Nvidia in full philosophical retreat for Tegra 3
The dreaded SIMD added to T30
Apr 6, 2011 in Desktop, Efficiency, Graphics, Microprocessors, Mobile
Hot on the heels of the short-lived but self-described success of Tegra 2, Nvidia (NVDA) is going to launch a new core in about a year cunningly called T30 or Tegra 3. This one adds one major feature that implies a completely new core.
Short story, the feature is Neon, a 128-bit SIMD engine that is integral to the core. Tegra 1 and 2 did not have it, Tegra 3 does. It takes up a lot of space, power, and most notably it lessens the need for GPU and GPU compute functions, which is why Nvidia treated it like an infectious rash up until now.
Unfortunately for them, their counter-offer, proprietary GPUs and proprietary tools to write with, didn’t go over well. If a coder has to choose between marginally faster for a short time with proprietary tools and a ubiquitous standard with lots of hardware options, guess which one they go for?
When Nvidia announces the Tegra 3 architecture, expect a full ‘triumphant retreat’ in to the ARMs, pun intended, of A9+Neon. This same core will be the basis for T40/Tegra 4, and T50/Tegra 5 is now known as Denver, but the x86 part has been dropped for legal reasons.
So, the more things change, the more they stay the same. People making devices like consistency, choices, and most of all, standards. If you buck that trend in an established segment, the market reacts. If you survive that, you don’t buck it again. Adding NEON shows that Nvidia learned the hard way.S|A
19 Responses to “Nvidia in full philosophical retreat for Tegra 3”
Trackbacks/Pingbacks
-
-
Jun 6, 2011
[...] upping the ante (namely, the Samsung Galaxy S II) and realized that Tegra 2 is a dying platform (for many reasons) that will soon give way to the likes of products from Qualcomm and Texas Instruments (Motorola’s [...]
-
-
Jun 6, 2011
[...] upping the ante (namely, the Samsung Galaxy S II) and realized that Tegra 2 is a dying platform (for many reasons) that will soon give way to the likes of products from Qualcomm and Texas Instruments [...]
obviously, the author never coded GPGPU. Coding under CUDA is not “marginally faster”. is not only far faster, but much easier to program thanks to lots of libraries, .NET extensions, graphic profiling tools, available code, and support.
Code is faster, easier to program, and much more easier to maintain.
Qualcomm Snapdragons are the only ones using Neon currently unless another has snuck out? That’s why Snapdragons do so well at Linpack
Oh dear Charlie come on and face it
Tegra family was a success .
and by the way who in mobile world care
that NVIDIA make 600W GPU all they care
is that how much Tegra consume and how fast
NVIDIA can provide it to them .
This chips will be success till the day NVIDIA
Can Ship them faster than other competitor
with the same TDP as their competitors
and also the same performance .
OpenCL, CUDA, Tesselation, Unified Shaders on a device that has to consume 1W at its maximum power… you guys are funny really…
Before Tegra2 there was not decent 3D acceleration on mobile device and now everyone want DX11-like features on their phones!
BTW, Tegra3 cores seems to use Unified Shaders.
Already 4 cores are a waste of space on the die, but we’ll see how new applications for tables will be able to exploit them (folding @ home or renderman on your tablet?)
Yes, this is the most clever thing of all the article (let’s figure the rest, then). In fact it is well known that nvidia competitors use “standard” GPU whose implementation and programming model can be learned from any electronics manual o kid painting book… bad boys those at nvidia that constantly invest in R&D to always come up with something new every generation.
By the way, the “short-lived” Tegra2 will be almsot 2 years old when Tegra 3 will be on market. And only lately competitor managed to pass its performances. Just before Tegra 3 will bring them again in the low level part of the market.
@noone
Sounds like you don’t understand the technology.
OpenCL has an embedded profile from Khronos. No different from OpenGL ES being the mobile variant of OpenGL. ARM is behind OpenCL, Imagination Technologies is support OpenCL on their GPUs which are used in rather popular ARM processors such as the A5 in the ipad2. Take a look at the graphics performance the PowerVR SGX 543MP2 in the ipad2 is giving! Blows Tegra2 out of the water. It makes perfect sense to have OpenCL support because done right, parallel hardware should lower power consumption for the same compute throughput. Take a look at NVidia’s own cuda marketing..
Tegra3 is NOT using unified shaders. NVidia themselves said this at the announcement, it’s just more of the ULP GeForce GPU.
You probably do not understand what’s usefulness.
Tell me a single application that can be profitably used on a tablet that takes advantage of a OpenCL acceleration.
I have not said that OpenCL is useless in general, I said it is completely useless on tablets that are devices not thought for content creation. Do you want to edit a photography using your fingers? Or model and render a 3D scene? Well, yes, maybe useful for real time presentation effects, I agree.
You can’t list 5 applications that are OpenCL accelerated on desktop/workstation and you believe that OpenCL will be the killer feature for tablets? Oh, come on. If it will be, it will be when SoC GPU will much much faster than they are now, as the parallelism they can offer now is very limited (yes, it is much better having a CPU integrated vectorized ALU).
And that’s what’s related to the OpenCL thing? Do you thing it is twice as fast because it uses OpenCL acceleration? Or maybe because it integrates four times the cores of the previous version? Are you aware that it came out more than a year after Tegra2? And Tegra3 is to be launched soon (and is already in production)? When do you think PowerVR will release is next SoC able to surpass Tegra3 power? In next 6 months? Or in a year?
As said, before Tegra there was not decent 3D graphics (nor 1080p playback) on any mobile device, and now after a single generation people think that without CUDA, DirectX 11 capabilities, OpenCL and such a device cannot be useful. We are long way distant when that will be true.
It is also a question if timing: when the fist tablet generation will finally be launched, all producers using Tegra2 will already find Tegra3 on their desktop ready to be swapped in the new products. Or possibly, they are already doing that right now.
I presume that Tegra 4 is a straight die shrink to 28nm of Tegra 3? This will reduce the cost of a large 40nm chip (Tegra 3 could be over 100mm^2 given it is quad-core and includes Neon, and the graphics is 50% larger), and also run faster.
It could also be the stage where the integrated graphics is enhanced to become unified and updated. Given low bandwidth on a mobile device, tessellation would be an obvious feature to include as well.
Besides that … no A15? Denver being a custom core, probably designed around the A15 capabilities…
@jjj Kal-El is quad core A9 but they’ve put 4 cores behind the same amount and bandwidth cache they had in Tegra2. The NEON unit is almost certainly a 64 bit implementation direct from ARM just as with their main A9 cores (unlike Qualcomm who do a custom 128 bit implementation of NEON).
Their GPU is 12 shaders but they are not unified. Tegra2 has 8 shaders split between 4 vertex and 4 pixel shaders and likely Tegra3 is 6 of each. Not unified, not programmable.
BUT IT’S OKAY GUYS ‘CUS AMD WILL BE DEAD BY THE END OF THE YEAR!!!!!!!!
(someone with a sense of humor) LOL! :-)
Another (broken) maddoctor..
Nvidia already announced and demoed Tegra 3.
Tegra 3 aka Kal-El quad core A9,12 GPU cores (shaders) started sampling in feb,production august for tablets,in smartphones for the holiday season(or so they say),on 40nm and yes it has NEON but all this has been public since mid february.
Let’s face it. nVidia has an image problem in the devices community. What are they best known for? Hot-running, power-hungry discrete GPU cards. Translates well to cool, power-sipping CPUs doesn’t it?
So, how poor have T2 sales been anyway?
None of the devices launched so far have moved big numbers, except maybe the Thunderbolt. Is that a T2 phone?
No, it isn’t.
Farcically, T20 not only didn’t have NEON, but also did not have a GPU based on unified shaders and thus no CUDA or OpenCL either!
Farcically, T20 not only didn’t have NEON, but also did not have a GPU based on unified shaders and thus no CUDA or OpenCL support either!