Mobile Power - Tegra vs. Cortex A8 vs. Snapdragon

B

bassballboy

Gawd
Joined
Apr 25, 2007
Messages
526
All,

I'm sure this will be a hot topic in the coming weeks, but does anyone with technical know-how have a good comparison of these three systems? From what I've read the Tegra platform has GPU that is ~50% faster than the PowerVR used with the A*, but the A8's CPU is atleast 2x as powerful as the OMP=AP11 CPU used in Tegra.

Any insight on this? It's rather confusing and there's a lot of fanboyism going around.
 
Without having the full tech specs in front of me:

Tegra is very good at media and very power-efficient at media and (I believe) is CUDA based which has proven itself to be a very good system

Cortex A8 and Snapdragon are more your basic low-power CPU, clearly evolutions of the brand.

The most important thing is really how well the software is written for any of those CPU/GPUs, good software will make the best use of the architecture and features a chip offers. This includes the basics of speed and battery life.

Personally I'm most excited about Tegra, mainly because it looks to offer 8-12 hours of HD media playback at smooth rates. However I do believe that this speed is largely due to the CUDA software that's been developed, if the same sort of API was developed for the A8 or Snapdragon, I believe the performance could be similar.

I really can't stress enough that this is going to come down to software. The '1st gen' smartphone CPUs were all in the 400-550 mhz range, good enough for smooth single-app operation. IMO the new CPU in the 3Gs is 'gen 1.5', sort of the same as the HD 4890 is to the 4870. '2nd Gen' is what's coming with Tegra, A8, and Snapdragon, all are ~1 ghz processors, all have more of a media focus (HD is the big thing), and all are promising more battery life (min 24 hours versus the 8ish we get now).
 
Well the Cortex A8 most definitely has a superior CPU vs the Tegra SoC, anandtech did a good write up of the differences between the new and old ARM cores when they did the iPhone 3GS review.... suffice to say, it is *not* a small, or incremental upgrade, it really is a massive generational shift (the Cortex A8 that is) ...

I dont really know much about the Snapdraon unfortunately, but I know HTC is coming out with a beautiful looking phone based on it later this year... so I'll keep my eyes peeled :)
 
Tegra is not just a CPU its a system on a chip.

It has different cores that do different things and are optimized for that task. That is what gives it insane battery life. You can run basic ARM software on it but the good stuff comes from writing software to take advantage of its abilities.

What I am very excited about is its ability to send out 1080P over HDMI while at the same time being able to do other things. It will be absolutely perfect for nettop setups.

Now don't get me wrong. The cortex series is an insanely powerful chip that will be a HUGE player in the market. http://www.theinquirer.net/inquirer/news/1533877/dual-core-arm-chips-apple-tablets

BRITISH CHIP SHOP ARM today announced development of two Cortex A9 dual core processors that are capable of running at up to 2GHz.

The processors, which feature hard macro implimentation, are built using the TSMC 40nm-G process and draw less than 250mw per CPU under full load, making them ideal for embedded applications and portable devices where heat, size and battery drain are crucial issues.
Click to expand...

http://www.itproportal.com/portal/news/article/2009/9/16/could-arms-cortex-a9-defeat-x86-cpus/

Performancewise, the new Cortex-A9 is certainly no slouch with an estimated Coremark score of around 10,500 (for the 2GHz model) which makes it more powerful than a 3GHz Intel Xeon 5160 while consuming a fraction of the power.

Even the power optimised version which runs at 800MHz is as fast as an Intel Core 2 Duo clocked at 1.2GHz with a Coremark score approaching 4000 points. Given that the A9 can be scaled speedwise even further, the next generation Cortex could be even more powerful.

Read more: http://www.itproportal.com/portal/n...ms-cortex-a9-defeat-x86-cpus/2/#ixzz0RI68f6Ts
Click to expand...
 
Snapdragon, Cortex A8 (as implemented on the Pre) and Tegra are all SoC (system on a chip) in that they all have integrated GPU + CPU + I/O, and they're all ARM architectures.

My understanding is the Tegra has the best GPU but the worst CPU of the 3, and from what I've seen, the CPU portion of the Snapdragon and Cortex are about the same, although Cortex is superior in a broad sense (depending on implementation - there are some pretty advanced Cortex implementations and the Snapdragon name will be applied to future implementations that are more advanced than currently available chips).

Tegra is not unique as some are painting it to be - the Open GL performance on the Pre and iPhone 3GS are pretty solid, although my guess is Tegra will still outperform them in applications where OpenGL performance is the main bottleneck. Tegra also has very good video decode / encode support, probably ahead of the other two in this regard, although they are no slouches.

There's a good table in this article to put it in perspective:
http://en.wikipedia.org/wiki/ARM_architecture
 
Last edited:
All three have good things about them as well as some drawbacks, others have posted lots of good info on the hardware so I will point to software again.

The Touch 2 that is replacing my Touch Pro right now has the same internal hardware, but thanks to software optimizing and better utilization of the hardware the new phone is almost twice as fast and snappy in doing anything.

After a day or two of heavy use my memory use is pushing 60% used with nothing open. This bogs down the rest of the phone until I reboot. Clearing bad leaks like that and really writing for the hardware you use can make your phone seem so much faster than it is. Same with anything really but a lot of phones are pushed out quickly...

Partially why the iPhone runs as well as it does consistently.

My next phone will probably be the Snapdragon based HTC Leo (new images and a long video released earlier tonight with more goodies on how it runs)

I just hope the software really takes advantage of it....
 
the Leo looks absolutely amazing....

I tried out a Touch Pro 2 myself, found it unbearably laggy.... the fact that they kept the old (ARM 11 based) internals really screwed that phone over....
 
brucedeluxe169 said:
the Leo looks absolutely amazing....

I tried out a Touch Pro 2 myself, found it unbearably laggy.... the fact that they kept the old (ARM 11 based) internals really screwed that phone over....
Click to expand...

Yeah, should try the original Touch Pro :p If you get one and remove some of the network tied crap the TP2 is faster due to software, but the hardware is no different from the older model so.....
 
"My understanding is the Tegra has the best GPU but the worst CPU of the 3, "

I'm afraid you are missing the point entirely. The Snapdragon chip running at 1GHz is going completely in the wrong direction. These are mobile devices! The ideal situation is to have the CPU running at SLOWER not faster speeds. i.e. 100-200 MHz would be more ideal.

Computational efficiency i.e. bit computation per unit energy gets worse as the frequency increases due leakage across the transistor gates. To increase efficiency you need to reduce speed and go parallel. In other words, the GPU is far more important than the CPU as it has multiple cores. increasing a CPU/GPU speed from 500MHz to 1GHz will improve things by a factor of 2, but at a similiar cost to battery life. Shifting operations from the CPU to the GPU will improve things by typically 30x (depending on the GPU and the parallelism of the code) at no cost to battery life. The next generation of GPU will provides speed-ups of 1000x and more!

The key thing as Arkangyl pointed out is the software. If the software does not take advantage of the parallel coding then it is pointless. Unfortunately the Tegra does not presently support CUDA (only openGL) so the real power of the device will not become apparent for some time.
 
paddyd said:
Unfortunately the Tegra does not presently support CUDA (only openGL) so the real power of the device will not become apparent for some time.
Click to expand...

Exactly why I am not that excited for it yet, love the possibilities but how long must I wait.......

Very well put and explained!
 
paddyd said:
"My understanding is the Tegra has the best GPU but the worst CPU of the 3, "

I'm afraid you are missing the point entirely. The Snapdragon chip running at 1GHz is going completely in the wrong direction. These are mobile devices! The ideal situation is to have the CPU running at SLOWER not faster speeds. i.e. 100-200 MHz would be more ideal.

Computational efficiency i.e. bit computation per unit energy gets worse as the frequency increases due leakage across the transistor gates. To increase efficiency you need to reduce speed and go parallel. In other words, the GPU is far more important than the CPU as it has multiple cores. increasing a CPU/GPU speed from 500MHz to 1GHz will improve things by a factor of 2, but at a similiar cost to battery life. Shifting operations from the CPU to the GPU will improve things by typically 30x (depending on the GPU and the parallelism of the code) at no cost to battery life. The next generation of GPU will provides speed-ups of 1000x and more!

The key thing as Arkangyl pointed out is the software. If the software does not take advantage of the parallel coding then it is pointless. Unfortunately the Tegra does not presently support CUDA (only openGL) so the real power of the device will not become apparent for some time.
Click to expand...

Hold on a second. You are posting quite a bit of misinformation here. Frequency increases leakage, but you can't just compare power efficiency by saying "oh Snapdragon runs at 1ghz, so it wastes more power". It depends on process, transistor density, and the CPU architecture. The Snapdragon @ 1ghz draws 350MW, while cortex a8 @ 600mhz draws ~354MW (<.59mw/mhz). So stop posting bs.

I love it when Joe public criticizes the engineers who make these very popular, well selling ASICS, saying, "Oh the fools, don't the know more X means Y? We need LESS Y (and hence LESS X), and more Z!" It's just damned stupid. If you're so good at making these things, why don't you explain to us how your processor, or the design you're working on, is beter?

Snapdragon & A8 are extremely even with regard to power draw. Not sure about performance, but the 3GS is performing about 50% better in OGL ES 1.1 bench right now (not sure if the latest driver for Snapdragon devices is used here).
 
Well, talking about the 3GS's chipset only makes marginal sense when we consider that most of the video (not 3D) silicon in it *seems* to remain dormant at this point. In terms of what is currently exposed, Tegra is obviously massively superior for video. If the rumours are true, the 3GS is quite a bit stronger though (and unlike the WiiHD/GPGPU, this is actually believable and has some real evidence ). Even if the 3GS was theoretically able to do 1080p HP Decode, it nearly certainly could not do 720p encode - so that's one thing NV wins on at least. Same for the camera, where the 3GS needs to offload much more to the camera module (still no integrated ISP)


As for 3D, Tegra's GPU has the same number of TMUs (2) as the SGX535, nearly certainly more raw ALU power, and a bit lower efficiency (TBDR vs IMR; MIMD vs SIMD shaders). In the end it comes down to drivers and clock speeds, but I have no idea what the 3GS is clocked at...
but the SGX535 should be more powerful than gpu in the Tegra

i don't nou that much about gpu in the Snapdragon so i cant comment on that

you people have to remember that Tegra is overhiped and is overall nothing special
maybe the Tegra 2 will be able compete better till then there are a lots of better SoC like the OMAP3
 
Would be great if there were places to read up on the relative IPC of each architecture, because hearing 550mhz vs 1ghz etc doesn't mean much of anything if we don't know the inherent capabilities of the processor per cycle. Benchmark suites would be helpful...
 
argor said:
Even if the 3GS was theoretically able to do 1080p HP Decode, it nearly certainly could not do 720p encode - so that's one thing NV wins on at least. Same for the camera, where the 3GS needs to offload much more to the camera module (still no integrated ISP)
Click to expand...

wait, what? my old 486 could do 1080p encode...
 
Yeah, but not at any sort of relevant speeds.

As far as NV's Tegra, yeah you're absolutely right. It is definetely slower on the CPU side, but seems to be around 50% "better" on the GPU side (it does something like 50% greater fillrate, and maybe 50% more triangles). I know that ARM11 is theoretically ~ 2x slower per clock, but Tegra clocks 600mhz (at least on Zune), and from benches I've seen on anandtech, real world increase of Iphone 3gs vs Ipod Touch 2g (at 533 mhz) isn't that great (like 30%). So personally, I'd rather have the Tegra at this point.

But with regards to Snapdragon (at least the CPU side), it's a good processor it seems. I don't know a ton about it, but do know that performance is roughly on par with the a8, but at 1ghz. Power draw is the same.

That being said, this press release makes me think that the Cortex A8 has a much higher power draw than .59mw/mhz like is stated. http://www.intrinsity.com/index.php/articles/64-hot-rodding

They're saying that at 1ghz with Fast-14 d.logic, the A8 gets ~ 750mw. But they're also saying this somehow fits within the power profile of Iphone 3gs. So that 354mw is clearly average use, TDP is probably double that. If this article is correct, or maybe it is just intentionally vague and I'm reading into things.

If Tegra with Cortex comes out, and can stay within a similar power envelope as the S5PC100 (SOC that Iphone 3gs uses), I'd be impressed. I can't find that figure anywhere. The Tegra is quoted as having a 1W ceiling.

I have an Iphone 3gs, and love the hardware (except for the lcd, for $600 this should be an OLED), but feel like I have Apple watching my every move. I hate this company. But it is very fast.

Here is an interesting slide from nvidia: http://www.bit-tech.net/news/hardware/2009/06/02/nvidia-s-tegra-cpu-has-a-mighty-battery-lif/1

Tegra @ 1W seems to murder Snapdragon, and by association the S5PC100.

Edit: One confounding factor: http://www.intomobile.com/2009/10/2...comm-snapdragon-and-marvell-armada-oh-my.html

Snapdragon at 1ghz, for the entire chipset, including graphics, is ~350mw average use (with max power I think being 500mw)? If so, then Cortex A8 pales in comparison.

Edit 2: Snapdragon is based on Cortex A8 it seems.

A good discussion, and some technical figures: http://forum.xda-developers.com/showthread.php?t=567401
 
Last edited by a moderator:
akotlar said:
It seems like IPC should be the same between the 2. Snapdragon seems to be based on Cortex A8. http://www.anandtech.com/gadgets/showdoc.aspx?i=3595&p=4
Click to expand...

Well, this is straight from Anand:

"For the most part, ARM’s licensees don’t modify the design much at all. There are a few exceptions (e.g. Qualcomm’s Snapdragon Cortex A8), but usually the only things that will differ between chips are clock speeds and cache sizes."

So while you're correct that they are definitely related, a significantly modified design could significantly change the IPC.
 
brucedeluxe169 said:
wait, what? my old 486 could do 1080p encode...
Click to expand...
in real time j i dont thing so
akotlar said:
but seems to be around 50% "better" on the GPU side (it does something like 50% greater fillrate, and maybe 50% more triangles). I know that ARM11 is theoretically ~ 2x slower per clock, but Tegra clocks 600mhz (at least on Zune), and from benches I've seen on anandtech, real world increase of Iphone 3gs vs Ipod Touch 2g (at 533 mhz) isn't that great (like 30%). So personally, I'd rather have the Tegra at this point.
Click to expand...
:eek: the gpu in the Tegra is somewhat weaker than the SGX535 i did explained little about it in my previous post
Cortex-A8 is about (2.0 DMIPS/MHz
ARM11 is about 1.2 DMIPS/MHz
But with regards to Snapdragon (at least the CPU side), it's a good processor it seems. I don't know a ton about it, but do know that performance is roughly on par with the a8, but at 1ghz. Power draw is the same.

That being said, this press release makes me think that the Cortex A8 has a much higher power draw than .59mw/mhz like is stated. http://www.intrinsity.com/index.php/articles/64-hot-rodding

They're saying that at 1ghz with Fast-14 d.logic, the A8 gets ~ 750mw. But they're also saying this somehow fits within the power profile of Iphone 3gs. So that 354mw is clearly average use, TDP is probably double that. If this article is correct, or maybe it is just intentionally vague and I'm reading into things.

If Tegra with Cortex comes out, and can stay within a similar power envelope as the S5PC100 (SOC that Iphone 3gs uses), I'd be impressed. I can't find that figure anywhere. The Tegra is quoted as having a 1W ceiling.


Tegra @ 1W seems to murder Snapdragon, and by association the S5PC100.
Click to expand...
nop wrong also the 3gs is not using the S5PC100 it is a custom design from apple
most people just assume that it was the S5PC100 as it some what early for a custom design from apple to appear
here is a hint
the S5PC100 use in-house Samsung Graphics core
here are the specks
&#1048707;Rendering performance @Max freq. (133MHz)
•Peak vertex geometry performance (transform only): 9.28M vertices/s
•Vertex geometry performance with single light condition: 7.55 vertices/s
•Shaded fill rate: 125.6M pixel/sec
•Bilinear-filtered textured fill rate with Alpha blending: 37.8M pixel/sec
•Floating-point pipeline & Object-order rendering
•4-Way SIMD vertex shader+ pixel shader
•ShaderModel 3.0: World 1st implementation
•128-bit (32-bit x 4) FP x 1 Vertex Shader
128-bit (32-bit x 4) FP x 1 Pixel Shaders
8-stage pipeline
512 Instruction Slots (configurable)
•Memory BW Optimization by Hierarchical Caching
then the 3gs
SGX535 (28 MPolys/s


Edit 2: Snapdragon is based on Cortex A8 it seems.
Click to expand...
no Qualcomm has a ARM architectural licensee based on ARMv7-A
 
argor said:
no Qualcomm has a ARM architectural licensee based on ARMv7-A
Click to expand...

Like I said it is based on the ARMv7-A. I never said that ARM is selling them the chips, or incorporating them into the Snapdragon SOC. Check the links I posted before you "refute" my information. Snapdragon is a custom ASIC, and the CPU is not identical to the A8, but uses the same instruction set.

http://news.cnet.com/8301-13924_3-10123149-64.html
"Qualcomm is able to achieve this relatively high speed (1.5GHz) for a low-power processor because it did more than simply get a license from ARM. "We went and got an architecture license from ARM. The architecture license was for their new instruction set, the V7 instruction set. There's a difference between getting an architecture license and just getting a core license. A core license means ARM does the (chip) core and they give it to you. The architecture license is different: the actual implementation is your own."
 
quadnad said:
Well, this is straight from Anand:

"For the most part, ARM&#8217;s licensees don&#8217;t modify the design much at all. There are a few exceptions (e.g. Qualcomm&#8217;s Snapdragon Cortex A8), but usually the only things that will differ between chips are clock speeds and cache sizes."

So while you're correct that they are definitely related, a significantly modified design could significantly change the IPC.
Click to expand...

You're right, but I doubt the IPC is significantly different between these two. I suppose they could have increase the # of pipeline stages to get the extra clocks, but I'd hope that they did something more impressive than that. There wouldn't seem to be much of an advantage doing that for such tight power envelopes.

But it sounds like Anand is saying the architecture was not significantly modified, so it could be some rearrangement of logic, or use of something akin to the Fast-14 tech. I think increasing the # of pipeline stages from 13 to something that would allow an extra 60% clock speed would constitute a significant design change wouldn't it? Maybe anand doesn't know all the details?
 
I'm new here, but have done a lot of dev work with ARM based chips.

It really depends on which Snapdragon chip you are referring to as Snapdragon is a brand name and not an exclusively locked chip arch. The current Snapdragons that clock over 600Mhz are definitively based on Cortex A8 cores.

ARM also provides the OpenGL engine IP to it's licensee's. Cortex A8 can not be considered a chip in itself as the licensee is required to layout the end product in a package of their own design.

Personally, I would suggest paying more attention to the upcoming Cortex A9. With the recent explosion of mobile devices it's adaption should be much quicker than the two years it took for the A8.

Sorry for any grammerical errors as I am typing this from my quite snappy 3Gs.
 
You must log in or register to reply here.
Back
Top