nVidia scientist on Larrabee

[thilanliyan]

http://bits.blogs.nytimes.com/...ence-criticizes-intel/

 

[SickBeast]

How did that NV guy "break his silence"?

The title should read "NV's marketing department pays scientist to bash Larabee and make the competition look bad". :roll:

He points out that Larabee's x86 cores are wasteful in terms of die size. This may hold true for graphics performance, but he fails to mention the benefit of having that many x86 cores in your computer. Any video encoding app would benefit without having to be patched or re-coded, for example.

For most people, it's better to have more general-processing power than it is to have a ton of graphics power. If you look at most laptop computers, it illustrates this quite clearly.

 

[thilanliyan]

Yeah it wasn't a great article...I'm almost ashamed I posted it.

 

[SickBeast]

Originally posted by: thilan29
Yeah it wasn't a great article...I'm almost ashamed I posted it.

I just don't know how a scientist who claims to have a passion for parallel computing can bash a product like Larabee, considering what companies like Pixar are doing with it.

The entire article reeks of bullshit and marketing.

 

[TidusZ]

Tune in next week when Ati scientists talk about physx!

 

[faxon]

Originally posted by: SickBeast

Originally posted by: thilan29
Yeah it wasn't a great article...I'm almost ashamed I posted it.

I just don't know how a scientist who claims to have a passion for parallel computing can bash a product like Larabee, considering what companies like Pixar are doing with it.

The entire article reeks of bullshit and marketing.

beat me to the punch

 

[SickBeast]

You would think NV would want to give their scientists legitimacy and credibility, particularly when they're trying to push their CUDA standard.

IMO NV needs a new CEO.

 

[Idontcare]

Originally posted by: SickBeast
You would think NV would want to give their scientists legitimacy and credibility, particularly when they're trying to push their CUDA standard.

IMO NV needs a new CEO.

So too did AMD when Sanders was getting long in the tooth...we saw how that ended. Sometimes its better to keep the devil you know than risk the one you don't.

 

[thilanliyan]

I doubt Jen-Hsun would relinquish control that easily. Doesn't he have a major financial stake in the company?

 

[SickBeast]

Originally posted by: thilan29
I doubt Jen-Hsun would relinquish control that easily. Doesn't he have a major financial stake in the company?

Oh they never go easily and you have to half-bankrupt the company to pay for their golden parachute.

He's actually not that bad in light of the GPUs they have been making. I just think he has a huge flaw when it comes to standards like DX and CUDA; he seems unable to reach a consensus with other companies to compromise and act in a reasonable fashion.

In short I think his ego is too big and that he has a personality disorder. He reminds me of Rollo.

 

[Wreckage]

Originally posted by: SickBeast

considering what companies like Pixar are doing with it.

Nothing. LarryB does not exist yet.

 

[BenSkywalker]

Any video encoding app would benefit without having to be patched or re-coded, for example.

That would ignore what a non x86 based alternative with comparable layout could do. Some here may demand ignorance to enter into a discussion, but in all seriousness x86 is about the poorest architecture you could imagine for this type of architecture. Decode hardware when tiny die space per core is the essence of your design goal is rather foolish. What makes this worse, far worse, is that the applications will still require a recompile in order to run on Larrabee, it isn't an OoO architecture- default x86 code would roll over and die running on it(wouldn't be surprising to see a normal processor be faster on anything with decent amounts of branching).

With several hundred thousand transistors per core wasted on decode hardware, more trasnsistors utilized to have full I/O functionality given to each core, a memory setup that is considerably more complex then any of the other vector style processor choices available Larrabee is making an awful lot of compromises to potential performance to be more Intel like then it needs to be.

Everyone seems to be taking the stance that Larrabee must have a lot going for it because of how much Intel is putting into it. Itanium anyone? Everyone with so much as an extremely small dose of understanding knew that Itanium was going to be a huge failure in the timeframe it hit. Sadly, a VLIW setup for something like Larrabee would end up being a much better option then where they are headed.

I guess, the best way to think of it is that Intel clearly sees a major movement as does everyone else in computing power. The problem is, Intel wants to take as much lousy outdated broken down crap with them as they can. We already have x86 as our main CPUs to handle that garbage, why do we need more of the same wasted die space on our GPUs? To make it so that lousy existing x86 code that isn't well suited for extreme levels of parallelization can be recompiled in an easier fashion? So let's prop up our outdated poorly structured code base for a short term gain and hold back everything else in the long term? Just doesn't make sense to me.

 

[SickBeast]

Originally posted by: Wreckage

Originally posted by: SickBeast

considering what companies like Pixar are doing with it.

Nothing. LarryB does not exist yet.

I'm pretty sure that on a prototype level production studios are working with it right now. There was some sort of demonstration done.

 

[SickBeast]

Originally posted by: BenSkywalker

Any video encoding app would benefit without having to be patched or re-coded, for example.

That would ignore what a non x86 based alternative with comparable layout could do. Some here may demand ignorance to enter into a discussion, but in all seriousness x86 is about the poorest architecture you could imagine for this type of architecture. Decode hardware when tiny die space per core is the essence of your design goal is rather foolish. What makes this worse, far worse, is that the applications will still require a recompile in order to run on Larrabee, it isn't an OoO architecture- default x86 code would roll over and die running on it(wouldn't be surprising to see a normal processor be faster on anything with decent amounts of branching).

With several hundred thousand transistors per core wasted on decode hardware, more trasnsistors utilized to have full I/O functionality given to each core, a memory setup that is considerably more complex then any of the other vector style processor choices available Larrabee is making an awful lot of compromises to potential performance to be more Intel like then it needs to be.

Everyone seems to be taking the stance that Larrabee must have a lot going for it because of how much Intel is putting into it. Itanium anyone? Everyone with so much as an extremely small dose of understanding knew that Itanium was going to be a huge failure in the timeframe it hit. Sadly, a VLIW setup for something like Larrabee would end up being a much better option then where they are headed.

I guess, the best way to think of it is that Intel clearly sees a major movement as does everyone else in computing power. The problem is, Intel wants to take as much lousy outdated broken down crap with them as they can. We already have x86 as our main CPUs to handle that garbage, why do we need more of the same wasted die space on our GPUs? To make it so that lousy existing x86 code that isn't well suited for extreme levels of parallelization can be recompiled in an easier fashion? So let's prop up our outdated poorly structured code base for a short term gain and hold back everything else in the long term? Just doesn't make sense to me.

I find it interesting that you first bash x86, and then you go on to bash itanium. Why so much hate for intel? If we shouldn't use x86 and we shouldn't use itanium, what should we use? The way I look at it is that if there was something so much better out there, people would use it. I remember apple used an alternative platform awhile back and it was considerably slower than x86.

The way I look at it, 80 atom cpus will encode video faster than my G80 based GPU with 80 stream processors running CUDA. I would go so far as to say that I think those same 80 atom cpus could compete with anything NV has available today.

You bash x86, yet it is the language upon which just about everything today is written. I would certainly call NV's approach more radical. They're doing what they are with CUDA because they do not have an x86 licence. If intel had more involvement in rasterization, all current GPUs would probably run some variant of x86 already.

From a graphics standpoint, I agree that intel's approach is problematic and wasteful, but I think there are side benefits that counteract this and actually make the idea quite viable. There's also the idea that games programmed specifically for Larabee will not take a performance hit.

 

[Wreckage]

Originally posted by: SickBeast

I'm pretty sure that on a prototype level production studios are working with it right now. There was some sort of demonstration done.

In other words....nothing.

Originally posted by: SickBeast


The way I look at it, 80 atom cpus will encode video faster than my G80 based GPU with 80 stream processors running CUDA. I would go so far as to say that I think those same 80 atom cpus could compete with anything NV has available today.

So 80 CPUs can beat 1 GPU?

 

[BenSkywalker]

I find it interesting that you first bash x86, and then you go on to bash itanium.

Probably because you are trying to read it a certain way. I bashed the success of Itanium- Intel spent billions on engineering and years hyping it. In no uncertain terms was it a faiulre and everyone outside of Intel loyalists knew it was going to be in advance and could explain why. Everyone except Intel loyalists were right.

Why so much hate for intel?

No hate at all for Intel. Sometimes I do something stupid too, my friends are usually the ones who point it out to me, the people who genuinely don't like me just sit back and wait to get a good laugh

If we shouldn't use x86 and we shouldn't use itanium, what should we use?

VLIW would actually work quite well in a Larrabee style setup, it sucked trying to replace x86 CPUs for the masses. It still wouldn't be ideal, but it would be much, much better then x86.

The way I look at it is that if there was something so much better out there, people would use it.

You are not even close to that ignorant Geothermal power production is vastly superior to every other method we use in the world today, outside of Iceland it isn't really used anywhere. There is the question of massive amounts of money that need to be put into something to make it viable for the mass market. Intel is banking on scales of economy making Larrabee a viable platform. The problem is, they need it to be competitive in the GPU space before they will realize that goal. If it isn't, their whole plan falls apart.

The way I look at it, 80 atom cpus will encode video faster than my G80 based GPU with 80 stream processors running CUDA.

80 Atom CPUs utterly dwarf the die space of 80 stream processors, by a huge margin. Besides that, 80 Atom CPUs come in about 1/6th the raw procession power of 80 stream processors. We aren't even close to comparable in terms of raw power or die size, x86 it utterly demolished in your given comparison on both fronts(Larrabee has far better designed cores then Atom which would be an abject failure in every way possible if it was the design route they were taking).

You bash x86, yet it is the language upon which just about everything today is written.

Because it has been around for so long and it works extremely well at general purpose OoO code. For the specific useage we are talking about, it sucks, badly. Intel already had to add a bunch of instructions to be able to push Larrabee with a straight face, it is simply a bad instruction set to use for a vector processor.

I would certainly call NV's approach more radical.

Not in the space they are in. HPC application architectures are closer to stream processors then x86 OoO based ones.

If intel had more involvement in rasterization, all current GPUs would probably run some variant of x86 already.

No chance honestly. x86 is flat out bad for highly parallel code. Different length instructions right off the top- bad idea for the type of architecture they are using. Requiring decode hardware to alter code to something that is machine level friendly isn't a good idea when you could just create an architecture that produces machine level friendly code in the first place. Millions of transistors per Larrabee chip will be spent on decode front end, that is required because of x86 and wouldn't be if they used an architecture better suited for it.

From a graphics standpoint, I agree that intel's approach is problematic and wasteful, but I think there are side benefits that counteract this and actually make the idea quite viable.

By the time Larrabee hits, will it have any advantage over the GPUs? Given that Larrabee requires a recompile and won't run OoO code properly, what edge do you see it having over the DX11 class GPUs from a non graphics standpoint? Larrabee can not run existing code, and even with a recompile outside of very limited exceptions it can't run existing code remotely decently either.

There's also the idea that games programmed specifically for Larabee will not take a performance hit.

They won't take a performance hit compared to software rendering. Developers could make games to run faster on Larrabee if they sought to cripple them on rasterizers, but that is what it would take for Larrabee to be competitive.

 

[Cookie Monster]

Originally posted by: Wreckage
Nothing. LarryB does not exist yet.

Actually it does exist because Pat Gelsinger showcased the Larrabee wafer at IDF Beijing not so long ago. Its a prototype of some sort but you can probably guess that they are working with these prototypes in their labs as we speak. They cant fake it either unless they want to miss the Q4/09 or Q1/10 launch of larrabee.

 

[SickBeast]

Ben, I'm going to respectfully disagree with a large portion of your post, particulary in regard to the "procession power". The NV CUDA cores may be great at parallel tasks, but when it comes to general processing, which most of us do a great deal of, I'm certain that the cores in the Larabee will be a great deal faster.

You failed to refute the fact that 80 atom cores would encode video much faster than 80 CUDA cores, yet that is indeed the reality of the situation, and that is indeed what people are going to want to do with these cores.

From my standpoint as someone who used to do a ton of 3D rendering, something like Larabee is an absolute godsend that people in the industry will pay huge money for. Do you really think that Larabee will not be able to do video encoding and 3D rendering? Don't kid yourself.

You mention VLIW CPUs, yet the only example I can think of that is the Transmeta Crusoe, which was slower than what Intel was offering at the time by a large margin. I'm not saying that x86 is all great or that it doesn't have flaws; simply that people have not yet invested the resources in developing a superior platform that could take over. The closest thing I can think of to that would be some of the technology that Sun and IBM have.

NV's approach *is* more radical in the sense that they seem to want to be able to run general tasks like Photoshop on their CUDA processors. At least intel is allowing normal x86 software to run on an x86 core; no middleware or emulation required.

Your comments on geothermal energy were irrelevant and debatable; see my comments above. I would talk about geothermal but it's beyond the scope of this forum.

 

[SickBeast]

Originally posted by: Wreckage

Originally posted by: SickBeast

I'm pretty sure that on a prototype level production studios are working with it right now. There was some sort of demonstration done.

In other words....nothing.

Originally posted by: SickBeast


The way I look at it, 80 atom cpus will encode video faster than my G80 based GPU with 80 stream processors running CUDA. I would go so far as to say that I think those same 80 atom cpus could compete with anything NV has available today.

So 80 CPUs can beat 1 GPU?

If Larabee is currently producing A-list Hollywood movies, how is that "nothing"? Because you want it to be? JSH's ego is too big to withstand the Larabee and you're afraid it will make him cry? Aww. :roll:

BTW it's an 80-core CPU vs. an '80-core' GPU. WTF are you smoking? Please give me some.

 

[BenSkywalker]

The NV CUDA cores may be great at parallel tasks, but when it comes to general processing, which most of us do a great deal of, I'm certain that the cores in the Larabee will be a great deal faster.

They are in order cores without trace cache. They would likely still be faster then GPU based cores for general purpose computing, but an Atom core processor will obliterate it. This has nothing to do with nV cores, anything they have out at this point isn't going to be going up against Larrabee anyway.

You failed to refute the fact that 80 atom cores would encode video much faster than 80 CUDA cores, yet that is indeed the reality of the situation, and that is indeed what people are going to want to do with these cores.

Forgive me, but in my world die space doesn't appear out of nowhere. The fact that using 20 times more die space will likely yield better results should be a given. What you would do well to remember is that you need to bring up such an absurd comparison to be competitive. Also, I'm not certain that Atom would be better in that situation, would come down to clockrate doing some quick ballpark estimates. At equal clocks I'm looking at the GPUs coming out ahead by a considerable margin.

From my standpoint as someone who used to do a ton of 3D rendering,

So did I, all the way back to the Irix machines, doesn't mean I'm going to shut off reasoning because of it.

something like Larabee is an absolute godsend that people in the industry will pay huge money for.

Let's look at the economics of that for a second- how's SGI doing? Intergraph? 3DLabs? Wildcat? Reality is that that market is, at best, miniscule. Besides that though, you are assuming that the GPUs won't be comparable to Larrabee when it hits in terms of what they can handle. Larrabee, giving Intel every benefit on what they have claimed, is going to struggle to have processing parity with the GPUs of that generation(AMD or nV). How limited Larrabee is in execution compared to how limited the GPUs are will likely be the deciding factor.

You mention VLIW CPUs, yet the only example I can think of that is the Transmeta Crusoe, which was slower than what Intel was offering at the time by a large margin.

Wow, I'm guessing you are much, much younger then I had thought, heh. Itanium is VLIW

NV's approach *is* more radical in the sense that they seem to want to be able to run general tasks like Photoshop on their CUDA processors.

Larrabee will run certain Photoshop filters, bank on it.

If Larabee is currently producing A-list Hollywood movies, how is that "nothing"?

A task that was previously handled by clusters of very cheap off the shelf CPUs.

 

[SickBeast]

Meh, you caught a typo at the end. I meant to say CPU vs. GPU. I edited it already.

I'm surprised you wouldn't want an 80-core CPU in your box to help render 3D Max scenes. I know I would. It was taking my Athlon XP 3200+ over two hours to render each frame. It took a week for my computer to render my university thesis. I wanted to make a video but didn't have time to hive my university's computers together.

You can say that Larabee is "nothing", yet it is impressive. How much would 80 computers with 80 cheap CPUs in them cost? Really, Larabee is going to be like having an entire university computer lab in your computer. It's incredible.

You point out that 3dlabs and such didn't fare well, yet I know for a fact that at one of the firms I worked at they paid huge $$$ for the rendering workstations. They were capable of photo-realistically creating interior renderings of hospitals including the elaborate equipment. I was completely blown away by it, and that was 5 years ago running on a quadruple CPU box with a now outdated GPU.

You keep talking about die space, but really, if intel is able to put Larabee onto a single die and sell it for $300, why do you care so much? When NV pushed the envelope with G80 I bought one. It's tech like this that pushes the boundaries and moves things forward.

 

[SickBeast]

After a PM exchange I think that Ben is one of the best posters on here and personifies elite membership here.

 

[novasatori]

Originally posted by: SickBeast
After a PM exchange I think that Ben is one of the best posters on here and personifies elite membership here.

Aww I was having fun reading, and I agree

 

[BenSkywalker]

This discussion really, truly is not about Intel versus nVidia, anyone who thinks it is you are very mistaken.

In essence, it seems that all this discussion is boiling down to is one thing that is only by extension related to Larrabee, a compiler.

Everything you have talked about is not unique to Larrabee in any whatsoever in the GPGPU space- it becomes possible on Larrabee because they have the compilers that will create machine level code Larrabee can run for most existing x86 based code(not all, Larry won't handle certain instructions at all native to C based languages).

With DX11 class GPUs, from either nVidia or AMD, or hell even BitBoys if they release one, the core functionality possible with any of the GPUs will be largely constrained by the level of support the compilers for them, not the hardware. Certain types of code are going to be much, much faster no matter the state of the compilers. Given that Larry is going to launch against DX11 class GPUs, I see it us rather dishonest to try and compare the limits of current hardware and compilers to compare to it. Yes, these DX11 GPUs aren't out yet- neither is Larry.

I'm surprised you wouldn't want an 80-core CPU

It isn't about not wanting one, it is about if I am given the choice between 80 stripped down CPU like functional units or a couple hundred functional units that are faster and double as a vastly superior graphics solution. That is what we are talking about in the timeframe we are looking at.

You point out that 3dlabs and such didn't fare well, yet I know for a fact that at one of the firms I worked at they paid huge $$$ for the rendering workstations.

Irix machines used to start around $10K for a very basic workstation. You know the problem with that business model? Raw material costs on even the highest powered solutions we have are less then $20. What we are paying for is R&D and production. If R&D and production costs have an initial impact of $1Billion dollars and you are going to sell 50,000 of them you are dealing with $20,000 cost per unit. If you sell 100 Million it's $10 per unit. I think everyone will be able to spot the difference in approach you take from a business perspective there. This is why AMD and nVidia between the two of them were able to shut down all of the dedicated high end 3D hardware companies, simple scales of economy. If Larry can not compete in the GPU market, then it will be forced to earn its' worth in the miniscule HPC segment. Given the amount of development time and effort spent on it it will easily run an order of magnitude more in cost then Tesla if it remains in production. This isn't a viable approach to the GPGPU model and why it has the potential to work.

When OpenCL/Brooks/CUDA offer support comparable to a run of the mill C based compiler then the only thing seperating Larry from the traditional GPUs will be performance- because of this I look at the enormous amounts of wasted die space on Larry and just shake my head. It utilizes resources poorly for every possible element it will handle. It isn't even a good compromise solution as from what we have seen both AMD and nV have been doing well with.

The proof will be when all the parts are out, but in order to hit DX11 standards the functionality of GPUs for GPGPU tasks is going to increase a significant margin, then it will just be a matter of support from the software devs to get their code base running on the GPUs(which I would wager both AMD and nV will be pushing rather hard).

 

[SickBeast]

Ben, you're knocking this thing without having seen what it can actually do.

DreamWorks animation said that Larabee allows them to increase what they can do by 20X. (!)

If Larabee really sucked, Sony would not want to put it into a PlayStation 4 console that they will be losing money on to begin with.

Intel has said that all of their driver programmers are focused on their current IGP solutions. They have usurped the 3d Labs people to work on Larabee. This tells me two things:

- Larabee will not need drivers
- Larabee will be a monster when it comes to 3D rendering (hence the DreamWorks comments)

Just think of all the time saved in eliminating DX and OpenGL. The Mac and Linux suddenly become gaming platforms. Games will be able to run just as fast (or faster) in software mode compared to running in OpenGL/DX.

You seem really concerned about this chip, but when credible sources like DreamWorks and Anand himself are excited about a product like this, it does make me scratch my head a little.

Thanks for the great conversation and for getting back to me via PM. You have my apologies for my initial comments.