NVIDIA Tegra Architecture

Well, most people's scepticism towards NVIDIA here is pretty easily understood. I mean, they've promised so much in the past, but they never ever really delivered.

Aside from NV30 and Fermi, when did they not deliver? AMD is notorious for over promising on performance and under performing. Titan Killer Anyone? Or how about Bulldozer? Steamroller? How bout 290x retail cards being slower than review cards? FX Processors? Or rebranding Fusion to HSA to fool consumers into thinking it's all new and different.
And this is only in recent years. I don't feel like going back into the ATi days with drivers that cheated on 3dmark, cut corners doing AA and AF, etc.
And don't get me wrong, nV has had their share of problems, but nowhere near as much of a track record of promising the world only to leave the fanboys in complete disappointment.
 
Aside from NV30 and Fermi, when did they not deliver? AMD is notorious for over promising on performance and under performing. Titan Killer Anyone? Or how about Bulldozer? Steamroller? How bout 290x retail cards being slower than review cards? FX Processors? Or rebranding Fusion to HSA to fool consumers into thinking it's all new and different.
And this is only in recent years. I don't feel like going back into the ATi days with drivers that cheated on 3dmark, cut corners doing AA and AF, etc.
And don't get me wrong, nV has had their share of problems, but nowhere near as much of a track record of promising the world only to leave the fanboys in complete disappointment.

We're talking about Tegra here. Tegra 1, 2, 3 and 4 were over-hyped and ultimately underwhelming.
 
We're talking about Tegra here. Tegra 1, 2, 3 and 4 were over-hyped and ultimately underwhelming.
can you be more specific ?
aside lack of Neon in T2 (problem discovered after the launch), I don't see any issue between specs and real world on T3/T4. The products are exactly what Nvidia said.
 
While the quality of the texture filtering isn't analyzed in the following points about mobile GPU rendering and image quality, the comparative precision of the prior generation hardware is explored in this blog entry and an article to which it links:

http://withimagination.imgtec.com/i...phics-we-go-hands-on-with-image-quality-tests

http://youilabs.com/blog/mobile-gpu-floating-point-accuracy-variances/

AF starts making really sense with at least 1 quad TMU; we're close that such GPUs become the lowest common denominator soon and if I have to have the dilemma between either AA or AF I'd rather pick the latter; supersampling isn't obviously affordable except some very rare cases and with MSAA or any derivative and/or approximation you still cover more polygon data with anisotropic filtering from a screen estate than with the former.

On top of that AF is adaptive while any form of AA isn't and in a power restricted environment I have my own priorities. Nonetheless AF quality and performance usually counts to NVIDIA's strenghts. As for precision with GK20A having both USC FP32 ALUs and additional FP64 SPs I don't think it's anywhere an issue anymore and shouldn't really.
 
can you be more specific ?
aside lack of Neon in T2 (problem discovered after the launch), I don't see any issue between specs and real world on T3/T4. The products are exactly what Nvidia said.

I thought T2 simply didn't implement Neon, not that there was a defect.
 
It was very silent about a defect that didn't exist? ;)

What he means is that nVidia didn't mention that Tegra 2 would be lacking NEON until someone asked them about in on their forum. Only then did they explain why they thought it wasn't worth including.

IMO it was a big mistake on their part. They probably feel the same way, or else I don't think they would have included it in Tegra 3 which had twice as many cores on the same process.
 
The lack of NEON in T2 was a real nuisance for video decoding.
I would've kept my Sony Tablet S for another year or so if I had the ability to decode H.264 720p high profile - something that even some previous-gen single-core SoCs with NEON could do without using dedicated hardware.

IMO, the video decoding deficiencies in T2 tablets ended up being the elephant in the room for T2.
If the SoC had decent dedicated hardware for H.264 decoding, most people wouldn't really notice the lack of NEON.
 
Yeah the deficient video processor was the real problem. Brute force CPU decode uses much more power so isn't a good solution anyway. I'm curious how valuable NEON is otherwise for typical tablet usage.
 
Yeah the deficient video processor was the real problem. Brute force CPU decode uses much more power so isn't a good solution anyway. I'm curious how valuable NEON is otherwise for typical tablet usage.

I can't say how useful it is in general, all I can tell you is my app gets a lot of benefit from it and until recently we didn't even support Tegra 2 at all. You wouldn't think there'd be a lot of people still using it but the e-mails came up, much more routinely than the x86 users even (which we also didn't support at the time).

NEON stuff did get asked about pretty frequently on Stack Overflow and ARM's forum.
 
Yeah the deficient video processor was the real problem. Brute force CPU decode uses much more power so isn't a good solution anyway. I'm curious how valuable NEON is otherwise for typical tablet usage.

I can't say how useful it is in general, all I can tell you is my app gets a lot of benefit from it and until recently we didn't even support Tegra 2 at all. You wouldn't think there'd be a lot of people still using it but the e-mails came up, much more routinely than the x86 users even (which we also didn't support at the time).

NEON stuff did get asked about pretty frequently on Stack Overflow and ARM's forum.
 
More Denver speculation(s)

As a follow up to my previous post:
http://forum.beyond3d.com/showpost.php?p=1821269&postcount=1993

Tegra K1 64-bit Denver core analysis: Are Nvidia’s x86 efforts hidden within?
http://www.extremetech.com/computin...nalysis-are-nvidias-x86-efforts-hidden-within


When Nvidia says that each Denver core is 7-way superscalar, it means that it has the hardware resources to perform seven instructions per clock cycle. Nvidia hasn’t said exactly what those hardware resources are (if it can decode seven instructions per cycle we’d be stunned), but it’s pretty clear at this point that Team Green has built an absolutely monstrous chip that should be capable of impressive performance. Maybe Nvidia’s claim that Denver is a “Super Core” isn’t just marketing fluff?
 
I do not want to further the derailment but I just want to add that they might still be using the fusion brand in japan, although it probably might not be directly from AMD.
Bd6md5sCEAIEaaN.jpg
 
Back
Top