Intel Larabee

[Klaim]

We talked a bit about Intel Larabee architecture previously, and it seems to be back on the stage for news :

An article about it's instruction set - I myself don't understand all about that because I'm not (yet) into assembly languages but I understand the concepts.

More : http://www.ddj.com/go-parallel/blog/arc ... new_i.html

We should keep an eye on this architecture.

[Kojack]

Yay! I've been waiting for them to release the presentations to the public since they took place a week ago.

*Goes off to read*

[_tommo_]

Definitely keeping an eye on that

I was experimenting the idea of a software-on-GPGPU rasterizer for a while now, and it makes even more sense with this
The problem with CUDA is that you can't use Rasterizer units to do the what they should do, instead Larrabees look much more optimized from that side... even if they lack rasterizers completely.
Now i want to get one of those

[Kojack]

Ok, that vector unit is damn powerful!
Every Larrabee core has 4 hardware threads, and each thread has it's own set of 32 vector registers. Each vector register stores 16 floats, or 8 doubles, or 16 ints, or 8 int64s. So you can fit a whole 4x4 matrix in a single register.

Then it's got the full range of maths operations on registers, including comparisons. It's also got 16 bit mask registers which can enable or disable outputting the result of an operation to individual elements of the destination register. Comparisons set the mask registers, so you can do something like only write elements with positive values and ignore the negatives (leaving the destination's element unchanged).

Loading and saving data to these registers can perform automatic conversions from and to all common directx/opengl pixel formats, and you can even use one vector register to hold 32bit offsets from a base register so each element comes from anywhere within a 4GB range instead of all 16 being contiguous. It can expand and compress during loads and stores too, so masking out 10 elements means it writes the remaining 6 (no matter where they are in the register) into contiguous memory.

It's going to take a while to get used to, but it looks damn cool. The vector unit alone is pretty awesome, but combine that with a hardware texture filter and somewhere from 16-32 cores and it's going to be an interesting graphics card.

Or for a better example of just how general purpose these cores are, Michael Abrash says you can run BSD unix on a core, while your pc is running windows.

Will it beat ATI or Nvidia? Well probably not for standard directx or opengl apps, but it's the potential to do different rendering styles or general purpose computing without the limits of gpu languages like opencl or cuda that make it so cool. I wouldn't mind having a geforce for general games and a larrabee as a physics / gpgpu card, like nvidia's tesla.

[_tommo_]

Seconded.

if they whole thing is powerful as they say, i'm afraid that CUDA and OpenCL are in serious trouble.
obviously Intel won't wipe out AMD and Nvidia from a day to the other, but they will become the ones following...
I think that now their objective is to place a Larrabee into a next gen console, so that any developer will be forced to learn its architecture!

Anyway from a coder standpoint the thing is damn interesting and i almost hope that Intel wins also this one

[xavier]

I think that now their objective is to place a Larrabee into a next gen console, so that any developer will be forced to learn its architecture!

Talking to someone who works for Intel, they are strongly pursuing the PS4 for Intel hardware.

[_tommo_]

I heard of that rumor but it sounds quite unlikely... Larrabee is a "wildcard" for costs and architecture novelty (and performances), just like the Cell in this generation... and we all know how it went for sony.
plus, PS4 that way could be the first computer ever to have a vectorial CPU and an x86 GPU

[xavier]

Well, from Intel employee to my ear, makes it not a rumour. It's just Intel throwing their hat into the console hardware ring, nothing more. There's also nothing there saying "Intel is offering Larrabee as-is" -- chances are, whatever Intel wants to propose, it's something (a) completely not intended for general public release, and (b) doesn't even exist yet in hardware.

[jjp]

I don't think Larrabee will go into a console because of two reasons. I heavily doubt the first iteration of Larrabee will be as efficient as solutions from AMD or NVidia in terms of transistors/performance and power/performance. And, even more important, I doubt Intel wants to sell Larrabee IP to a console manufacturer. Which is very important to a console manufacturer from a cost perspective. We will see

Apart from that: who knows if the successors of PS3 and XBox360 actually will have hardware that goes a major step further in terms of performance? But this is going OT.

[_tommo_]

Well, thinking better at this, Larrabee has a chance to make its way to the console market...
Intel could be actually paying console manufacturers to have their Larrabee in a next generation machine... this because it would be a major step towards market acceptance.
I mean, consoles would be the ideal environment for Larrabee, because they are a fixed platform that you can tune to every fine detail... so you can adapt to its programming model and get all the possible power from it.
A ground-up software render would make much more sense on a console, for example.
Instead on PC Larrabee would be forced to be a DX emulator, because its market share will be certainly too small to target it directly for most software houses.

So in fact, as Xavier said, the costs could be not a problem, because Intel could even sell its processor at a cost lowest than its production price, at least for the first period.
IMHO it's possible, as they would cut by many years the "natural" acceptance time, like sony did with Cell and Blu-Ray (again, with questionable results).
Also, probabily power won't be a major point in next generation, so huge flexibilty could be >>>> than 20% more power.

Still, i don't know if Sony has the funds to give Intel this chance, given its current situation... maybe MS is a better candidate?
Or maybe wii2 will have it... think at the surprise

[jjp]

Also, probabily power won't be a major point in next generation, so huge flexibilty could be >>>> than 20% more power.

Why the hell wouldn't it?

Edit: at first I thought you mean power consumption. But if you mean rendering power then this is a point where Larrabee looks good. But because of the IP issues I don't see it happen. My guess is we'll see a chip based on AMD or NVidia design in the next consoles which will look rather slim in terms of transistor count and power consumption compared to Larrabee.

[xavier]

I doubt Intel wants to sell Larrabee IP to a console manufacturer

On the contrary, Intel sees the volume of console sales over the past N years and wants a piece of that action.

Bear in mind, that IBM makes the PowerPC and the Cell -- processors used in the 3 main consoles right now. However, while you can buy a Cell (actually, have a system made around it) or buy a PowerPC, you *can't* buy the ones in the consoles -- the specific designs are not available to the public. The same would be true of a Larrabee-based console system -- you can certainly buy Larrabee systems, but you won't be able to buy the one in the console.

[xavier]

Or maybe wii2 will have it... think at the surprise

Funny thing is, the Wii demonstrated something that Sony (and to a lesser extent, Microsoft) both completely missed -- that console gamers want enjoyable, fun games.

By way of illustration, here's a (theoretical) conversation with a Sony console sales rep:

Sony Rep (SR): "Just look at the incredible power and technology available in the PS3 -- who *wouldn't* want to buy one (if not more) of these?"
Average Joe (AJ): "Hmm. What kind of games can you play on it?"
SR: "Games that push the furthest boundaries of technology!"
AJ: "Sure, but are they any fun to play?"
SR: "But just consider the incredibly advanced technology you get in the PlayStation 3!"

Actually, that may not be so much theoretical, as it is a summation of Sony's approach to the PS3 for the past three years...

Point being, Larrabee may be great technology, but if Sony still doesn't understand why to this day you can't find a Wii to buy, while Sony can't even give away PS3's, then Intel could very well be pursuing a ticket on the Titanic by courting Sony...

[jjp]

I doubt Intel wants to sell Larrabee IP to a console manufacturer

Bear in mind, that IBM makes the PowerPC and the Cell -- processors used in the 3 main consoles right now. However, while you can buy a Cell (actually, have a system made around it) or buy a PowerPC, you *can't* buy the ones in the consoles -- the specific designs are not available to the public.

Microsoft and Sony have bought the design of the CPUs their consoles use. This is important because it means they control the manufacturing of the chips themselves. I.e. they (among other things) can definitely reduce their costs as manufacturing these chips gets cheaper over time. If they didn't have the IP they'd have to buy chips from IBM. A monopoly situation. That is one of the mistakes Microsoft has made with XBox 1, for example.

[_tommo_]

IMHO Intel, to gain the market control it needs, could also accept to share the IPs of a recent technology such as Larrabee...
in fact this would allow them to place millions of larrabee and to shift programming method... while on pc it could be as well a flop.

Anyway i know that Nintendo has got to the point much more than the others in this gen, in fact i bought a wii first.
Too bad that now they are too concerned on crap and showelware to make games you actually want to play
So now i'm not that convinced about "power is not needed", it WAS not needed, now it's an invitation to third (and first) parties to produce crap.

Unfortunately the market now requires cheap waggle, so i think that next gen will be focused more on that... and also because of this Larrabee looks like an overshoot: i mean, you can put an HD4850 in a new console and still have a margin selling it at 250$.
I think we only need to wait and see...

[xavier]

I doubt Intel wants to sell Larrabee IP to a console manufacturer

Bear in mind, that IBM makes the PowerPC and the Cell -- processors used in the 3 main consoles right now. However, while you can buy a Cell (actually, have a system made around it) or buy a PowerPC, you *can't* buy the ones in the consoles -- the specific designs are not available to the public.

Microsoft and Sony have bought the design of the CPUs their consoles use. This is important because it means they control the manufacturing of the chips themselves. I.e. they (among other things) can definitely reduce their costs as manufacturing these chips gets cheaper over time. If they didn't have the IP they'd have to buy chips from IBM. A monopoly situation. That is one of the mistakes Microsoft has made with XBox 1, for example.

I hope you are not suggesting that Microsoft, Sony and Nintendo are all operating their own fabs -- that would be insanity. They are all still buying the chips from IBM.

The original XBox used off-the-shelf parts in order to minimize time-to-market.

[jjp]

I hope you are not suggesting that Microsoft, Sony and Nintendo are all operating their own fabs -- that would be insanity. They are all still buying the chips from IBM.

Of course they don't have to own fabs. But they can choose where to have their stuff manufactured. And: no, they are not buying chips of IBM. For example the current XBox 360 CPU is manufactured at Chartered Semiconductor (while the old 90nm version has been manufactured at IBM).

[xavier]

Fair enough. The original point (which was in answer to the question raised about someone using the consumer variant of the Larrabee architecture) is that Intel can do the same thing with their latest processor technology -- create domain-specific versions for a single customer.

[jjp]

Yes, but I doubt they want to do that Then again it would be a new situation because for Larrabee most of the interesting IP won't be the hardware but the driver software. But as I already said there is another heavy argument that speaks against seing Larrabee inside a console: it is very unlikely that it will be as transistor-efficient and power-efficient as solutions offered by NVidia or AMD.

[volca]

Maybe it could be. Some say Larrabee will drain about 200-300 Watts of power, right - but if one core would drain 3 watts of power as Atom does, and some surplus, it could also be around 100-150 watts for 32 core larrabee (just speculating here, but it seems logical to me). And you'd get a GPU/CPU hybrid this way - the console could work without any aditional CPU (as they say you can run a BSD kernel on a core of larrabee).

[_tommo_]

Turns out that who says that Larrabee will be a power drain could be right:
http://www.tomshw.it/articles/20090410/ ... -die_c.jpg
Apart from the sleepy guy, that should be a Larrabee wafer... and they said that a single die has an area of 600 mm2!

That's much more than the current monolithic design by Nvidia, the GTX 295... that isn't exaclty cost and energy efficient.

[jjp]

Maybe it could be. Some say Larrabee will drain about 200-300 Watts of power, right - but if one core would drain 3 watts of power as Atom does, and some surplus, it could also be around 100-150 watts for 32 core larrabee (just speculating here, but it seems logical to me). And you'd get a GPU/CPU hybrid this way - the console could work without any aditional CPU (as they say you can run a BSD kernel on a core of larrabee).

Using Larrabee as a cpu replacement would be stupid, you'd get even less sequential processing power than with an Intel Atom. And Atom is not much faster than the CPU used inside the Nintendo Wii (Atom @1.6Ghz approximatly equals a Pentium 3 Celeron @933Mhz).

You have to consider what kind of a traditional GPU can be built with the transistor & power budget of Larrabee. Larrabee's strong point is it's flexibility. It is an interesting architecture for any kind of highly data-parallel computations. But it doesn't even have fixed function logic for rasterization. Which means it will have to spend quite a bit of it's processing power on stuff a GPU does at no cost with it's fixed function logic. Whatever they do, Larrabee cannot be competitive in terms of power/performance and transistors/performance (when we look at traditional 3d realtime rendering).

[volca]

A fair point.

Wait a moment - do you mean it will have no texture units or just missing fixed function pipeline? I thought it will have texture units as dedicated hardware. Fixed function pipeline is dead, so I don't see a point in trying to implement it, it can be emulated through shader code anyway. Texture units are a must, though. Today's games don't use fixed function any more usually, so I don't see that as a problem.

[jjp]

It does have texture units but nothing else. Apart from the vertex/pixel/geometry-shader everything in the graphics pipeline still is fixed function GPUs implement these steps in hardware. And they do it for a reason.

[Klaim]

Ok now we can read something about rasterizing on Larabee.

Reading now.