Intel Larrabee discussion thread

[Mr. Bluntman]

There has been a discussion that has spurred up in this thread, on page three. I'm moving it here since it's the proper venue for it.

Have at it boys.

 

[ebeattie]

So.... what is this... Larrabee you speak of?

 

[dirtypop]

From Wikipedia: "Larrabee is the codename for a discrete graphics processing unit (GPU) chip that Intel is developing as a revolutionary successor to its current line of graphics accelerators. The video card containing Larrabee is expected to compete with the GeForce and Radeon lines of video cards from NVIDIA and AMD/ATI respectively. Intel plans to have engineering samples of Larrabee ready by the end of 2008, with public release in late 2009 or 2010."

Link

 

[leSLIe]

intel in the gpu arena, hmm... are they gonna use ray tracing ?

 

[dirtypop]

intel in the gpu arena, hmm... are they gonna use ray tracing ?

Seems that way. Might take a while to catch on though. Might not catch on at all. They say raytracing performance scales up really nicely, but it takes some growing pains to get to that point.

 

[Mr. Bluntman]

Seems that way. Might take a while to catch on though. Might not catch on at all. They say raytracing performance scales up really nicely, but it takes some growing pains to get to that point.

And for some things, rasterization is still more efficient. David Kirk, the chief scientist from NVIDIA said something along the lines that the future of graphics lies somewhere between raster graphics and raytracing, combining the best of both worlds for the most efficiency and lowest computational cost, with the end result being the same as pure raytracing.

Intel wants to push pure raytracing. Maybe in 10-20 years.

Read it here.

 

[Hypernova]

Ray-tracing is as close to real life simulation as it is physically possible. With the rate Intel is stacking up cores it was only a matter of time before they decide to take a jab at this.

Intel's already go the biggest share in the graphics market, they are just using that to it's fullest.

 

[Scali2]

They are NOT going to use raytracing, at least not for realtime graphics. It may be used sparingly in renderfarms though.
But raytracing is too limited a technology for regular games and things. There are various issues relating to common things like texture filtering and skinned animation that cannot be solved efficiently with a raytracer in realtime.

The future will be a hybrid approach with a rasterizer using raytraced effects where applicable. But that is something that not only Larrabee would be capable of.

 

[Scali2]

Ray-tracing is as close to real life simulation as it is physically possible.

This is not true. At least, not for standard Whitted raytracing algorithms.
They do not solve the phenomenon of multipath because they trace from the eye.
Which is why raytracers also need Monte Carlo routines, photon mapping and/or radiosity algorithms to aid the regular raytracing algorithm. None of which are very efficient or suitable for realtime use.

I don't know where the idea of 'raytraced games' comes from, but it is not going to work anytime soon. I spoke to the guys that made that Quake raytracer and working on raytracing hardware, asking how they would solve character animation in realtime... They had no answer. All they can do is precalc some animation frames, which is worse than even the animation of the original Quake. Compare that to the near-cinematic animation quality of HalfLife 2 and Crysis... and raytracing doesn't stand a chance in gaming.

nVidia is right in that article, rasterizing isn't going away anytime soon. Even Pixar movies are still mostly rasterized because it delivers better image quality and smoother animation at lower processing costs. Pixar only uses raytracing for certain special effects, eg a closeup of a crystal glass or so.
nVidia also points out that the acceleration structures are the achilles heel of raytracing. This is why you can't do animation... Every update of an object's shape requires the acceleration structures to be updated, nullifying their gain for realtime purposes (each frame is rendered only once, then the animation status is updated).

 

[Hypernova]

This is not true. At least, not for standard Whitted raytracing algorithms.
They do not solve the phenomenon of multipath because they trace from the eye.
Which is why raytracers also need Monte Carlo routines, photon mapping and/or radiosity algorithms to aid the regular raytracing algorithm. None of which are very efficient or suitable for realtime use.

I don't know where the idea of 'raytraced games' comes from, but it is not going to work anytime soon. I spoke to the guys that made that Quake raytracer and working on raytracing hardware, asking how they would solve character animation in realtime... They had no answer. All they can do is precalc some animation frames, which is worse than even the animation of the original Quake. Compare that to the near-cinematic animation quality of HalfLife 2 and Crysis... and raytracing doesn't stand a chance in gaming.

nVidia is right in that article, rasterizing isn't going away anytime soon. Even Pixar movies are still mostly rasterized because it delivers better image quality and smoother animation at lower processing costs. Pixar only uses raytracing for certain special effects, eg a closeup of a crystal glass or so.
nVidia also points out that the acceleration structures are the achilles heel of raytracing. This is why you can't do animation... Every update of an object's shape requires the acceleration structures to be updated, nullifying their gain for realtime purposes (each frame is rendered only once, then the animation status is updated).

True... true...

But perhaps Intel is banking on making a GPU powerful enough to brute force through a frame in real time.

Now can Intel make such a chip in 2 years? Fat chance.

But one can dream...

 

[LanceDiamond]

Interesting stuff in this interview with the guy that created Unreal about his feeling towards integrated graphics.

http://www.tgdaily.com/content/view/36390/118/

from the end of this interview with him, he says:

Intel’s integrated graphics just don't work. I don't think they will ever work.​

 

[Scali2]

True... true...

But perhaps Intel is banking on making a GPU powerful enough to brute force through a frame in real time.

But so is nVidia.
A few Teslas can do an incredible amount of raytracing if you want.
nVidia even has its own GPU-accelerated renderer by the name of Gelato, which can do raytracing. It's had that since the GeForce 6-series, if I'm not mistaken.
nVidia just doesn't use the marketing buzz of 'raytracing' to promote their products.
Currently, nVidia is still ahead of Intel in this whole area, and I don't think that the first generation of Larrabee will actually be able to compete with nVidia's GPUs. In fact, not even Intel thinks that. They stated somewhere that they hope to compete with mid-end GPUs with Larrabee at first, and move to high-end in the second generation.

 

[Scali2]

Intel’s integrated graphics just don't work. I don't think they will ever work.​

The irony is that Intel's GPU architecture is basically cutting-edge. They have SM4.0-capable unified shaders and hardware-accelerated MPEG4 and that sort of thing.
It's just that they are designed so cheaply that they only have very little shader units (my X3100 only has 8, compare that to the 96 my 8800GTS has), and the biggest problem is Intel's drivers at this point.

I think that if you'd take Intel's current GPU technology, create a fullblown GPU out of it with dedicated graphics ram, and have proper drivers, it'd actually be competitive with AMD and nVidia.
But as ATi can tell you, you don't just fix your drivers in a few weeks. It took years for ATi to really catch up with nVidia in terms of driver quality... It wasn't until the Catalyst driver series that ATi reached the same level of compatibility and performance as nVidia had in D3D, and OpenGL took ATi even longer. Many other brands never even got that far... eg S3 or XGI, who had promising D3D9-hardware on paper, but just wouldn't run most software properly.
Intel is still struggling there aswell. I hope they are wise enough to invest a lot into the development of drivers for Larrabee, because good drivers are at least as important as a good hardware-design.

 

[LanceDiamond]

Come to think of it, why wouldn't Intel just goble up Nvidia? Do they really think they can catch up in some reasonable amount of time instead? I guess this kind of thinking may be one of the many reasons I don't have $10billion to spend - because if I were Intel and had the cash, I'd think "screw it thats like one fab, write up the NVidia sale!"

 

[Hypernova]

Come to think of it, why wouldn't Intel just goble up Nvidia? Do they really think they can catch up in some reasonable amount of time instead? I guess this kind of thinking may be one of the many reasons I don't have $10billion to spend - because if I were Intel and had the cash, I'd think "screw it thats like one fab, write up the NVidia sale!"

If Intel takes out nVidia for the sake of taking out the competition they are going to have a very long tea chat held in court rooms from Anti-trust regulators. Remember that they already got ~50% of the GPU market with nVidia ~25%. They are already under a lot of heat for their dominance over CPU markets from EU and Korea etc. I doubt they want to go through that with GPU's.

Besides, if they truly believe that they can do things better then I say let them try. More choices are always good. Things haven't looked quite the same since the likes of S3 (The first GFX card in our family was a S3 from 1995) got wiped off the main stream GPU map.

 

[veedubfreak]

Intel holds about 38% of the market, with nvidia holding about 35%. But with the profits that Nvidia is pulling down, i cant see Intel even with their budget, being able to come up with the amount of cash that Nvidia would want if they tried to sell.

 

[leSLIe]

But perhaps Intel is banking on making a GPU powerful enough to brute force through a frame in real time.

Now can Intel make such a chip in 2 years? Fat chance.

well that is indeed unlikely.
Intel has been researching in photonics CPUs instead of the normal electronics, with their new silicon lasers. Who knows maybe we're gonna see a photonic GPU in 5 years

 

[Mr. Bluntman]

well that is indeed unlikely.
Intel has been researching in photonics CPUs instead of the normal electronics, with their new silicon lasers. Who knows maybe we're gonna see a photonic GPU in 5 years

Or carbon nanotube transistor/interconnect based ASICs.

 

[defaultluser]

The irony is that Intel's GPU architecture is basically cutting-edge. They have SM4.0-capable unified shaders and hardware-accelerated MPEG4 and that sort of thing.
It's just that they are designed so cheaply that they only have very little shader units (my X3100 only has 8, compare that to the 96 my 8800GTS has), and the biggest problem is Intel's drivers at this point.

I think that if you'd take Intel's current GPU technology, create a fullblown GPU out of it with dedicated graphics ram, and have proper drivers, it'd actually be competitive with AMD and nVidia.
But as ATi can tell you, you don't just fix your drivers in a few weeks. It took years for ATi to really catch up with nVidia in terms of driver quality... It wasn't until the Catalyst driver series that ATi reached the same level of compatibility and performance as nVidia had in D3D, and OpenGL took ATi even longer. Many other brands never even got that far... eg S3 or XGI, who had promising D3D9-hardware on paper, but just wouldn't run most software properly.
Intel is still struggling there aswell. I hope they are wise enough to invest a lot into the development of drivers for Larrabee, because good drivers are at least as important as a good hardware-design.

Yes, this is why I'm not holding my breath for Larabee. Intel has released FOUR generations of pixel shader-capable hardware without proper vertex shader support. And because lighting gets more and more complex with every release, each generation without hardware vertex shaders has put them deeper in the hole.

Intel has basically been relying on CPUs to do this all these years, and although Core2 provided a (much needed) boost over the Pentium 4, it wasn't nearly enough to keep up. Basically, I'll beleive Intel can make a real GPU with real drivers when I see it.

Incidentally, ATI actually has a real GPU in the 780G that will light a fire under Intel's ass. Finally, somebody besides Nvidia is offering an IGP with hardware vertex shaders. and this solution even beats out Nvidia's IGP.

 

[StealthyFish]

The irony is that Intel's GPU architecture is basically cutting-edge. They have SM4.0-capable unified shaders and hardware-accelerated MPEG4 and that sort of thing.
It's just that they are designed so cheaply that they only have very little shader units (my X3100 only has 8, compare that to the 96 my 8800GTS has), and the biggest problem is Intel's drivers at this point.

I think that if you'd take Intel's current GPU technology, create a fullblown GPU out of it with dedicated graphics ram, and have proper drivers, it'd actually be competitive with AMD and nVidia.
But as ATi can tell you, you don't just fix your drivers in a few weeks. It took years for ATi to really catch up with nVidia in terms of driver quality... It wasn't until the Catalyst driver series that ATi reached the same level of compatibility and performance as nVidia had in D3D, and OpenGL took ATi even longer. Many other brands never even got that far... eg S3 or XGI, who had promising D3D9-hardware on paper, but just wouldn't run most software properly.
Intel is still struggling there aswell. I hope they are wise enough to invest a lot into the development of drivers for Larrabee, because good drivers are at least as important as a good hardware-design.

With the kind of money Intel is able to invest into a graphics program such as Larabee and the man-power and brain trust they have available to them, I'm sure they can design something. Everyone uses Intel's integrated graphics as a show of how good Intel's graphics design group is. Do you really think that Intel has put it's entire effort into designing a good graphics processing unit? It's much like taking Nvidia's onboard graphics chips and generalizing that Nvidia's cards would be just as much junk as their on-board graphics chips. I have strong confidence in Intel's larabee group just because Intel has the money, man-power, and experience to design something. I know larabee is already in development and isn't just an estimated future program anymore.

 

[Ranari]

Some of you are forgetting that Intel has already taken a stab at the mid/high end GPU market, and failed miserably competing against Nvidia/ATI at the time. I have little faith in Intel's ability to compete against the current GPU hardballs; both of which have years of experience designing their hardware. Discrete GPUs are like a Formula-1 race cars; they're only as good as their driver.

If anything, Intel's ability to respond to the 780G will show just how well they'll be able to compete with Nvidia and AMD in the discrete GPU market.

 

[dirtypop]

Some of you are forgetting that Intel has already taken a stab at the mid/high end GPU market, and failed miserably competing against Nvidia/ATI at the time.

Have they?

 

[Mr. Bluntman]

Everyone uses Intel's integrated graphics as a show of how good Intel's graphics design group is.

Or because most people that buy PCs get them from brick and mortar stores, and that the majority of these PCs (and most all Intel laptops) just so happen to have Intel Celerons or Core 2's that have integrated graphics. The average joe doesn't know what integrated graphics is. And mostly what they look at is price, accesories, and appearance. Most of all, they trust the opinion of the salesman that it's "good enough" for what they do. So they buy it none the wiser.

Your logic:

Some of you are forgetting that Intel has already taken a stab at the mid/high end GPU market, and failed miserably competing against Nvidia/ATI at the time. I have little faith in Intel's ability to compete against the current GPU hardballs; both of which have years of experience designing their hardware. Discrete GPUs are like a Formula-1 race cars; they're only as good as their driver.

If anything, Intel's ability to respond to the 780G will show just how well they'll be able to compete with Nvidia and AMD in the discrete GPU market.

Am I the only one that remembers the i740?

Back in the Riva 128 and 3dfx Voodoo Graphics days this sucker was very competitive with the Voodoo in both speed and image quality, and it had the ability of higher resolutions like the Riva. ATi's Rage Pro was a pile of suck in comparison. Intel just dropped out of the market because they weren't interested, or was it Real3D that folded (Intel's partner in developing the i740) and Intel gave up on AIB cards?

I think now that Intel is devoting lots of resources on Larrabee, it won't resemble anything that we have seen from their integrated chipsets.

 

[Scali2]

Some of you are forgetting that Intel has already taken a stab at the mid/high end GPU market, and failed miserably competing against Nvidia/ATI at the time.

Well ofcourse, back then they didn't have any experience in 3d acceleration. Most of the competitors that were around at that time, have now 'failed' against nVidia/ATi aswell.
But that was ancient history. Today's hardware is completely different from those early trifillers. And Intel is different aswell.
Firstly, today's GPUs are very programmable and use high-performance SIMD floating point processing, making them very similar to CPUs. Secondly, Intel now has plenty of experience in designing GPUs.
Basically they 'just' need to scale up their IGP-designs and combine it with their CPU technology and other research projects, such as Terascale. You could say they have all the pieces of the puzzle, they just need to put it together.

It them comes down to the driver. But Intel wants to have working DX10 drivers on their IGP before the end of this quarter (technically Intel's IGPs was the first fully DX10-capable hardware on the market, they just didn't have the drivers ready, unlike nVidia). It seems they're taking drivers more seriously these days. Not too long ago they were showcasing their overhauled DX9 drivers for Windows XP, wtih proper vertex shader support, capable of playing games like Far Cry at reasonable framerates and quite good visual quality.
If they keep these driver improvements up, not only will their IGPs become more competitive, but I'll be getting more and more confident about Larrabee and its successors aswell. I'm quite sure that Intel is at least aware of the driver issue by now. They also have the money, so if they could just hire the right people to get their driver team going, we'll be in for some interesting times on the GPU-market.

I have little faith in Intel's ability to compete against the current GPU hardballs; both of which have years of experience designing their hardware. Discrete GPUs are like a Formula-1 race cars; they're only as good as their driver.

Well, ATi was very late to the 3d-game aswell, and their first few generations of hardware were quite lackluster. Then ATi realized that they are going to have to take 3d acceleration seriously if they want to remain in business. And that eventually led to the Radeon 9700 series, which is arguably the biggest landmark in the history of the GPU.
I don't think many people were expecting anything like that to come out of ATi (although the Radeon 8500 was a silent hint of things to come already.. I actually got one of those over a GF3 back then... which I initially regretted somewhat, because of driver issues. But as the drivers improved, it became a fantastic card).

 

[DaNotoriousDOM]

With Nvidia using 2y old tech it should take them no time to catch up.

 

[Simpson5774]

Just because people speculate that Intel could buy NV (which they wouldn't) doesn't mean its going to happen.

My question is, what kind of API could take advantage of Ray tracing? does DX9-10 have anything for ray tracing built in? I mean how does intel expect us to get from raster to ray tracing w/o an API to drive it?

 

[Menelmarar]

The irony is that Intel's GPU architecture is basically cutting-edge. They have SM4.0-capable unified shaders and hardware-accelerated MPEG4 and that sort of thing.
It's just that they are designed so cheaply that they only have very little shader units (my X3100 only has 8, compare that to the 96 my 8800GTS has), and the biggest problem is Intel's drivers at this point.

They aren't design cheaply, they are designed for their target market.

That is why they have the largest market share in GPU. They know what the average joe blow, internet email, office document processor needs. Low cost, low power. And they have an integrated product that meets this very well. They really don't even need a hardware 3D engine in my opinion. They do for Vista's new hardware accelerated GUI now, and maybe if users want to run old games. My dad's G965 can limp along in Half-Life2 engine games, Quake 3 engines, UT2004, and such. It's doable for previous generation games on low resolution 1024 displays.

As you mentioned, they have the shader units for SM4.0 and all already designed. They just need to scale it up from 8 units on an integrated solution, to 128 units on a PCI-e expansion card solution like nVidia has. They may already have groundwork for a competitor right now depending on how efficient their shader units are... No one can really know that though

 

[Scali2]

They aren't design cheaply, they are designed for their target market.

Yea, and the target market is cheap
That's what I meant to say anyway. They are designed to be cheap. Which is why they have so little shaders onboard.

As you mentioned, they have the shader units for SM4.0 and all already designed. They just need to scale it up from 8 units on an integrated solution, to 128 units on a PCI-e expansion card soultion like nVidia has. They may already have groundwork for a competitor right now depending on how efficient their shader units are... Noone can really know that though

Mind you, the Larrabee is a completely different beast.
Their current GPU solutions are much like the ones from ATi and nVidia. Using relatively low clock frequencies, trying to reach for very high IPC with massive parallelism.
Larrabee instead uses the x86-approach of speeds over 2 GHz, and not so much parallelism (it is rumoured to have 'only' 16-24 cores, a far cry from nVidia's 128). This might solve a problem in graphics... The more parallelism you add, the less efficient it gets. Generally the pipelines are arranged to render in 2x2 blocks... However, you don't always render 4 pixels (edge of polygon, occlusion etc). Perhaps Intel will go for a more serial approach, which allows for more efficient detection of which pixels to render, leaving less idle units.
On the other hand, this more CPU-like design will also be more suited to certain GPGPU problems, because it is more serial in nature, and can respond better to branches and things like that.
Thing is, nobody ever designed a GPU with the approach that Larrabee is taking. It will be interesting to see how well this approach compares to the conventional GPUs.

 

[defaultluser]

Am I the only one that remembers the i740?

Back in the Riva 128 and 3dfx Voodoo Graphics days this sucker was very competitive with the Voodoo in both speed and image quality, and it had the ability of higher resolutions like the Riva. ATi's Rage Pro was a pile of suck in comparison. Intel just dropped out of the market because they weren't interested, or was it Real3D that folded (Intel's partner in developing the i740) and Intel gave up on AIB cards?

I think now that Intel is devoting lots of resources on Larrabee, it won't resemble anything that we have seen from their integrated chipsets.

They dropped out of the market because they couldn't compete. Intel took the technology from the i740 and turned it into the integrated graphics for the i810.

Yes, the i740 was a a decent performer, but it didn't sell well. Within six months of release, you could buy 8MB i740 AGP cards for 40 bucks, because nobody wanted them.

The reason the i740 didn't do well was because Intel did their usual bang-up job of forcing their standards on everyone else. Intel was trying to force the entire graphics market to go AGP.The i740 was an AGP-only part, and by that I mean it used the onboard 8MB memory as a frame buffer only. This means that the card used DME texturing exclusively, which used system memory and hurt performance on systems with 66MHz bus (every system available at the time). It also meant you couldn't make a PCI version of the card, unless you were willing to tack-on expensive texture memory just like the Voodoo/Voodoo2 series.

AGP was good for Intel because they had THE BEST implementation, in their 440LX. If you wanted AGP from anyone else, you were stuck with the buggy chipsets from ALI and Via. Thus, iof you felt you needed AGP, you were going to buy an overpriced 440LX board and Pentium II CPU (Celeron still wasn't on the market).

The only way you could make a PCI version of the i740 was if you emulated DME memory with some dedicated texture memory. Unfortunately, since the card thought it was using MAIN memory for textures, it wouldn't fall-back to DMA from main memory when it ran out of local texture memory; the card had to have a big enough texture memory to hold ALL the textures required by the game, or it would have errors.

Real3D made such a PCI card, the 12, 16 and 24MB Starfighter PCI cards, with a massive 8-16MB texture buffer (read the reasons above), and because of the large amounts of expensive memory, these also did not sell well.

Did the graphics world eventually go AGP? Yes. Was the time of the i740 a good time to release an AGP-only graphics card? Hell no. And now, Intel is trying to force the graphics world to follow once again, but I expect another miserable failure.

 

[Mr. Bluntman]

Well, after some leaked slides hit the interwebs, it's time to resurrect this thread.

It will be interesting to see if this "radical departure" from the norm is actually going to work or faceplant into a big steaming pile of EPIC FAIL.

 

[defaultluser]

Well, after some leaked slides hit the interwebs, it's time to resurrect this thread.

It will be interesting to see if this "radical departure" from the norm is actually going to work or faceplant into a big steaming pile of EPIC FAIL.

I like how you labeled it a "radical departure," because it isn't, really. In the same way that GPU architectures are becoming more and more open and flexible, Larrabee is a multi-CPU x86 architecture that is made LESS flexible to improve performance.

What you end up with is a powerful vector unit with some flexible operations, just like we already have with modern GPUs. By the time Larrabee has any market prescence, there won't be any difference...just choose which architecture you want to develop for, because they're not compatible outside of OpenGL/DirectX.

Intel is betting that people will go with their mangled x86 concept, even though they're years late to market, and it's not like these are fully-x86 compatible. I don't see that happening.

 

[refraxion]

Can't wait to see how this comes out, really interesting indeed.

 

[Nahid]

For example, Larrabee's first product may be fabricated on a 32nm process.[5]