Article: Inside the NVIDIA APX 2500

[B3D News]

NVIDIA has just announced the 65nm APX 2500, an application processor supporting 720p H.264 video, OpenGL ES 2.0, and HDMI output. On the processing side, it sports an ARM11 core at 750MHz. We had a quick chat with Mike Rayfield and touched on a variety of subjects and interesting design choices...

Read the full article

 

[mczak]

Read the full article

Intereseting read. Though the talk about "extremely low power consumption" is pretty much meaningless since you can't get any comparisons to the competition.
And, the cpu is not really twice as fast as the iPhone cpu as such, the iPhone uses an arm 11 too and clocks it at 400Mhz, but the cpu is rated for 620Mhz it seems - so at maximum clocks the APX 2500 would only be marginally faster (and, of course, we've no idea how power consumption would compare at these clocks, obviously the reason the iPhone doesn't clock it that high).

 

[Arun]

Intereseting read. Though the talk about "extremely low power consumption" is pretty much meaningless since you can't get any comparisons to the competition.

Broadcom's BCM2727 is claimed to be capable of 5 hours of 720p video playback. It's not clear what codec that refers to though, and whether it includes the screen or was measured via HDMI streaming. As for NVIDIA's other direct competitors - the OMAP3430/3440 and the STw4820 - none of them release any power consumption number at all, so who knows. I'd tend to argue however that if it was lower, they'd be bragging about it right now...

And, the cpu is not really twice as fast as the iPhone cpu as such, the iPhone uses an arm 11 too and clocks it at 400Mhz, but the cpu is rated for 620Mhz it seems - so at maximum clocks the APX 2500 would only be marginally faster (and, of course, we've no idea how power consumption would compare at these clocks, obviously the reason the iPhone doesn't clock it that high).

Rated at 620MHz? Not at all; that rumour comes from ARM's claims for ARM11 in general, which is just an indication for potential licensees. That doesn't mean the specific chip in the iPhone could reach that frequency - it very likely could not. However, I was also wrong and the iPhone's correct frequency is 412MHz. Strange, my initial googling turned out a reliable source with a number in the 300s - oh well, fixed!

 

[mczak]

Rated at 620MHz? Not at all; that rumour comes from ARM's claims for ARM11 in general, which is just an indication for potential licensees. That doesn't mean the specific chip in the iPhone could reach that frequency - it very likely could not.

That doesn't seem to be true, it looks like this incarnation of the arm 11 core (a samsung S5L8900 apparently) really DOES support 620Mhz (some sources claim 667Mhz, but I was not able to find sources which claim anything lower than 620Mhz). Of course, maybe it's not true there's no official data available on that chip (or it could be true but the chip would need higher voltage for that). But anyway, if others can produce arm 11 cores running at 667Mhz easily (samsung S3C6400 for instance, 90nm), one at 750Mhz just doesn't sound like that much of an achievement (only if it has lower power draw, which we don't know).

However, I was also wrong and the iPhone's correct frequency is 412MHz. Strange, my initial googling turned out a reliable source with a number in the 300s - oh well, fixed!

Well, the original clock was apparently 400Mhz, got raised with 1.1.2 firmware to 412Mhz.

 

[Arun]

That doesn't seem to be true, it looks like this incarnation of the arm 11 core (a samsung S5L8900 apparently) really DOES support 620Mhz (some sources claim 667Mhz, but I was not able to find sources which claim anything lower than 620Mhz). Of course, maybe it's not true there's no official data available on that chip (or it could be true but the chip would need higher voltage for that).

I can't find any official or at least highly reliable evidence of that. Don't trust everything you read on the internet... (there's another mobile phone that some claim use the same SoC and it's clocked in the 500s there, but I'm not very vonfident in the reliability of that claim...)

But anyway, if others can produce arm 11 cores running at 667Mhz easily (samsung S3C6400 for instance, 90nm), one at 750Mhz just doesn't sound like that much of an achievement (only if it has lower power draw, which we don't know).

Given some of the frequencies I've seen on 90nm, it is true that I wouldn't have expected 750MHz to be so unusual on 65nm. As I said though, it turns out that this is the highest-clocked ARM11 that was announced so far on 65nm, so in practice it's definitely quite high!

Well, the original clock was apparently 400Mhz, got raised with 1.1.2 firmware to 412Mhz.

Yeah, still doesn't explain why I thought it was in the 300MHz ballpark though!

 

[crystall]

As for NVIDIA's other direct competitors - the OMAP3430/3440 and the STw4820 - none of them release any power consumption number at all, so who knows. I'd tend to argue however that if it was lower, they'd be bragging about it right now...

The power consumption of a single SoC is not a very useful metric considering the differences between them. However a solution which requires less chips for building a whole phone is almost always going to be less power hungry than one which requires more given a more-or-less comparable feature set. Off-chip comunication requires a lot of power by embedded standards.

 

[RussSchultz]

acquired PortalPlayer, the two had very different architectures for multimedia processing. The former used extended Tensilica cores while the latter used a DSP attached to fixed-function units. The former is just like AMD, while the latter is much more similar to Texas Instruments’ approach. And it turns out that the DSP approach won. Of course, we presume quite a few implementation details changed.

AFAIK, Portal had 2 ARM7's in their chips, and no DSP attached to fixed function units.

Though I never paid much attention to the product that came out right when NVIDIA purchased them, which probably had video hardware as fixed function unit or accelerator (from Hantro/On2, maybe?)

IMHO, the audio processing work is very very very very minimal (<25MIPs on an ARM11) so putting in dedicated hardware to handle that is just going to complicate the design for very little benefit. (Particularly when the real 'problem' of audio decode isn't the algorithm, but the data handling, including play queues, FFWD, RWND, etc, and neither the tensilica solution, or dedicated hardware will give you that.

 

[Arun]

The power consumption of a single SoC is not a very useful metric considering the differences between them. However a solution which requires less chips for building a whole phone is almost always going to be less power hungry than one which requires more given a more-or-less comparable feature set. Off-chip comunication requires a lot of power by embedded standards.

No offense intended, but that's wrong and/or way too simplistic. The power consumption for video has absolutely nothing to do with the level of integration (unless you count SRAM/eDRAM buffers as 'integration' but I don't); it's all process and architecture. Pretty much every single SoC in the world released in the last few years has all the video decoding logic and the video output blocks on the same chip, and the differences in power consumption are absolutely huge between different products. Remember: in terms of non-wireless digital logic, all of these solutions are 'true' single-chips.

AFAIK, Portal had 2 ARM7's in their chips, and no DSP attached to fixed function units.

Though I never paid much attention to the product that came out right when NVIDIA purchased them, which probably had video hardware as fixed function unit or accelerator (from Hantro/On2, maybe?)

Yeah, they changed their technology platform completely and it's because their new technology impressed NVIDIA that they got acquired, AFAIK. The GoForce 6100 taped-out a short time before the acquisition. The fixed-function stuff is definitely in-house, but I think the DSP might be licensed. My guess is the DSP(s) in the APX 2500 is in-house too (Stexar...), but I didn't ask that explicitly.

IMHO, the audio processing work is very very very very minimal (<25MIPs on an ARM11) so putting in dedicated hardware to handle that is just going to complicate the design for very little benefit.

I'm not sure how much of it is done in fixed-function units, my guess would be it's fairly DSP-centric but who knows. Either way, the fact remains that going from 50 hours of battery life for audio to 100 hours and eventually 200 hours - well, every single milliwatt starts making an impact. Even if it might look like overkill, it might not be if you've got enough units to amortize your R&D cost over.

 

[Npl]

wonder if Sony is (partially) funding the research in expecting a future PSP2 Chipset from them. Would suck for them if now Competitors get a headstart on Sony from their own money

But then I expect the PSP2 GPU to be more powerful and having a seperated CPU (MIPS-Derivate - Dual 74K maybe? ). But its still the same branch of Media-Handheld this chip is going to cover.

 

[crystall]

No offense intended, but that's wrong and/or way too simplistic. The power consumption for video has absolutely nothing to do with the level of integration (unless you count SRAM/eDRAM buffers as 'integration' but I don't); it's all process and architecture. Pretty much every single SoC in the world released in the last few years has all the video decoding logic and the video output blocks on the same chip, and the differences in power consumption are absolutely huge between different products. Remember: in terms of non-wireless digital logic, all of these solutions are 'true' single-chips.

Maybe I missed your original point but I was thinking of the whole system power consumption, not only that required for video decoding. And that's because in a smartphone, or a PDA, you're going to have other parts working while you watch a movie.

 

[Arun]

Maybe I missed your original point but I was thinking of the whole system power consumption, not only that required for video decoding. And that's because in a smartphone, or a PDA, you're going to have other parts working while you watch a movie.

Oh, well, you were replying to a paragraph where I was basically only talking of 720p video decoding so I didn't think you meant to be more general than that...

Either way, I don't really see what difference there is between the APX 2500, the STn8820 and the OMAP3 Family there. All the application/multimedia processing digital logic is on a single chip, and the power management & audio analogue is off-chip. You could make a case that application processor + baseband integration saves a little bit of power, but the level of integration for digital logic is really similar for all the products we're discussing here; the difference is only in the analogue part of the equation.

 

[tangey]

The author makes a mistake in comparing this Nvidia Chip with TI's 3430, and concluding that the TI chip is just marginally quicker. The 3430 was announced over a year ago and will start appearing in phones 1st half of this year. A more correct comparsion would be with TI's recently announced 3440 chip which clocks the A8-core at 800Mhz instead of the 550Mhz in the 3430. This gives the TI chip roughly twice the performance of the Nvidia one, along with an SGX core from PowerVr, which I suspect will be the SGX540.

 

[Arun]

Just in case that's not completely obvious, I'm the author - anyway, I disagree wisth that position because: 1) I suspect TI isn't going to stop trying to sell the OMAP3430; they might emphasize the 3440, but the former should still be cheaper. 2) The 3440 will start sampling in Q2, while the APX 2500 and the STn8820 are both sampling today. So that's 4-5 months behind ST & NV...

I certainly agree that the performance comparison is a very hard thing to do fairly; I've tried to be objective without being overly verbose, but apparently you don't think that was a success! Thanks for the feedback anyhow though I'm definitely very curious what SGX is present in the OMAP3440 BTW, if anyone knows about that I'd love to know even if it had to be off-the-record.

 

[tangey]

HI,
I didn't realise you were the author !

Yes, TI will continue to push the 3430, I was thinking of the "fairer comparision" in terms of Nvidia stated '09 for mobile phones and I imagine that the 3440 will be a similar timeframe for phones.

The SGX540 is of course a guess of mine, based on some not totally unconnected "happenings".

1) TI licensed another SGX core from IMG in Oct last year
2) In and around the same time IMG started talking about 540 for the 1st time
3) IMG states that 540 is for "very high-performance mobile phones, in-car, DTV, and mobile computing devices"
4) In the same website update that TI put up the OMAP3440, they also put up a page specifically for Mobile Internet Devices.

 

[Ailuros]

According to TI's website there's OMAP 3410, 3420, 3430 and 3440. Lowest end SGX510 has been scratched afaik. Albeit just some funky speculation from my side:

3410 = SGX520
3420 = SGX520
3430 = SGX530 or 535
3440 = SGX535 or 540

The difference between 535 & 540 should be merely in amount of units (ie performance) IMHO, since up to 540 Direct3D compliance is marked at D3D9.0L and only =/>SGX545 is D3D10.1. I wouldn't be in the least surprised though if 520 is twice as fast than MBX Lite+VGP and 530 twice as fast as MBX+VGP.

Anyway why would anyone make any performance estimates without knowing further details anyway? The so far supplied data tells me that the APX 2500 is mostly concentrating on multimedia. The first question which exact 3D capabilities the graphics part of that thingy has, before even someone touches the performance part. IMHLO the tiniest SGX starts at SM3.0 and if I'm not mistaken that's a lot more than OGL_ES2.0 requires.

Once some of the above details have been clarified, we can then expand to performance or more specifically performance/capabilities/power consumption ratios.

Finally CSAA on APX2500 sounds highly interesting; I assume it comes with some portion of MSAA. If there's something like an absolute baseline mode (something like 2xMSAA + =/>4xCSAA) the bandwidth and memory footprint consumption for something like that might be highly competitive to what a TBDR would consume with Multi- or Supersampling.

 

[Arun]

Yes, TI will continue to push the 3430, I was thinking of the "fairer comparision" in terms of Nvidia stated '09 for mobile phones and I imagine that the 3440 will be a similar timeframe for phones.

Sure, that's a fair point. However, I'd argue that there could still be new design wins for the OMAP3430 decided today - and NVIDIA is competing directly against TI there to get those design slots. I agree that OMAP3440 is a direct competitor, but so is the 3430 IMO. Regarding SGX, if that license was only in October 2007 then this seems much more likely to be aimed at the OMAP4 family (Mobile World Congress 2009?) than the OMAP3440. We'll see!

Ailuros: That sounds fairly reasonable (although I wouldn't exclude the 530 completely for the 3440 *if* that's also what's used in the 3430). Anyhow, I'll make sure to ask NV about Handheld CSAA when I can.

 

[tangey]

According to TI's website there's OMAP 3410, 3420, 3430 and 3440. Lowest end SGX510 has been scratched afaik. Albeit just some funky speculation from my side:

3410 = SGX520
3420 = SGX520
3430 = SGX530 or 535
3440 = SGX535 or 540

The difference between 535 & 540 should be merely in amount of units (ie performance) IMHO, since up to 540 Direct3D compliance is marked at D3D9.0L and only =/>SGX545 is D3D10.1. I wouldn't be in the least surprised though if 520 is twice as fast than MBX Lite+VGP and 530 twice as fast as MBX+VGP.

3410 doesn't have any 3D acceleration.

The 3420 and 3430 most likely have the same core as they were announced long before the 2nd SGX core license from IMG was anounced.

3440 either has the new core, or is using the same core as the 3420 and 3430.

 

[Loewe]

IMHO, since up to 540 Direct3D compliance is marked at D3D9.0L and only =/>SGX545 is D3D10.1.

I think thats not right!
All SGX cores support OpenGL 2.0, OpenGL ES 2.0 and Direct3D Mobile. AFAIK the last is something like DX7/DX8, I would say at max SM2.0.

Only SGX535 is DX9 and SGX545 is Direct3D10.1. I think both cores especially made for one vendor, who need such cores for windows systems.
Yes I know NEC use the SGX535 for the NaviEngine1.

 

[roninja]

For Omap 3440 my guess would be SGX 540. 3430 and 3420 most likely use the same core either 530 or 535.

Where that leaves OMAP 4xxx is anyones guess but the Cortex A9 core is not due till 2010 AFAIK.

Did not know 510 was canned as it has only made brief appearances.

 

[Loewe]

I don't know for sure but I think Omap3420 and 3430 both will be SGX530. The 535 is to big for this cores, there are to many transistors they don't need.
4340 I think SGX540, I can't give reason for this.

And yes there is no SGX510, I have the confirmation from PowerVR.