Tegra 4 Likely To Include Kepler DNA

MrSeb writes "Late last week, Jen-Hsun Huang sent a letter to Nvidia employees congratulating them on successfully launching the highly acclaimed GeForce GTX 680. After discussing how Nvidia changed its entire approach to GPU design to create the new GK104, Jen-Hsun writes: 'Today is just the beginning of Kepler. Because of its super energy-efficient architecture, we will extend GPUs into datacenters, to super thin notebooks, to superphones.' (Nvidia calls Tegra-powered products 'super,' as in super phones, super tablets, etc, presumably because it believes you'll be more inclined to buy one if you associate it with a red-booted man in blue spandex.) This has touched off quite a bit of speculation concerning Nvidia's Tegra 4, codenamed Wayne, including assertions that Nvidia's next-gen SoC will use a Kepler-derived graphics core. That's probably true, but the implications are considerably wider than a simple boost to the chip's graphics performance." Nvidia's CEO is also predicting this summer will see the rise of $200 Android tablets.

Comments

Paper tiger

[Anonymous Coward]

So will this version be something more than a paper tiger? So far the Tegras have sounded better on paper than their real world performance ends up being.

Re:Paper tiger

[Lunix Nutcase]

Yes, which means it should be far more advanced than the A5X. The A5X is the same A5 from March of last year doing nothing but bumping the core count in the GPU and adding some memory. The GPU in it is just the 3 year old SGX543 that is also underclocked ain comparison to something like the PS Vita that uses the same GPU core but at a higher clock speed. So the Tegra should have had no issues beating the A5X in every test hands down.

Wow

[Hatta]

Where did they get Johannes Kepler's DNA?

Re:Wow

[Anonymous Coward]

From the outside of Tycho Brahe's fake nose.

Don't ask.

Re:Wow

[bacon.frankfurter]

They cloned it from Kepler's blood taken from a mosquito fossilized in amber, obviously. DNA on a chip. Makes perfect sense. Kepler's laws of planetary motion probably add a significant boost to pipeline performance. And what better way to integrate that functionality than by cloning Kepler himself, and regrowing his brain on a chip! I was wondering how long it would take to grow a brain on a chip, after they successfully created a gut on a chip

Re:Wow

[Ironchew]

Nvidia's next-gen SoC will use a Kepler-derived graphics core.

When did Johannes Kepler solve a graphics core?

Changing marketing terms

[T.E.D.]

Nvidia calls Tegra-powered products 'super,' as in super phones, super tablets, etc, presumably because it believes you'll be more inclined to buy one if you associate it with a red-booted man in blue spandex.

Wayne-powered products will of course be called "bat" instead.

nVidia's CEO is a little behind the times.

[bistromath007]

The droidpad I'm posting this from cost $200.

Re:nVidia's CEO is a little behind the times.

[Anonymous Coward]

I have a Nook Tablet, and I'm very happy with the hardware, happy with it as a book reader, but I wish it was a real Android tablet (with Bluetooth and Android Market^H^H^H^H^H^H^H^H^H^H^H^H^H^HGoogle Play).

So? Pick from CM7 or CM9.

Enjoy.

GPU programming is a nightmare.

[140Mandak262Jamuna]

It is possible to use the GPU effectively to speed up some scientific simulations. Usually in fluid mechanics problems that could be solved by time marching (or physics that obey hyperbolic governing differential equations). But working with the GPU is a real PITA. There is no standardization. There is no real support for any high level languages. Of course they have bullet points saying "C++ is Supported". But you dig in and find, you have to link with their library, there is no standardization, you need to manage the memory, you need to manage the data pipe line and fetch and cache, the actual amount of code you could fit in their "processing" unit is trivially small. All it could store turns out to be about 10 or so double precision solution variables and about flux vector splitting for Navier Stokes for just one triangle. About 40 lines of C code.

On top of everything, the binary is a mismash of compiled executable chunks sitting in the interpreted code. Essentially the if a competitor or hacker gets the "executable" they can reverse engineer every bit of innovation you had done to cram your code into these tiny processors and reverse engineer your scientific algorithm at a very fine grain.

Then their sales critter create "buzz". Make misleading, almost lying, presentations about GPU programming and how it is going to achieve world domination.

Re:GPU programming is a nightmare.

[PaladinAlpha]

Half of our department's research sits directly on CUDA, now, and I haven't really had this experience at all. CUDA is as standard as you can get for NVIDIA architecture -- ditto OpenCL for AMD. The problem with trying to abstract that is the same problem with trying to use something higher-level than C -- you're targeting an accelerator meant to take computational load, not a general-purpose computer. It's very much systems programming.

I'm honestly not really sure how much more abstact you could make it -- memory management is required because it's a fact of the hardware -- the GPU is across a bus and your compiler (or language) doesn't know more about your data semantics than you do. Pipelining and cache management are a fact of life in HPC already, and I haven't seen anything nutso you have to do to support proper instruction flow for nVidia cards (although I've mostly just targeted Fermi).

Re:GPU programming is a nightmare.

[Anonymous Coward]

This is so far detached from reality it almost makes me wonder if it is someone shilling for Intel or another company that wants to defame GPU computing.

First, the claim that it is effective "usually" in PDE solving is absurd; examine the presentations at any HPC conference or the publications in any HPC journal and you will quickly find numerous successful uses in other fields. Off the top of my head, I have seen or worked on successful projects in image processing and computer vision, monte carlo simulations of varying complexity in any number of fields, optimization problems and risk analysis in computational finance, statistical analysis of data from experiments or observations, and (less commonly) in non-numerical-computing applications (some graph algorithms map fairly well to a GPU).

Second, the claim that GPU programming is a pain in the ass -- in general, there is no doubt. It requires more time investment due to the optimization required to get performance that justifies the hardware cost, and far more difficult debugging than a CPU program (particularly a serial CPU program). The rest of your claims here, however, are once again nonsense. You say there is no standardization, but it is not even clear from context where you found standardization lacking -- the hardware? There are new architectures every few years, but they are almost always backwards compatible (ie, they will run old code). The only thing the architecture updates do is expose new features, much like new CPU architectures may add to the instruction set. Or is it software standardization you're looking for? OpenCL and CUDA are both open standards -- one controlled by a very slow moving board of industry representatives, one controlled by a fairly fast moving single company. This is very, very similar to the state of graphics APIs, where OpenGL and DirectX fill similar roles. The idea that they don't support high level languages is questionable -- what defines a high level language? If C or C++ qualify than definitely CUDA and possibly OpenCL do as well. If they don't, then no -- but expecting to write for hardware targeted solely at high performance and scientific computing in Ruby or whatever is idiotic. I would also love evidence (meaning official claims from a manufacturer, not some 3rd party who is just as misinformed as you) of claims to support C++. The closest I can think of is nVidia expanding the support for C++ features (templates, classes, ...) in CUDA -- but I've certainly never seen them or AMD claim to support C++ on the GPU. It sounds like you were expecting to take a C++ program, hit "Compile for GPU", and get massive parallelization for free, which reinforces the idea that you are did not do a shred of research into GPU computing. That you need to link with their library is obvious -- did you expect you would be communicating with the GPU without going through the driver? You're either going to use a library and compiler directly in the build process (CUDA) or indirectly at runtime (OpenCL). That you need to manage the memory is once again idiotic whining -- the people targeted by this are already managing their memory. This isn't to accelerate your shitty Web2.0 application, it's for serious numerical computation of the sort that is almost always written in C++, C, or (ugh) Fortran. That you manage the data pipeline and fetch and cache is at best a half truth; the shared/local memory is similar to a cache, but both AMD and nVidia GPUs use an actual cache that is not controlled at all by the programmer (unless possibly you are modifying the .ptx files for a CUDA kernel? I have not gotten into that sort of thing). The claim about being able to fit a limited amount of code is utter horseshit. And, finally, with the claim about how the code is stored -- this is true for OpenCL typically (there may be obfuscation methods or something along those lines as used by other languages which allow easy retrieval of code -- I'm not sure), but CUDA can be compiled to binary files that are no m

Re:GPU programming is a nightmare.

[maccodemonkey]

In my experience, GPU programming works exactly like you'd expect it to work. Your nightmare doesn't sound like it's with GPU programming, it sounds like it's with NVidia's marketing.

GPU processors are really small, so everything you've listed here is expected. The code size, variable limits, etc etc. The advantage is you have thousands of them at your disposal. That makes GPUs extremely good when you need to run a kernel with x where x is from 0 to a trillion. Upload the problem set to VRAM, and send the cores to work.

Stuff like C++ and high level languages is also not good for this sort of work. I'm not even sure why people are bothering with C++ on GPGPU to be perfectly frank. Again, you're writing kernels here, not entire programs. C++ is honestly bulky for GPGPU work and I can't imagine what I'd use it for. Both CUDA and OpenCL are already pretty high level, any further past that and you're risking sacrificing performance.

Interpreted code is also good. It's usually JIT compiled for the architecture you're working on. In the case of OpenCL and CUDA, it could be recompiled to run on an ATI card, NVidia card, or local CPU, all of which have different machine languages that you won't know about until runtime.

It sounds like you're angry because GPU programming isn't very much like programming for CPUs, and you'd be right. That's the nature of the hardware, it's built very different and is optimized for different tasks. Whether that's because you were sold a false bill of goods by NVidia, I don't know. But it doesn't mean GPU programming is broken, it just may not be for you. It mostly sounds like you're just trying to cram too much into your individual kernels though.

GPU programming nightmare: improvements are commin

[IYagami]

It is possible to use the GPU effectively to speed up some scientific simulations. Usually in fluid mechanics problems that could be solved by time marching (or physics that obey hyperbolic governing differential equations). But working with the GPU is a real PITA. There is no standardization. There is no real support for any high level languages. Of course they have bullet points saying "C++ is Supported". But you dig in and find, you have to link with their library, there is no standardization, you need to manage the memory, you need to manage the data pipe line and fetch and cache, the actual amount of code you could fit in their "processing" unit is trivially small. All it could store turns out to be about 10 or so double precision solution variables and about flux vector splitting for Navier Stokes for just one triangle. About 40 lines of C code.

  On top of everything, the binary is a mismash of compiled executable chunks sitting in the interpreted code. Essentially the if a competitor or hacker gets the "executable" they can reverse engineer every bit of innovation you had done to cram your code into these tiny processors and reverse engineer your scientific algorithm at a very fine grain.

Then their sales critter create "buzz". Make misleading, almost lying, presentations about GPU programming and how it is going to achieve world domination.

According to wikipedia, there are frameworks (like Open CL http://en.wikipedia.org/wiki/OpenCL ) in order to program in high level languages and have compatibility through various platforms

Re:And now for the obvious question...

[Anonymous Coward]

WTF is Kepler?

A new gaming-oriented GPU from NVidia that can't compete with even the previous generation of GPUs (Fermi) on many compute applications, namely integers. It's fine if you do single-precision floating point stuff all the time, but terrible if you want to work with integers or double-precision floats.