop there. Philosophically one is design with much more legacy and hand holding in view--and a completely different entry into the market (not to mention the use as a graphics adapter!)--and the other is clean slate design being pushed through the pipe as a completely new platform.

I think the philosophical concerns, philosophical approach, and hence the implimentation are quite distinct. Enough to obscure the conceptual similarities at a high level.

Philosophically, I think Larrabee is actually addressing a lot of those areas Cell's design intentionally avoided... all the while actually addressing a different market: GPUs.

Granted Cell isn't targeted directly at GPUs as Larrabee will be, but remember that Larrabee, even though competing there, won't really *be* a GPU as we understand them today, but will very much be a CPU with direct lineage to the x86 ISA targeted towards graphics. Cell 2 could easily follow the same path

Even if that happened, that wouldn't be philosphically like Cell, but Cell2

All semantics and emphasis on terms aside... and no intent to derail the thread...

I think we would agree that Larrabee's general goal is similar to Cell, but with key concessions on both sides in terms of design which play a significant role in terms of the execution of the underlying architecture. Big picture Cell and Larrabee belong to a host of designs in this category, but I cannot shake the reality that Larrabee is targetting the market in a way that reverses the tide of Cell in terms of legacy and is actually taking on the larger, more successful stream processors: GPUs.

 

[Shifty Geezer]

I would say perhaps the place to look if they are philosophically similar is in the approach to handling the bottlenecks. Cell design wasn't just lots of cores for lots of FP, which was a no-brainer design choice for any high-performance CPU, but a memory system to deal with data throughput so those cores don't choke. If Larrabee supports the same principles of interconnected per-core storage with a ring type distribution of data throughout than I'd say the philosophy is the same. If Larrabee just sticks to the existing tiered cache design and just improves the performance of these caches, then the designs are fundamentally different apart from the obvious points that all high-performance parts will share (as you may have noticed, I'm not clued up on Larrabee's architecture!). The fact Larrabee is targeted at a different principle market is neither here nor there IMO. Cell on a PCIe card isn't impossible and Toshiba's SpursEngine does pretty much that already.

That said, if Larrabee is trying to be a GPU and CPU/accelerator, that is a philosophical difference. Cell wasn't designed to be a GPU replacement and has some major shortcomings in that respect, which agrees with your point above. From that POV, ignoring architectures and looking at what the processors are trying to be, which I suppose is the real philosophy behind them and the rest is actually engineering, then, then they are different.

 

[one]

but at the same time, fact is for XBox 1 Intel viewed MS as just another OEM customer; this time would be different, as Intel needs MS more than the other way around for Larrabee.

Can Larrabee run Windows on it? If it can, Intel doesn't need a game console. If a completely homogeneous version of Larrabee is too slow to run Windows, then they can release a heterogeneous version with a few fat IA cores. MMX/SSE is totally redundant on Windows if no apps are programmed to use it, but Larrabee cores in a CPU are not if they can act as slower IA cores. I think mainstream CPU for PC will be like that in 5 years.

 

[V3]

I've seen the propose diagram of Larrabee but anyone can link me up to real info ?

You guys are talking about design philosophy and stuff, where are you guys getting the info from ?

 

[liolio]

I agree with both of you the larrabee has nothing to do with cell from an architecture point of view.

Anyway I could see Intel mix some gemeral purpose cores in larrabee to fit some part of the market.

Larrabee may come with 16 24 and 32 cores, the 16 cores part could come somes general purpose cores (gesher or nehalem I haven't watched the Intel roadmap for some time).

And By the way I have a question, I read some time ago a new here about how fast communication between the cpu and gpu could allow for some interesting improvement in graphic, do you think it would be better to have:
A multi core cpu say 4 cores (nehalem gesher whatever) & a 32 cores larrabee
or
2 identical chips which include 2 gesher/nehalem cores and 16 larrabee cores

I failed to not be exited about the whole larrabee thing.
I remember back in 2003/2004 the excitement was really high about cell!
When I think of so many cores sharing the same ISA with Intel owning Havoc with the man power of both Intel and Ms I'm really exited (even if I don't that much hope) about what the top devs team could came with! No matter the larrabee could slightly underperform in regard to what other vendors provide, the flexibility and cheer power for physic and IA make me drool.

 

[liolio]

I've seen the propose diagram of Larrabee but anyone can link me up to real info ?

You guys are talking about design philosophy and stuff, where are you guys getting the info from ?

You should read that thread:
http://forum.beyond3d.com/showthread.php?t=46393
I think there some interesting links while the discussion is awesome (I fail to understand properly most of the post but it give a general idea of the pro/con of the design (minus what we know about it so far).

 

[Carl B]

I wouldn't argue that it doesn't have similar goals (e.g. significant floating point performance through many core design) but I would content that "philosophical similarities to Cell" is a stretch.

I think you and I have different views on what the word "philosophy" means then.

First I think it is fair to note that the many-core high-flops approach isn't new or peculiar to Cell at all. Within just the last couple years alone we have seen products from Ageia, Clearspeed, even IBM with "Waternoose" derivatives like Xenon have gone with an approach of simpler, if not specialized cores, with beefed up vector performance.

Absolutely; and I would say that they are all philosophically similar, save perhaps for Xenon which exists in a grey area IMO in terms of its shifting towards the 'many simple core' paradigm. In fact, I would say you neglected to mention some of the most relevant participants in the revolution, that being Sun's newer chips.

In fact GPUs have been pretty much the posterchild for this general trend. I think the goal, and the general road thereto, has been pretty clear for a long time.

As GPUs become more general, at the same time there are some processors that are shifting to their massive parallelism/streaming model... and how could anyone deny that Cell has been most prominent among them? Clearspeed is useless in any sort of general environment, and Ageia - I don't even know, I feel sorry for them. Cell is an actual viable chip with widespread (qualified) support across a number of industries.

"Philosophically" I think Larrabee is quite distinct from Cell. You noted the heterogenous/homogenous cores, but it goes a lot deeper than that.

First and foremost is that Larrabee is going the add-in route, instead of a new platform. It seems, as of this moment, Larrabee is going to be a PCIe card which can be utilized in already established infrastructures. (Not true of Cell). Taking it one step further with x86 legacy Intel is leveraging an existing ISA and legacy code (something SPE's didn't attempt to do). Philosophically, as a product, Larrabee is quite different from Cell.

Well again I disagree; the chips try to achieve the same thing, and via same manner... as it's just the implementation that's different - and yes the implementation is very material - but what is left but to say that they share a similar philosophy? Ask a Cell designer what upcoming chip most contends with their own niche and you'll hear that it's Larrabee/Terra-scale out of Intel, and GPGPU on the GPU front. I'm not going to get into it in depth here, but I've spoken to a number of folk at a couple of institutions and they all view Cell, GPGPU, Larrabee, etc... as part of a coming revolution in programming philosophy. Where they each lie on the curve one can debate, but it's the same curve for sure.

And while the trend is toward "more minimal cores over and against fewer robust cores" Larrabee appears to be more inline with chips like Xenon than with Cell in this regards. Cell is very clean and has a high peak flops/mm2 approach. Based on the little we know about Larrabee it seems to make key concessions here. First is going with x86 and generating necessary performance from new vector units (my guess would be the cleanslate design of the SPEs is more transistor effecient) but more importantly is that Intel appears to be avoiding the scratchpad design of Cell and instead going with managed memories. x86 with managed memories, at the sacrifice of peak performance, appears philosophically quite distinct from Cell with its own SPE ISA, low legacy, and hands on approach to memory utilization. They have overlap in their general goal but are going about it quite differently in terms of architecture (new and minimalist for highest gain versus the pros/cons of legacy) as well as market approach.

Well Xenon is three fairly robust cores, whereas Larrabee is anything but... so although the legacy ISA support, cache support, and quasi-homogeneous nature (I say quasi since presumably heterogeneous is planned for future revisions), I'd still say that Intel is making a break for something that moves more towards Cell than say Xenon. Or one could view it as a further radicalized Xenon concept, but things start to blend.

And I didn't even get to the specialized graphics hardware. Cell isn't a GPU. You aren't going to toss it into a PC for this task. Larrabee, to some degree, will. It also appears to be leveraged toward workstations as an add-in product.

Well I want to point out that Cell *is* available as an add-in product, that can be plugged into PCI-e slots, and that can be placed on-board independent of CPU; Roadrunner will use the PCI-e version paired to Opteron boards and Toshiba recently demo'd the SpursEngine working in conjunction with an x86 environment on their laptops.

Beyond that I'll just say that the assumption of specialized graphics considerations - we'll assume texturing units - for now is just that; an assumption for which we have no solid evidence. I personally would expect it, but at the same time since it's being targeted at that market, what's the surprise? I view it as a hallmark of the initial application revision rather than of the architecture itself.

As Barbarian noted a number of months ago, the ability to run x86 code on Larrabee is a "big deal" and I think pivotol to what Intel's goal and philosophy: get the best performance possible while support some degree of legacy support (through x86, manged caches, etc). Running your "legacy score board update" on Larrabee may be slower than running it on a more robust core -- and definately slower than a custom app for Cell or Larrabee -- but something like this would be such low usage to begin with that you can let a low level engineer be sloppy with it or, a high level designer to gain more time to make the performance intensive code swifter.

I agree with you here, but that's a different "philosophy" than the one I'm discussing. That has more to do with the ease-of-approachability goals at Intel rather than anything inherent to the high-level architecture. They could have gone a totally different route on the ISA and everything that makes Larrabee distinct/unique would still be there.

But it all comes at a cost, in his case HW. Per mm^2 I wouldn't bet against Cell having twice the peak floating point performance when laid side by side Larrabee.

Is mm^2 what you want to use as your metric though given Intel's process advantage? On a serious note though, the truth is that's just an arbitrary guess; obviously Cell will increase their density going forward, and perhaps enhance the architecture in general in a future revision, but neither is an FP slouch so I'm not ready to crown either the winner.

So is that how we have come to "philosophically" view Cell? If so, how does Cell even have a footprint on this approach of bundling together a lot of cores?

I was pretty sure I got quite a few lecturings, with little protest from others, how the EIB, SPE non-legacy and notably the LS design and the like were what defined Cell apart from the general theme of "Stream Processors" which is quite a large umbrella of thought.

Well, not sure why anyone would want to define Cell as different than 'stream processors' in general from within that context, but those reasons are why Cell was revolutionary, with that I agree. The SPE non-legacy doesn't have any direct correlation btw, rather simply its achieved performance. If they had done it with legacy or in any other manner, the achievement would not have been diminished. It's simply that a from-scratch ISA yielded the results they wanted in that vein. But Larrabee will share some of those considerations; not the EIB per se, but certainly a wide/fast ring bus and access to a lot of external bandwidth. The LS is a unique-to-Cell situation, definitely. But Larrabee goes for the ame basic goal of basically trying to reduce latency to the lowest level possible.

Exactly... they are the same... but completely different. If they are architecturally quite distinct, as you just outlined, and how they fit into the market (platform versus addon card and the like) I think talking about Larrabee being "philosophically" similar to the same as Cell a real stretch.

Well, on this we'll agree to disagree.

They have a similar goal: high FP performance.

Generalized (many cores) they are similar, but the similarities stop there. Philosophically one is design with much more legacy and hand holding in view--and a completely different entry into the market (not to mention the use as a graphics adapter!)--and the other is clean slate design being pushed through the pipe as a completely new platform.

I think the philosophical concerns, philosophical approach, and hence the implimentation are quite distinct. Enough to obscure the conceptual similarities at a high level.

Hmmm... well, I've stated my case to the contrary already, so needless to say I disagree with your view of their philosophical 'distinction.' I'd go further to the above and add that the only markets we can actually surmise Larrabee will be likely strong in are the same areas in which Cell is strong in; physics, HPC, modeling problems, etc... graphics truth be told is a total unknown. It could completely suck there.

Philosophically, I think Larrabee is actually addressing a lot of those areas Cell's design intentionally avoided... all the while actually addressing a different market: GPUs.

I would almost want to ask at this point what you consider Cell's own philosophy to be; if they put out a chip with rasterization APUs mixed in with the SPEs, would it not still be the same architecture at its core? Certainly in my mind the design premises would still be the same. It's just that not explicitly targeting that market at the moment, such a chip doesn't exist.

Even if that happened, that wouldn't be philosphically like Cell, but Cell2

I think again it goes to the definition of 'philosophy' here. Even knowing nothing about what Cell 2 will be like, I can already proclaim it as being philosophically similar; else it wouldn't even be called "Cell" 2 to begin with.

All semantics and emphasis on terms aside... and no intent to derail the thread...

I think we would agree that Larrabee's general goal is similar to Cell, but with key concessions on both sides in terms of design which play a significant role in terms of the execution of the underlying architecture. Big picture Cell and Larrabee belong to a host of designs in this category, but I cannot shake the reality that Larrabee is targetting the market in a way that reverses the tide of Cell in terms of legacy and is actually taking on the larger, more successful stream processors: GPUs.

Well, I stated similar above and don't disagree; but it's hard to put semantics aside when apparently it's exactly semantics we're embroiled in discussing here.

 

[Carl B]

Can Larrabee run Windows on it? If it can, Intel doesn't need a game console. If a completely homogeneous version of Larrabee is too slow to run Windows, then they can release a heterogeneous version with a few fat IA cores. MMX/SSE is totally redundant on Windows if no apps are programmed to use it, but Larrabee cores in a CPU are not if they can act as slower IA cores. I think mainstream CPU for PC will be like that in 5 years.

I think in terms of breaking into the GPU space, having the architecture form the basis of the next XBox' graphical sub-system would prove invaluable in terms of getting developers familiar with working with the chip in the role of a GPU. I think it's an important distinction, because I don't believe that Intel will just go ahead and tout it as their next general PC basis, even were it to run Windows (which it seems like it should be able to). I think it'll remain exotic/specialized at least for some time upon its introduction, until yes finally they do go the heterogeneous route, but even then...

I just have to hold to the notion that it does a lot for Intel were the chip to be included in the next XBox revision, and insofar as it's actually a chip MS is consdering - and frankly who knows if it is, it might not be at all - I believe Intel will fight to win that contract. Especially considering that it might give them an outsized voice in having future revisions of DirectX go in a direction that naturally favor the strengths of the architecture itself.

 

[Shifty Geezer]

I'd go further to the above and add that the only markets we can actually surmise Larrabee will be likely strong in are the same areas in which Cell is strong in; physics, HPC, modeling problems, etc... graphics truth be told is a total unknown. It could completely suck there.

Keeping this on topic, what exactly is Larabee's position likely to be? Surely in the graphics department it'll be trumped by specialised GPUs, so a Larabee only neXtBox could be an underperformer. Then again perhaps diminishing returns will will make Larabee 'good enough' and provide an exceptional flexibility? This would reflect the situation with Cell2 in PS4, if a single chip solution becomes suitable providing just a lump pool of resources for developers to distribute between game code and graphics as they choose.

Also if Larabee is being offered as a plug-in accelerator, would we be looking at a separate CPU in neXtBox, and if so, would that be a symmetrical multicore system? That could prove a headache - balancing code over multiple instances of multiple cores!

 

[liolio]

It all depends on how far larrabee will land from more specialised parts.
We still don't if larrbee will include texture sampling units as this looks to be the most discused parts in the dedicated larrabee thread.

 

[Carl B]

Keeping this on topic, what exactly is Larabee's position likely to be? Surely in the graphics department it'll be trumped by specialised GPUs, so a Larabee only neXtBox could be an underperformer. Then again perhaps diminishing returns will will make Larabee 'good enough' and provide an exceptional flexibility? This would reflect the situation with Cell2 in PS4, if a single chip solution becomes suitable providing just a lump pool of resources for developers to distribute between game code and graphics as they choose.

Also if Larabee is being offered as a plug-in accelerator, would we be looking at a separate CPU in neXtBox, and if so, would that be a symmetrical multicore system? That could prove a headache - balancing code over multiple instances of multiple cores!

Well, I bolded the above because that's really what it comes down to... people who believe that Intel will fall short on the GPU front and folk that think it has a shot at greatness. But yes indeed though, the context of the discussions/theorizations is just that: Larrabee as the GPU.

So, we'll see. There's a lot of excitement (at the flexibility) and doubt (knowing the strength of the incumbents) on either side of the aisle.

*If* Larrabee was in the next XBox, I think the form it would take would depend on its own development cycle, XBox' release date, and the needs to be filled. Here we're discussing it as a GPU, and in that context I just assume whatever Intel's leading edge OOE core at the time might be for the CPU. But yeah, it could be two heterogeneous Larrabee cores, one homogeneous and one hetero, one massive chip overall... who knows.

There really are so many unknowns right now, not the least of which is whether this is even a possibility, but adding to the engine of speculation of course is the fact that the architecture could theoretically be massaged to fit whatever role XBox required this far out from launch.

Obviously until Larrabee proves itself, MS has to look at the thing as a large gamble to bank on, so honestly I would find a PPC/AMD extension more likely for 720.

For PS4 though, whatever form the Cell chip might take, I do expect dedicated graphics hardware rather than a Cell-based approach. NVidia seems to be a partner Sony is going to stick with for awhile. (Not that things can't change quickly.)