NVIDIA Tegra Architecture

[DSC]

http://www.sweclockers.com/nyhet/16627-nvidia-kepler-och-maxwell-i-framtida-tegra

Sweclockers speculates that Logan will have Kepler unified shaders GPU and Stark will have Maxwell GPU.

 

[Ailuros]

Heard it before, doesn't make sense since there are worlds of differences between a SFF mobile SoC and a high end standalone GPU architecture. It just won't fit in some people's heads.

 

[Alexko]

Wasn't Kepler originally advertised as capable of fitting into phones?

 

[Ailuros]

Wasn't Kepler originally advertised as capable of fitting into phones?

Officially by NVIDIA? Could be I've missed it but I don't recall such a thing. What various infamous rumor mongerers have cooked up in the meantime is obviously a completely different chapter; for some of them it's a miracle they can keep their left from their right foot apart. If you'd tell me how you'd fit even a reduced Kepler cluster into a smartphone/tablet SoC I'll come up with a viable theory how you can fit a full grown elephant in your home refrigerator

There's a reason ULP GeForce blocks in so far Tegras (and for that ANY GPU blocks from any IHV in this market) have been designed from ground up with power consumption being the top most priority. Simple example: texturing is de-coupled on desktop GeForces since G80, for Radeons much earlier. Find me one SFF mobile GPU where it's the case. By the time you start scratching differences and tradeoffs for power consumption in a SFF mobile design its fairly absurd to call anything a "Kepler" or a "Maxwell". What marketing will do is of course a completely different story; when they go as far and call each ALU lane a "core" I don't think the sky is the limit in any case.

 

[Arun]

Officially by NVIDIA? Could be I've missed it but I don't recall such a thing.

http://www.anandtech.com/show/5703/jenhsuns-email-to-nvidia-employees-on-a-successful-kepler-launch

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

The quote seemed unambiguous when I first read it, but I now realise it's slightly vague: it never confirms that that it is Kepler itself that will go into superphones, only that its architecture will help GPUs penetrate superphones. So it could *theoretically* be a loose derivative or it could even wait until Maxwell (as Maxwell would naturally still benefit from the architecture R&D, just like Kepler benefited from the R&D in Fermi/Tesla).

 

[Ailuros]

Loose derivative is one way to describe it obviously; it'll just come down to how "loose" things can be. Both are GPUs and both are at least partially called "GeForce", therefore it's the same thing; case closed

 

[OlegSH]

Some interesting reading - http://blogs.nvidia.com/2013/03/tegra-4-tegra-4i-win-praise-from-leading-semiconductor-analyst-firm/

 

[Nebuchadnezzar]

Some interesting reading - http://blogs.nvidia.com/2013/03/tegra-4-tegra-4i-win-praise-from-leading-semiconductor-analyst-firm/

Aren't they vastly underestimating Krait 400's performance here? They're stating a 10% IPC increase over the S4 Pro, but the Snapdragon 600 / Krait 300 already has a proven 15%+ increase over that. Krait 400 goes another step above that.

Other tidbits: 9.4W top TDP for Tegra 4, likely will never reach those 1.9GHz clocks on all 4 cores in any real situation.

They estimate 5W for 2.3GHz T4i, again something they say will never be reached in reality and 1.8GHz would be the more realistic target.

 

[Alexko]

Aren't they vastly underestimating Krait 400's performance here? They're stating a 10% IPC increase over the S4 Pro, but the Snapdragon 600 / Krait 300 already has a proven 15%+ increase over that. Krait 400 goes another step above that.

Other tidbits: 9.4W top TDP for Tegra 4, likely will never reach those 1.9GHz clocks on all 4 cores in any real situation.

They estimate 5W for 2.3GHz T4i, again something they say will never be reached in reality and 1.8GHz would be the more realistic target.

I bet if I put a rocket engine on my bicycle, it will go really fast too. Someone should declare me winner of the next Tour de France right away, clearly there's no need to even compete for it.

 

[ams]

When they say "no thermal constraints", don't they simply mean in a tablet form factor?

 

[ams]

Aren't they vastly underestimating Krait 400's performance here? They're stating a 10% IPC increase over the S4 Pro, but the Snapdragon 600 / Krait 300 already has a proven 15%+ increase over that. Krait 400 goes another step above that.

That is a possibility, but even so, it appears that the majority of the CPU performance increase in S800 vs. S4 Pro is from an increase in CPU clock operating frequency. NVIDIA estimates that S800 will be ~ 50-75% faster than S4 Pro in most CPU-intensive applications: http://forum.beyond3d.com/showpost.php?p=1715441&postcount=844.

Other tidbits: 9.4W top TDP for Tegra 4, likely will never reach those 1.9GHz clocks on all 4 cores in any real situation.

That may be true, but it is also not likely that S800 with reach 2.3GHz clock operating frequency on all 4 cores in any real situation too. In fact, it is unlikely that any quad-core SoC will reach their max rated CPU clock operating frequency on all 4 cores in a handheld device (although Project Shield may be an exception).

 

[Exophase]

When they say "no thermal constraints", don't they simply mean in a tablet form factor?

Not exactly:

In the larger tablet form factor, the chip should have more thermal room to perform. Even a typical tablet, however, will not support the full power of Tegra 4, so some throttling will still be required when all four CPUs are running.

If it really needs > 9W with all four CPU cores running then I'm inclined to agree. That said, if 2.35W per core at 1.9GHz would be a lot worse than what Samsung is claiming for their Cortex-A15 at 1.8GHz on what should be a similar/competitive process. The numbers they gave are more like 1.2-1.3W. That last 100MHz would cost you but shouldn't cost this much. Still, they have a point - if the power optimized core at 825MHz uses only 40% less power than a Tegra 3 A9 at 1.6GHz then that's not a good sign of what the other cores will use at 1.9GHz.

I'm not sure if that 670mW @ 845MHz figure was for the power optimized core or not (you'd think it would be), but that's indeed much worse than what Samsung claims. Unless I'm reading that graph badly they seem to be claiming under 400mW for a similar clock speed. Samsung hasn't exactly impressed anyone with power figures on Exynos 5250 so this doesn't look good.. 5410 should at least be somewhat more efficient even before taking into consideration the A7s but still..

 

[ams]

The full power of Tegra 4 will probably be reserved for Project Shield (ie. clamshell type device), but many of the benchmark results listed above were run on a reference Tegra 4 tablet at MWC, with essentially the same results as what we see in the list above.

 

[french toast]

The full power of Tegra 4 will probably be reserved for Project Shield (ie. clamshell type device), but many of the benchmark results listed above were run on a reference Tegra 4 tablet at MWC, with essentially the same results as what we see in the list above.

The question is what will they have to cut it down to, to be able to fit inside a smartphone?

 

[mboeller]

Aren't they vastly underestimating Krait 400's performance here? They're stating a 10% IPC increase over the S4 Pro, but the Snapdragon 600 / Krait 300 already has a proven 15%+ increase over that. Krait 400 goes another step above that.

Other tidbits: 9.4W top TDP for Tegra 4, likely will never reach those 1.9GHz clocks on all 4 cores in any real situation.

They estimate 5W for 2.3GHz T4i, again something they say will never be reached in reality and 1.8GHz would be the more realistic target.

....and the 9.4 W are only for the 4 A15 cores without GPU and I/O....
.....and only if the core reach the 1.9GHz with only 1.0V


Within the Tegra4i document they say that the Tegra 4i at 2,3GHz is 50% faster than Tegra3 (SpecInt). Assuming the Tegra3 runs at 1,7GHz it seems that the A9r4 cores are around 10% faster than the normal A9 cores.

 

[Ailuros]

....and the 9.4 W are only for the 4 A15 cores without GPU and I/O....
.....and only if the core reach the 1.9GHz with only 1.0V


Within the Tegra4i document they say that the Tegra 4i at 2,3GHz is 50% faster than Tegra3 (SpecInt). Assuming the Tegra3 runs at 1,7GHz it seems that the A9r4 cores are around 10% faster than the normal A9 cores.

How many T3 based smartphones shipped exactly with a 1.7GHz frequency (wasn't it just one HTc variant or am I wrong?) how often if ever did that one hit 1.7GHz and how did battery life look on that one despite that?

Of course will be 4i be quite a bit faster than T30/33 on many fronts, but that kind of comparisons are completely worthless.

By the way despite that they note that the 30fps for the T4i are just a "preliminary result" for it in GLB2.5 even then somewhat less than 3400 frames isn't exactly something to write home about. It seems single channel memory might be limit things more than I would had imagined.

 

[mboeller]

How many T3 based smartphones shipped exactly with a 1.7GHz frequency (wasn't it just one HTc variant or am I wrong?) how often if ever did that one hit 1.7GHz and how did battery life look on that one despite that?

Ahh yes, you are correct. According to the Tegra4i paper the improvement is 15% (SPECInt) not 10%.

 

[Ailuros]

Self correction it seems that the scores marked with that paragraph symbol, are the author's own estimates. Either way with 30fps estimated for the T4i in GLB2.5 (1080p offscreen) and the Qualcomm S600 in the Samsung Galaxy SIV yielding already 42fps in a product that will launch in a couple of weeks, NV better optimize the living hell out of their drivers since the challenge for next year's smartphones won't be "just" 40% higher than the current estimates.

 

[Exophase]

The full power of Tegra 4 will probably be reserved for Project Shield (ie. clamshell type device), but many of the benchmark results listed above were run on a reference Tegra 4 tablet at MWC, with essentially the same results as what we see in the list above.

Most of the numbers were also for single threaded tests, AFAIK

 

[french toast]

Reference tablet also carried high end sandisk nand flash storage and ddr3L 1866 to go along with a 9w tdp.

Forget about seeing nvidias marketing benchmarks inside any smartphone without handling it with a pair of oven gloves.


Edit..question. .putting costs aside and looking at purely performance metrics...why would you pick a full power tegra 4 in your tablet over say a 10w Haswell??..