NVIDIA Tegra Architecture

[Deleted member 13524]

Right, because the scale doesn't start at 0.5 or 0.8 like the AMD marketing graphs?

No, trusting that graph to compare Krait 300 to Cortex A15 is ridiculous because:

1 - The S600 performance is "estimated" from a S4 Pro. God knows how they "«estimated»" that.

2 - The "estimated" results are probably coming from a shipping smartphone like Xperia Z or Nexus 4, which also carry rather conservatively-clocked LPDDR2 (800-1066MT/s), low-power/slow mass storage and can even throttle down due to thermals.
The Tegra 4 engineering board they used for comparison could (and probably) have some DDR3 2133MT/s, a high-end SSD and a heatsink with a fan to prevent throttling.


Anyone looking to compare Krait 300 to Cortex A15 in application performance should be looking at Galaxy S4's S600 vs. Exynos scores.

Look at the data. Do you think it makes sense that the GS3 (with quad-core Cortex A9 @ 1.4GHz) has the same Sunspider score as the HTC One (with quad-core Krait 300 @ 1.7GHz) if there was no tuning? Do you think it makes sense that the GS4 (with quad-core Krait 300 @ 1.9GHz) has a 38% higher Sunspider score than the HTC One (with quad-core Krait 300 @ 1.7GHz), even though the difference in operating frequency is only ~ 12%?

The HTC One uses dual-channel LPDDR2. At best, it clocks at 1066MT/s so 8.53GB/s total bandwidth.
The Galaxy S4 uses dual-channel LPDDR3, which starts at 1600MT/s so a total of 12.8GB/s.

The difference in total bandwidth between the two is about 50%. Considering they're both handling a huge 1080p resolution and are therefore bandwidth-starved, I'm actually surprised how the Galaxy S4 isn't much faster than the One in almost everything.

 

[french toast]

No, trusting that graph to compare Krait 300 to Cortex A15 is ridiculous because:

1 - The S600 performance is "estimated" from a S4 Pro. God knows how they "«estimated»" that.

2 - The "estimated" results are probably coming from a shipping smartphone like Xperia Z or Nexus 4, which also carry rather conservatively-clocked LPDDR2 (800-1066MT/s), low-power/slow mass storage and can even throttle down due to thermals.
The Tegra 4 engineering board they used for comparison could (and probably) have some DDR3 2133MT/s, a high-end SSD and a heatsink with a fan to prevent throttling.


Anyone looking to compare Krait 300 to Cortex A15 in application performance should be looking at Galaxy S4's S600 vs. Exynos scores.





The HTC One uses dual-channel LPDDR2. At best, it clocks at 1066MT/s so 8.53GB/s total bandwidth.
The Galaxy S4 uses dual-channel LPDDR3, which starts at 1600MT/s so a total of 12.8GB/s.

The difference in total bandwidth between the two is about 50%. Considering they're both handling a huge 1080p resolution and are therefore bandwidth-starved, I'm actually surprised how the Galaxy S4 isn't much faster than the One in almost everything.

Correct..apparently nvidia uses some high end sand disk..they have a deal for dev platforms (the one those optimistic results are from)

@ exophase: only s800 is HKMG..not s600...so greater gains are to be expected with s800.

Havnt got time for links but I posted them a few pages ago I think.

 

[Helmore]

875ms for KRAIT..wow.

Sunspider is too dependent on software, but you already know that.

A Nokia Lumia 920 scores 920 ms for example. That's on a now pretty old MSM9690, dual core Krait 100/200 @1.5 GHz.

Ah well, not very on-topic.

 

[ams]

No, you didn't say IPC advantage.. it just so happens S600 is out in the wild running at 1.9GHz.

Again, the fact that Samsung has a 1.9GHz variant is not important because we were talking about differences in performance per watt for S600 vs. T4. Performance per watt for Krait 300 doesn't materially change with a small bump up in frequency from 1.7GHz to 1.9GHz. And when it comes to application performance, I believe that there is more than enough evidence to show that the quad-core Cortex A15 has up to 1.5-3x performance advantage in a variety of different applications, irrespective of which S600 variant is used for the comparison. Hopefully Anand can confirm this in the coming months.

 

[ams]

No, trusting that graph to compare Krait 300 to Cortex A15 is ridiculous because:

1 - The S600 performance is "estimated" from a S4 Pro. God knows how they "«estimated»" that.

Actually I believe it was Qualcomm who stated that S600 has "up to 30% higher performance" than S4 Pro. Anandtech's data on the HTC One clearly shows that, on average, the S600 @ 1.7GHz does have ~ 30% performance advantage over S4 Pro across a variety of different apps. To compare to a 1.9GHz variant, simply set the S600 scale on the chart to ~ 1.12 instead of 1.0 across the board, that is easy enough to do. Even when used in a tablet form factor (which is bound to happen later this year), it is unlikely that S600 will be using a max CPU operating frequency above 1.9GHz.

 

[Helmore]

My expectations for Tegra 4 are that it'll be around 30-40% faster in CPU intensive tasks and around twice as fast in GPU intensive tasks compared to a Snapdragon S600. Tegra 4 will require at least twice as much power to do so though.

At least with my expectations won't set me up for too much disappointment. I hope. No actually, I think my expectations for GPU performance are a little optimistic.

This is all conjecture on my part of course.

 

[Exophase]

As you are pointing out and many here seem to be missing..tegra 4 reference tablet has a very high tdp headroom, ddr3 1866? And likely other enhancements I cant remember (I mentioned these a few pages back but cant remember the exact details)

I think pretty much everyone but ams is pointing out the big potential difference in testing environment

Seems you're right, Krait 300 (S600) isn't HKMG, Krait 400 (S800) is. So the relevant question becomes as follows: timeline-wise which will be the more suitable comparison, Tegra 4 vs S600 or vs S800?

 

[Helmore]

I think Tegra 4 vs. Snapdragon S800 will be more relevant. One oddity between the S600 and the S800 is that the S800 has an integrated baseband while the S600 has to use a separate baseband. Which is odd, as I thought the S800 is aimed at tablets and very high end phones.

 

[Ailuros]

I think pretty much everyone but ams is pointing out the big potential difference in testing environment

Well he'll also discredit the value of GLB2.7 because for some reason synthetic benchmarks that aim to predict the future should run at at least 50fps average to have any relevance. For GPU performance NV predicted the majority on GLB2.5 for which Adrenos live up to their hw expectations.

Seems you're right, Krait 300 (S600) isn't HKMG, Krait 400 (S800) is. So the relevant question becomes as follows: timeline-wise which will be the more suitable comparison, Tegra 4 vs S600 or vs S800?

Does anyone know when S800 sampled to Qualcomm's partners?

 

[Helmore]

Does anyone know when S800 sampled to Qualcomm's partners?

Early January according to EETimes. They sampled at around the same time as the S600.

Linky:http://www.eetimes.com/electronics-news/4404437/Qualcomm-samples-next-gen-Snapdragon-processors

 

[Jubei]

I think Tegra 4 vs. Snapdragon S800 will be more relevant. One oddity between the S600 and the S800 is that the S800 has an integrated baseband while the S600 has to use a separate baseband. Which is odd, as I thought the S800 is aimed at tablets and very high end phones.

S800 = the artist formerly known as MSM8974 from the old roadmap a few years back

So it was always designed with an integrated baseband in mind. And looking at Qualcomms current PR for the chips, S600 is their "mainstream" chip while S800 is for flagships

 

[3dcgi]

If you leave the trisetup part of a GPU untouched and it'll only increase as much as core frequency it's inevitable that you end up with a question mark what the rather big increase in VS ALUs is actually good for.

The decision to increase VS ALUs without changing the setup rate is likely because Nvidia expects VS's to get more complex rather than games pushing more simple triangles.

 

[Ailuros]

The decision to increase VS ALUs without changing the setup rate is likely because Nvidia expects VS's to get more complex rather than games pushing more simple triangles.

Or they just kept conveniently the 2:1 PS to VS Vec4 ALU ratio T3 introduced. Vertex rate isn't going to increase by a 6x fold within just one year in mobile games.

 

[Ailuros]

http://dxbenchmark.com/phonedetails.jsp?benchmark=dxpro27&D=NVidia+Wayne

Given that it's again probably clocked at somewhere over 300MHz, more than a nice showing in GLB2.7 probably too.

 

[french toast]

http://dxbenchmark.com/phonedetails.jsp?benchmark=dxpro27&D=NVidia+Wayne

Given that it's again probably clocked at somewhere over 300MHz, more than a nice showing in GLB2.7 probably too.

Sorry but I may have missed something? Those are lower than s600...if this is representative of wayne smartphone performance then it is very disappointing from my perspective.

Its also a far cry from nVidia's leaked marketing scores...pretty much how I thought it would turn out to be honest. :/

I know its DX 2.7 but they should be quite similar to gl 2.7.
If anything you would expect tegra to have even better drivers than qualcomm for DX XX.

 

[Helmore]

http://dxbenchmark.com/phonedetails.jsp?benchmark=dxpro27&D=NVidia+Wayne

Given that it's again probably clocked at somewhere over 300MHz, more than a nice showing in GLB2.7 probably too.

How do you know it's clocked at those speeds? I can't find anything about it's clocks.

 

[french toast]

Ive comprised a comparison of differing gpu uarchs.
http://dxbenchmark.com/compare.jsp?...P8510+ATIV+Tab&D7=Microsoft+Surface+RT&cols=7.

If that is at 300mhz as ailuros alludes to then it actually is very good in DX benchmark..anywhere between 2-3.5x better than tegra 3...

It trounces adreno and sgx 5 series. .which is to be expected considering nVidia's superior microsoft drivers.

 

[OlegSH]

How do you know it's clocked at those speeds? I can't find anything about it's clocks.

look at filtering rate numbers, they are quite low compared to theoretical peak at 670 Mhz ( 670 * 4 TMU = 2680 MTexels/s ), I don't say that theoretical numbers are achievable, but the real numbers should be at least two times higher than 516 MTexels/s in DXBenchmark low level tests

 

[Helmore]

look at filtering rate numbers, they are quite low compared to theoretical peak at 670 Mhz ( 670 * 4 TMU = 2680 MTexels/s ), I don't say that theoretical numbers are achievable, but the real numbers should be at least two times higher than 516 MTexels/s in DXBenchmark low level tests

I guess you're right. Tegra 3 scores 334 MTexels/s, which makes 516 MTexels/s for Tegra 4 seem rather low.

BTW, Snapdragon MSM9690 with an Adreno 225 scores 133 MTexels/s in the same test. I'm taking the scores for Tegra 3 and Adreno 225 from the Surface RT and the Dell XPS 10. I couldn't find and Adreno 320 scores on DXBenchmark.

 

[Ailuros]

How do you know it's clocked at those speeds? I can't find anything about it's clocks.

Because the DXB2.5 1080p score is at just 26.8 fps while NV has mentioned for GLB2.5 57 fps. While there might be slight differences in performance between the DX and OGL_ES version of the benchmark I don't expect them to be in the >2x fold ballpark.

It's the Convington dev platform and the GPU should be clocked at around 315MHz on estimate; scale up to 672MHz and you get the possible DX2.7 score for it.