Cuda Supercomputer

[WmH]

Nvidia's hitting the market hard with their GPU based processing systems. This may be the way to speed 3D rendering in the years ahead.

Nvidia Supercomputer (http://www.nvidia.com/object/personal_supercomputing.html)

and a link to the build your own ideas - Build Your Own (http://www.nvidia.com/object/tesla_build_your_own.html)

I understand that '3D Coat' works with Cuda for better performance at high polygon levels.

CUDA is propritary and (as such) is limited to Nvidia GPU's (and a fairly narrow subset of that) and is unlikely to gain any widespread support.
OpenCL on the other hand, being an open standard and hardware agnostic (like OpenGL) is being supported by both AMD (ATI) and Nvidia and just might gain enough momentum to achieve critical mass.

Wiki (http://en.wikipedia.org/wiki/OpenCL)

Cronos (host orginazation) (http://www.khronos.org/opencl/)

We should see at least some initial application support soon as Apple has included it as a core technology in the soon to be released snow leopard.

[futagoza]

We should see at least some initial application support soon as Apple has included it as a core technology in the soon to be released snow leopard.

Maybe Juanxer can pop in here..., so how does it work then for EIAS in the future V9 etc.? Can 32bit Carbon apps "speak" OpenCl or CUDA then too under Snow Leopard?

Regards
Stefan

[juanxer]

I am trying to learn more about that: at the Game Developer Conference 2009 this week there have been some sessions regarding these technologies. Anyway, my guess is one can write OpenCL code portions and call them from any OS X API, be it Carbon, Cocoa or anything else, the same way one can call pure C or C++ code. After all, OpenCL is multiplatform, so it must be compatible with everything out there in that sense.

It seems CUDA and OpenCL are quite similar, coding-wise. What ought to be far simpler is the way one "compiles" it: in CUDA you have to play with at least one compiler for your normal code plus another for the CUDA code with a profile matching your GPUs and things. OpenCL is supposed to be more transparent and rely on non-conventional compiling techniques or something (which is not quite a Just-In-Time compiler, I read somewhere).

Anyway, the thing has just been released more or less, and the last Snow Leopard timetable says the final version ought to be released around August. Realistically, nobody will be able to show anything major OpenCL-wise until, I don't know, year's end?

[WmH]

Apple will likely be ready out of the gate with some of it's applications (remember that OpenCL was based on the work Apple did in it's pro apps (like motion) directly utilizing the GPU for computational (not just display) work, so they already have a framework in place)
Also remember that the OpenCL API is likely already in the hands of developers and being used (or will be after the ADC this spring) so I would guess that we will see both 3rd party and internal (Apple's pro Apps) at, or shortly after, the launch of Snow Leopard.

[plsyvjeucxfw]

Sounds like interesting (and wonderful) times ahead.

I take it this OpenCL will recognize and use all CPU / GPU chips within a given machine? That is, once the applications are written to take advantage of the technology.

[WmH]

Yes, provided there are drivers (just like OpenGL there is a software abstraction layer required) That is what makes OpenCL so interesting, is that the driver can be written atop almost anything, for instance Nvidia is writing the OpenCL driver on top of the CUDA construct they already have in place.

Interestingly enough there is already chatter on the Blender3d development boards (Ton, the Blender project leader is said to be very interested) Not any word from commercial developers, but that would be expected.
MS would of course like to throw a monkey wrench into this as they don't like open (non MS proprietary) industry standards. Perhaps they will be able to twist Autodesk's arm not to support it, at least initially, but it is likely that the smaller developers will quickly try to gain an performance advantage (Maxon, NewTek and Luxology) once the GPU drivers become available.
The one nice thing is that if ATI and Nvidia want to sell Apple GPU's (and they do, Apple is one of their largest and most important customers) they will NEED to have the OpenCL drivers written by the time Snow Leopard ships. (and porting the drivers back to Linux and Windows will be considerably easier than writing from scratch (porting OS X to linux might even be trivial)

[juanxer]

Not only GPUs and CPUs but specialized thingies such as DSPs, other special designs, FPGAs (sort of a processor construction kit on a chip which allows one to flash a processor design into it sort of like one flashes the EPROM in a videocard), etc., too, as long as the builders provide OpenCL drivers. If I am not wrong, this first release will cover GPUs and CPUs only, but the whole idea is to be able to harness any kind of kit available, do truly heterogeneous computing. Also, there is OpenCL for Mobiles, too.

I think what could delay things a bit is nVidia and ATI/AMD getting their drivers ready. They are working on it.

EDIT: WMH answered all that already :)

Microsoft has some sort of GPGPU thing via Directx, but I think it was something they defined years ago to provide some route for programmers to do this kind of things, as they were overhauling their DX APIs anyway. I don't see them trying to kill CUDA, STREAM or OpenCL with that.

The most interesting product to get OpenCL'd would be Intel's Larrabee.

[aleph-zahir]

CUDA is propritary and (as such) is limited to Nvidia GPU's (and a fairly narrow subset of that) and is unlikely to gain any widespread support.
OpenCL on the other hand, being an open standard and hardware agnostic (like OpenGL) is being supported by both AMD (ATI) and Nvidia and just might gain enough momentum to achieve critical mass.

Wiki

Cronos (host orginazation)

We should see at least some initial application support soon as Apple has included it as a core technology in the soon to be released snow leopard.

It would be nice if that was the case, but I'm not convinced that's how the industry works.

Advertising budget, kickbacks and under the table deals count for a lot.

Nvidia have a pretty big head start with 70+ cards, wrappers for the most widespread end user and industrial languages, applicationsalready shipped for video encoding, cryptography, bioinformatics et cetera.

My point is; just because opencl is better, don't assume it will be more successful. For example, Intel is the most successful cpu manufacturer, microsoft is the most successful OS producer, and they have a history of being pretty incompetent compared to their smaller - less successful - rivals.

Maybe Apple can pull something amazing out of the bag, they've done it before, but I wouldn't count on it. Their record isn't exactly 1:1 after all.

Or maybe I'm just getting cynical in my old age.

[WmH]

Cynical? No, I don't think so. Healthy skepticism is a benefit of experience.

However you have mixed examples, Intel has been the leader (except for the P4 series which was admittedly a dark time for them), though AMD and IBM have temporarily eclipsed them on occasion (and normally very short lived) they have maintained the leadership role in microprocessor design throughout most of the last 15 years.

MS's ride to prominence when Apple floundered (the sales bozo years) was not without some merit. Don't judge MS by the worthless load of spaghetti coded bloat-ware they peddle today. At one time MS was an impressive company with solid (if not terribly innovative) products. Even in the OS market, NT3,4 and the first release of NT5 (aka Win2000) were impressive and it was at a time when Apple's original OS (which was fairly limited) had been allowed to languish for years (amid a series of failed project's to modernize it) (though admittedly I never thought much of W95/98/Me which were their most popular, guess that goes to support your point ;-)

It is unfortunate that they have been permitted to bypass the laws (apparently money talks, both in the US and the EU (see now we are both cynics)) that prevent abuse of market-share. (BeOS was such a loss, think of what that OS could do in a net-book or with a minimalist Net-appliance! (even in the state it was many years ago when MS scuttled it)) They are currently strong-arming manufacturers who want to sell net-books loaded with android. (again that would likely improve netbooks usability 10-fold over XP)

P.S. Forgot to address the original point, yes while Apple has not had a 100% success record (nobody's perfect) their record (since the turnaround) has been pretty impressive. I wouldn't bet against OpenCL being successful. Even at ten-to-one odds ;-)

[aleph-zahir]

However you have mixed examples, Intel has been the leader (except for the P4 series which was admittedly a dark time for them), though AMD and IBM have temporarily eclipsed them on occasion (and normally very short lived) they have maintained the leadership role in microprocessor design throughout most of the last 15 years.


Personally I think any other chip design than x86 would kick the shit out of it on a level playing field. In "their day" (before the economic realities squeezed the life out of them) PPC, ARM, 68k, Alpha and SPARC all had better designs than the equivalent intel offering.

In my opinion Intel's "leadership role" comes from their talent at business and marketing rather than chip design. Some of it, of course, comes from better fabrication. But since that's just a secondary effect of larger capital, I don't think that's really relevant to the issue of design either.



MS's ride to prominence when Apple floundered (the sales bozo years) was not without some merit. Don't judge MS by the worthless load of spaghetti coded bloat-ware they peddle today. At one time MS was an impressive company with solid (if not terribly innovative) products. Even in the OS market, NT3,4 and the first release of NT5 (aka Win2000) were impressive and it was at a time when Apple's original OS (which was fairly limited) had been allowed to languish for years (amid a series of failed project's to modernize it) (though admittedly I never thought much of W95/98/Me which were their most popular, guess that goes to support your point ;-)



Personally I've never seen a microsoft OS which was the best available, even in a direct comparison with the OS they were plagiarizing their "new" ideas from.

DOS was a bad copy of CPM.

Windows 3.1 was functionally equivalent to the Apple, Commodore, SUN and IBM offerings from a decade before, but more resource hungry and less well integrated.

Windows NT didn't quite catch up with System 7, Solaris or RISCOS from 5 years before, even though it was pretty good for an MS product. Perhaps the first (and only so far) stable kernal they developed.

XP and OS/X ? Pull the other one, it's got bells on. XP is basically NT with a tarted up GUI.

Vista? Barely even as good as XP, and frankly best forgotten.


So I don't think you can argue that MS ever had the best designs. What they had is a virtual monopoly, backscratching deals with hardware manufacturers, good lawyers, and gigabucks of marketing money.

I'd still stick to my original point; "better" seldom translates into "more successful".

Which is not a million miles away from CUDA vs. OpenCL.

[juanxer]

But then the problem with CUDA is that it is nVidia compatible-only (the same way STREAM is for ATIs), and it requires recompiling each time you move your CUDA apps to a different nVidia GPU setup. Not very average Joe-friendly, really.

The fun point is that nVidia was all high-ending CUDA until ATI announced that it would put STREAM and OpenCL in its base drivers for all its modern GPUs, from the lowest of the lowest (including integrated shared memory ones) to the FireGL ones. Add to that all the marketing talk Intel was doing about Larrabee. That's when nVidia started making CUDA for game physics and so noises. OpenCL, if anything, is going to "democratize" GPGPU (either by becoming "it" or by kicking everyone else into it), and the industry really needs that. Because it wants to sell multicores and has no real case to justify it to most consumers. And because it is really useful for lots of things.

I'll see if I can do an infodump in a few days, but a couple of examples of pedestrian GPGPU uses:

-Antivirus: imagine loading a GPU with all the virus profiles of an antivirus as shaders. Basically, you could test a single file or even every TCP/IP packet your PC moves for all these virus definitions in a single siimultaneous pass, instead of consecutively.

-Cameraphone: image processing, sound processing, etc.


I'd say Vista was Microsoft's OS X 10.0: mostly blah. A necessary step, though. One has to remember OS X was a painful thing to use until Panther. My impression is its aim is to eventually build a .Net-only OS, the same way Apple wishes to eventually become a Cocoa-only thing.

And regarding Intel, I'd say all these factors (design, fabrication, marketing decisions and so) count in defining what a good chip builder is. For example, their strategy regarding multicore chips and on-chip memory controlers (delaying implementing them until their multicore fabrication tech had matured) has resulted in them regaining the crown with Nehalem while AMD was stumbling because of trying to be "design-pure" before being really capable of fabbing fourcores and so. Of course, Intel's ISA is a horror compared to a RISC one (although nowadays Intel chips are RISC cores plus a x86-to-RISC translator), but it is results what counts, and I couldn't be happier in a MacTel world after years of PPC horror stories

[aleph-zahir]

I personally hope that OpenCL wins out, because it's better for the users and developers. If it does it'll be among the few but significant standards where the marketing team and IP lawyers don't get their way.

As I see it there are, broadly speaking, two ways to bring new hardware/software to the market.

A)

Plagiarize an idea, supercharge it using a brute-force-and-ignorance "more power" approach, ram the result down user's throats with saturation advertising and FUD marketing, invest hugely in fabrication to take advantage of economies of scale, make anti-competitive deals with related products and OEMs to stunt support for competitors, take advantage of existing capital so you don't have to profit on the first version to aid market penetration, use every dirty trick you can think of to keep a strangle-hold on the market.

B)

Make a product that's better than the competition.


In a rational world approach B would be the winner every time, but you only have to look at the market leaders to see which really works the best.

[juanxer]

But then that bag of tricks is used by just everyone :) .

Isn't Apple subsidizing all its Pro apps in order to sell hardware, Sony Nintendo and Microsoft subsidizing their consoles in order to sell software, nVidia ATI and Intel FUDing each other to death, and the same goes for Apple lying through its teeth in most of their "I'm a Mac" adverts or their PPC claims in the last gasps of the architecture? Lots of pots and kettles around, I'd say.

Let's be real: as nice as the RISC architectures are, they were just a means for workstation builders to not depend on the usual providers those days (Moto, Intel, IBM, etc.) and gain true hardware independence. The main claim, that RISC's reduced instruction set, single instruction+data formats and such would inherently mean faster chips was nullified by more relevant "dirty tricks" like multistage caches, superscalar architectures and so that provided more horsepower, not just to the x86 family but to the RISCers themselves, actually. That, and industry politics that often included companies pulling a Darwin Award out of the jaws of success, was what reduced the field to x86 (Intel, AMD, Via and a few others), IBM's POWER, SPARC, ARM and some others... which are not so few, considering. The same goes for OSes: Microsoft bought its way into PC-DOS; Apple bought its way into GUIs via Xerox (paying for access to their research); Microsoft did the same (by licensing Apple's, actually); Apple bought NeXT and destroyed its multiplatform nature; Final Cut Pro and the rest of the Pro Apps were born from the corpses of at least eight companies and killed product lines. Etc.

Also, each player has gone through highs and unexpected lows before. Intel is not that unassailable (they can't compete with ARM without abandoning their ISA, the IA64 fracas, the stagnation of the original Pentium line). Microsoft isn't so, too (Linux servers, Linux netbooks, the security mess XP is). But they aren't without technical merit where it is due (Intel's Core line has worked fantastically, all signs suggest Windows 7 is going to be their Panther, etc.). And everyone here plays hard, sometimes risking violating the law. Even Apple has had its borderlines in the EU with its iPod/iTunes/iTMS bundling practices).

(and what is new, really, I guess :) )

I know what you mean by those A) and B) options. The thing is, they aren't mutually exclusive. The "bad guys" can do brilliant products too. Those of us that are Apple users have been through at least three or four Ministry of Truth reversals in this regard :rolleyes: .

[juanxer]

ATI demoed their Havok physics simulation middleware for games reimplemented in OpenCL, which means it has its OpenCL drivers in working shape already.

[Navstar]

I think Intel's Laurabee holds more promise that just CUDA alone. But then again, CELL held lots of promise and where has that gone. (Besides Roadrunner & PS3)

[juanxer]

Now that you mention Cell, ultimately it ought to be possible to put it under the OpenCL umbrella, too.

Larrabee's advantage ought to be that it's sort of a Pentium farm on a chip: you can run both conventional apps on it (slower in a per core basis than in current Intel processors) and many-core-optimized apps, everything being x86 code instead of, say, PPC code here and SPE code there in the case of the Cell.

Intel is supposed to build PCIe Larrabee-based graphics cards first, but I read somewhere that one idea was to have Larrabee be just one more component in the set they use when they put together a new variant of a processor ("here we put a couple of cores, here a memory controller, here a cache unit and, hey, let's put a dozen Larrabee cores as a GPU/many-core unit").

Anyway, they say they expect to have product out by this year's end. First of all, it will have to be competitive with traditional 3D cards, as they are primarily marketing it as such and as a GPGPU-killer as secondary (they are a bit unclear in that, which feels a bit silly). If it manages to do so, things will become very interesting.

[CGTalk Moderation]

This thread has been automatically closed as it remained inactive for 12 months. If you wish to continue the discussion, please create a new thread in the appropriate forum.