Apple A5X SoC

[ChronoReverse]

Why are people even discussing whether A5 is faster than Tegra3 in graphics? Isn't this indisputably fact? Is there even a single situation where the Tegra3 cleanly beats the SGX543MP2?

 

[lothar]

Why are people even discussing whether A5 is faster than Tegra3 in graphics? Isn't this indisputably fact? Is there even a single situation where the Tegra3 cleanly beats the SGX543MP2?

I don't see anyone debating that in this thread.

The question is...Is SGX543MP4 4x faster than Tegra3 as Apple's chart claimed?
IMO, it will probably be 2-3x max faster. Nowhere near 4x.
But it's still faster than anything on the market by far, yes.

 

[Mopetar]

I don't see anyone debating that in this thread.

The question is...Is SGX543MP4 4x faster than Tegra3 as Apple's chart claimed?
IMO, it will probably be 2-3x max faster. Nowhere near 4x.
But it's still faster than anything on the market by far, yes.

Of course it's 4x faster, for at least some specific benchmark(s). Otherwise it's probably somewhere between 4x and no difference. This is no different than any other device manufacturer picking the best results to use when presenting their product.

 

[lothar]

Of course it's 4x faster, for at least some specific benchmark(s). Otherwise it's probably somewhere between 4x and no difference. This is no different than any other device manufacturer picking the best results to use when presenting their product.

That's the same thing I mentioned in my earlier post.

http://forums.anandtech.com/showpost.php?p=33115920&postcount=51

I don't get the point of doing "Offscreen" benchmarks...
If "Offscreen" benchmarks are the holy grail, then why don't we start using them in benching CPUs/GPUs instead of doing numerous resolution changes, numerous AA/AF/HDR changes, numerous game platform changes, and so on?

It seems to me that "Offscreen" benchmarks benches theoretical performance, not actual performance.

It's safe to say that Apple just took a single benchmark that had 4x performance out of the dozens of other benches they tested.
Nvidia does that a lot on the PC discrete graphics side too. In fact, they've been doing that for decades. It's an art for them.

 

[poofyhairguy]

Whichever the case, A5 is faster than Tegra 3.

For the GPU. Lets be clear, it is faster than Tegra 3 on the GPU.

I have seen my Tegra 3's four cores going at it at 1.6ghz- the A5 can't compare.

 

[soccerballtux]

It remains as useless, though. We don't use our computers, tablets, and smartphones without seeing the output.

I'm blind, you insensitive clod.

 

[soccerballtux]

Uh no, that's not what you were arguing before. You were whining that the high-res display needs a more powerful GPU, and that Apple was somehow being deceptive by not mentioning that fact. Then when someone brought up the fact that Apple does indeed mention that in their promo video, you whined some more about how it needs to be "omnipresent" in their marketing. I was pointing out how completely ridiculous that suggestion is.

Oh well. I'm sure if Samsung or Asus released a $500 tablet with an insane 3.1 megapixel screen and GPU equal to the Playstation Vita's (or hell, even one that managed to match the A5's), everybody here would be trumpeting it to the heavens.

If Samsung or Asus did that I'm sure it would be laggy as hell because of Android's inefficiencies at rendering the UI...

 

[soccerballtux]

I would assume that Tegra 3^H^H^H^H ANDROID would fall flat on its face running 1536p. As for A5X, Anand predicts similar performance to the iPad2 A5 at 720p, but a little slower.

ftfy

 

[runawayprisoner]

Just noticed something interesting in Anand's iPad (3) analysis:

He mentioned Tegra 3 GPU was running at 500MHz on the Transformer Prime and A5 GPU was running at 250MHz.

At which point, the performance of each chip is thus: Tegra 3 12GFLOPS, A5 16GFLOPS.

That translates to about a 30% difference, which is shown in GLBenchmark nicely. (though for some reason, the Pro benchmark still shows a roughly 80% difference)

Supposing Apple ups the clock speed of A5X GPU to... say... 333MHz, what happens? The peak theoretical performance of A5X GPU jumps to 42GFLOPS. On paper, that means compared to tablet Tegra 3, it'll be:

42 / 12 = 3.5x

Or just about over 3x the performance of Tegra 3 at 500MHz.

But why 333MHz for GPU clocks? Recall the A5 running its GPU at 250MHz. A5 runs LPDDR2-800, so the bus width for RAM should be 200MHz. Assuming the ratio between GPU clock vs MEM clock at 1.25.

If A5X runs LPDDR2-1066, that means bus width for RAM is about 266MHz. Use the previous ratio and we get GPU clock speed of 333MHz, which fits perfectly into the scheme.

At 333MHz GPU clock, Apple can still assume the same CPU clock of 1GHz just fine (333MHz x 3 = ~1GHz), so it all makes sense if they want to go that route.

What do you think? Is 333MHz the likely clock speed for A5X GPU?

Also, consider that A5 runs a lower resolution natively (1024 x 768 vs 2048 x 1536), that would explain why A5X vs A5 is not more than a 2x difference.

 

[Mopetar]

That's a possibility, but it just seems more plausible that it's still at 250 MHz and that Apple based their comparison off of a single benchmark. It's a much simpler explanation.

 

[runawayprisoner]

I'd think they wanted to do more than just adding 2 extra GPU cores in there. A bump in memory bus wouldn't be too far off either, considering they are doubling RAM capacity.

Even if Apple based their comparison off of a single benchmark, an iPad 3 at 2048 x 1536 simply can't outperform an iPad 2 at 1024 x 768 by a factor of 2x if both run at native resolution. That's quite a tall order.

Not to mention that at the same clock speed, they're assuming completely perfect 100% scaling in performance... for 4 GPU cores!? I'm sure nVidia or ATI would kill to have that kind of scaling for their SLI and Crossfire tech.

 

[Mopetar]

Even if Apple based their comparison off of a single benchmark, an iPad 3 at 2048 x 1536 simply can't outperform an iPad 2 at 1024 x 768 by a factor of 2x if both run at native resolution. That's quite a tall order.

But that's the trick. They never claimed that the new iPad would be twice as powerful as the previous generation. Once again they've merely pointed to a benchmark that shows the A5X's GPU is twice as powerful as the GPU found in the A5. They're not claiming anything beyond that, but people get easily fooled and infer far beyond the claims that Apple has made. It's the same reason that other companies use the same tactic. Most consumers latch on to the one piece of information without fully considering precisely what it means.

 

[runawayprisoner]

But that's the trick. They never claimed that the new iPad would be twice as powerful as the previous generation. Once again they've merely pointed to a benchmark that shows the A5X's GPU is twice as powerful as the GPU found in the A5. They're not claiming anything beyond that, but people get easily fooled and infer far beyond the claims that Apple has made. It's the same reason that other companies use the same tactic. Most consumers latch on to the one piece of information without fully considering precisely what it means.

If this was Apple of 2009, I would agree with you, but Apple from 2010 onward has maintained a pretty good track record of showing fairly accurate and to the point benchmark numbers.

Just like when they said the GeForce 320M would be faster than the older GeForce 9400M, it was faster.

And also just like when they quote 10 hours of battery life on the iPad, it's indeed 10 hours.

There really is no reason for them to release wrong numbers. Even Anand confirmed that Apple was right when they said the A5 was 7x faster than the A4.

If they release wrong numbers, it'll just cause backlash for them. In that case, if the A5X wasn't any faster than the A5, I'd think they would just shut up about it.

 

[Mopetar]

Yes, they always say up to X% better, which was also true for at least one benchmark. If you look at Anand's review of the iPad 2 you'll see that yes, in at least one case it really is seven times faster, but in a lot of others it's not nearly that good.

Apple has merely stated that the A5X is more powerful than the A5 and haven't even made any claims about overall performance, which I'm not even worried about because Apple wouldn't ship it if they thought it would provide a miserable user experience.

The simplest, and therefor the most likely, explanation is that it's just two more graphics cores running at the same speed. This assumption could certainly be wrong, but as there's no solid evidence at this time to suggest that it's faster, there's no real reason to accept an alternative that requires a lot of assumptions.

 

[runawayprisoner]

Well, that's why we have a discussion. It's to speculate.

If all we have to do is just accept it for what it is, then I guess I can just wait another week and tell you what the clock speed is after I run OpenGL Extension Viewer on my iPad 3, but hey... where's the fun in that?

I do recall us having a discussion last year about the iPad 2's A5, too. Probably not here on Anandtech, but I assumed the GPU in the iPad 2 would be running at either 250MHz or 300MHz. I wasn't too far off the mark. But that's enough about me.

Even if we were to assume the same 250MHz clock speed, I think it's interesting to note that that's still faster than 200MHz on the PS Vita.

If Anand's assumptions are right, upscaled games may become the norm, but then I don't doubt that the iPad 3 is somewhat (slightly) more capable of outputting better graphics than the Vita. Assuming it's the same clocks as the iPad 2, that is.

 

[Steelbom]

Well, I think it's about time we take our minds off the iPad 3 for a second, and... discuss its SoC!

So, what do you think of the new dual-core king? Apple boasts 4x graphics performance of Tegra 3, which is just about right but in my opinions, not so much considering they threw A5 being 2x as fast as Tegra 3 in there. Isn't A5 only about 1.5x faster than Tegra 3?

Also another thing I wanna discuss is... how much RAM do you think this thing has? Mike Capp of Epic Games blurted out during Infinity Blade Dungeons demo that it had "more memory than PS3 or XBox 360". So if I were to trust that, it means this thing definitely has more than 512MB of RAM. But how much more?

The screen has 4x more pixels. I'm pretty sure that that would translate to a 4x increase in interface image size as well, unless I'm missing something. So how much RAM should it have to maintain all those high-res graphics assets at 60fps smooth?

It's definitely got 1GB RAM, but it shader bound tests it's going to be more like 2.5x faster than Tegra 3 not 4x as Apple has claimed. (They're probably comparing it against texture bound scenarios.)

No confirmation of clock speeds on A5X yet. But I'm sure it'll be the same 1GHz, otherwise they would have touted it as a "new feature".

I'm sure PowerVR SGX543MP4 would leave Tegra 3 behind a sizable amount. But I'm just not sure it's a 4x increase.

On a side note, I'm actually wondering how the iPad 3 stacks up to Vita in terms of graphics performance. Based on what Epic Games showed during the Vita presentation, and what they showed today with Infinity Blade Dungeons, my wild guess is that... iPad 3 may just pull ahead. I have seen what the Vita is capable of, and I'm honestly not very impressed. But it might just be that current Vita games are not so well-optimized yet.

Regarding RAM, I guess 1GB is reasonable. I was kind of hoping for more, though, since it's a 4x increase in pixel count, and thus graphics asset size.

The iPad 3 won't compare against Vita at all. They may be running the same GPU but the Vita may be running it at a higher clock speed, on top of that, the Vita is powering a 960x544 display, and the iPad 3 is powering a 2048x1536 display which is over six times as powerful.

Just noticed something interesting in Anand's iPad (3) analysis:

He mentioned Tegra 3 GPU was running at 500MHz on the Transformer Prime and A5 GPU was running at 250MHz.

At which point, the performance of each chip is thus: Tegra 3 12GFLOPS, A5 16GFLOPS.

That translates to about a 30% difference, which is shown in GLBenchmark nicely. (though for some reason, the Pro benchmark still shows a roughly 80% difference)

Supposing Apple ups the clock speed of A5X GPU to... say... 333MHz, what happens? The peak theoretical performance of A5X GPU jumps to 42GFLOPS. On paper, that means compared to tablet Tegra 3, it'll be:

42 / 12 = 3.5x

Or just about over 3x the performance of Tegra 3 at 500MHz.

But why 333MHz for GPU clocks? Recall the A5 running its GPU at 250MHz. A5 runs LPDDR2-800, so the bus width for RAM should be 200MHz. Assuming the ratio between GPU clock vs MEM clock at 1.25.

If A5X runs LPDDR2-1066, that means bus width for RAM is about 266MHz. Use the previous ratio and we get GPU clock speed of 333MHz, which fits perfectly into the scheme.

At 333MHz GPU clock, Apple can still assume the same CPU clock of 1GHz just fine (333MHz x 3 = ~1GHz), so it all makes sense if they want to go that route.

What do you think? Is 333MHz the likely clock speed for A5X GPU?

Also, consider that A5 runs a lower resolution natively (1024 x 768 vs 2048 x 1536), that would explain why A5X vs A5 is not more than a 2x difference.

Apple's claimed that the iPad 3's GPU is twice as powerful as the A5 which tells us clock speed is the same. (Remember, they never said it performed at native resolution twice as well, only that it's twice as powerful, as it has doubled how many SGX543's they have in the A5.)

>>>>>

We'll definitely see 1GB of RAM. There's plenty of evidence of that. Something I'm trying to figure out is whether or not the iPad 3 has enough memory bandwidth to run games at native resolution, Anand doesn't believe so and it's quite possibly and sadly true. Games of course would still be rendered at a resolution higher than 1024x768 which would still look better than they do on the iPad 2, but I'm hoping we'll see games at 2048x1536. I think, considering Infinity Blade: Dungeons, and Sky Gamblers: Air Supremacy, that it'll work out just fine. They both look great, so if they can get better graphics looking awesome on the display, I'm happy.

 

[dagamer34]

I like that Apple made sure to up the GPU for its Retina Pad. I was scared that they would pull another iPhone 4- up the res with the same GPU. It is kinda disappointing they didn't throw in two mores cores, or better/faster cores but in iOS I guess that isn't as needed as in Android.

What is much more interesting to me is the single-core AppleTV. That tells me that now that Apple is getting closer to the fab process by moving away from Samsung THEY have to find something to do with all the silicon that is not 100%. We don't know yet but I bet the AppleTV is an A5 with a (faulty) core disabled. That plus the fact that the iPad 2 is sticking around at a lower price point might indicate that Apple is planning on making the fab switch with A5 parts instead of waiting for A6.

Finally there is the discussion about what this new iPad means for the next iPhone. I disagree with Anand that the A5X will be in the iPhone 5- the iPhone doesn't need the same beefy GPU (unless again we are doubling the resolution for it) especially when it will hurt the battery life so much.

That implies that the iPhone 5 will be much like the iPhone 4- almost no upgrade on the SoC side (maybe even no upgrade except for the fab process) but a different form factor.

I wouldn't assume that the 2nd core is faulty, just disabled. The Apple TV has no real use for a 2nd core, since there aren't any major secondary CPU tasks (in fact the only primary task is the UI thread and downloading of the movie since the actual playback is handled by a hardware decoder). However, stating that it's a dual core chip means you can't use any chips where only a single core isn't working. I half expect some chips to have both CPUs working perfectly.

The only reason why Apple would bother telling us that there is only one core enabled is if they intended to had an SDK in the future, otherwise they could have just said "A5" and none would be the wiser.

 

[runawayprisoner]

The iPad 3 won't compare against Vita at all. They may be running the same GPU but the Vita may be running it at a higher clock speed, on top of that, the Vita is powering a 960x544 display, and the iPad 3 is powering a 2048x1536 display which is over six times as powerful.

Actually, there are indications that the PS Vita's GPU runs at 200MHz to save battery life, because any higher and it'll drain battery life faster and make the handheld miserable.

If the iPad 3 runs at the same clocks as the iPad 2, which is 250MHz, then it has a 20% advantage over Vita.

Also iPad 3 games don't have to render at 2048 x 1536. They can be rendered at a lower resolution, like 800 x 600, and then upscaled. The same thing is done on the Vita because Vita doesn't have enough VRAM (clocks in at only a measly 128MB) whereas on the iPad 3, since the GPU shares system RAM, it has almost 1GB at its disposal.

 

[Steelbom]

Actually, there are indications that the PS Vita's GPU runs at 200MHz to save battery life, because any higher and it'll drain battery life faster and make the handheld miserable.

If the iPad 3 runs at the same clocks as the iPad 2, which is 250MHz, then it has a 20% advantage over Vita.

Also iPad 3 games don't have to render at 2048 x 1536. They can be rendered at a lower resolution, like 800 x 600, and then upscaled. The same thing is done on the Vita because Vita doesn't have enough VRAM (clocks in at only a measly 128MB) whereas on the iPad 3, since the GPU shares system RAM, it has almost 1GB at its disposal.

Ah, I see. That's interesting...

That's true, but I'd not want them rendered at 800x600. At least 1024x768, but preferably 1536x1152 or even higher.

The iPad 2's memory read speed is only 330MB/s, that'd still limit the quality of the image significantly right?

 

[runawayprisoner]

Actually, iPad 2 uses LPDDR2-800 in dual-channel config, which gives it 12.8GB/s theoretical bandwidth. That's the max access speed. If it seems slower in benchmarks, it's likely because the benchmark is measuring the CPU's calculative performance, so 330MB/s means the iPad 2's CPU can only calculate 330MB worth of data per second, but if it's just purely writing data, then the RAM should be able to take 12.8GB per second.

 

[Steelbom]

Actually, iPad 2 uses LPDDR2-800 in dual-channel config, which gives it 12.8GB/s theoretical bandwidth. That's the max access speed. If it seems slower in benchmarks, it's likely because the benchmark is measuring the CPU's calculative performance, so 330MB/s means the iPad 2's CPU can only calculate 330MB worth of data per second, but if it's just purely writing data, then the RAM should be able to take 12.8GB per second.

Mmm, interesting... thanks for the info. So there should be sufficient bandwidth to handle games at 2048x1536 with good graphics?

 

[BenSkywalker]

The GPU being twice as fast as the iPad 2 leads to the obvious question- WTF are they thinking?

Exclude that any other mobile devices exist from this argument, it doesn't matter.

The new iPad has *four* times the pixel count of the previous. If we take everything Apple says as gospel truth then the new iPad has two choices- run things at half the speed of the 'old' model, or scale the resolution which completely negates the advantage of having the higher resolution display in the first place.

Again, forget anything not made by Apple exists in the world. How is moving to a new iPad an upgrade over the iPad 2 for any sort of GPU intensive task?

The same thing is done on the Vita because Vita doesn't have enough VRAM (clocks in at only a measly 128MB)

I keep seeing the bar lowered for how little people know about GPUs on this forum- the degree of ignorance is honestly astounding.

http://www.anandtech.com/show/429/13

The fastest boards listed in that chart have 32MB of RAM. 128MB is *far* more then required for the Vita's rather low resolution. 1600x1200 is 1.92MPixels, the Vita is running 544,000Pixels. With one quarter of the Vita's VRAM you can drive four *times* the amount of pixels it has. Your numbers are off by at least a factor of sixteen.

For the rest of the hype over the GPU, just wait until any bench not named GLBenchmark are released, don't be too let down

 

[Steelbom]

The GPU being twice as fast as the iPad 2 leads to the obvious question- WTF are they thinking?

Exclude that any other mobile devices exist from this argument, it doesn't matter.

The new iPad has *four* times the pixel count of the previous. If we take everything Apple says as gospel truth then the new iPad has two choices- run things at half the speed of the 'old' model, or scale the resolution which completely negates the advantage of having the higher resolution display in the first place.

Again, forget anything not made by Apple exists in the world. How is moving to a new iPad an upgrade over the iPad 2 for any sort of GPU intensive task?



I keep seeing the bar lowered for how little people know about GPUs on this forum- the degree of ignorance is honestly astounding.

http://www.anandtech.com/show/429/13

The fastest boards listed in that chart have 32MB of RAM. 128MB is *far* more then required for the Vita's rather low resolution. 1600x1200 is 1.92MPixels, the Vita is running 544,000Pixels. With one quarter of the Vita's VRAM you can drive four *times* the amount of pixels it has. Your numbers are off by at least a factor of sixteen.

For the rest of the hype over the GPU, just wait until any bench not named GLBenchmark are released, don't be too let down

We'll see soon enough what the deal is, but Infinity Blade: Dungeons and Sky Gamblers: Air Supremacy which were both demoed at the Keynote look quite impressive graphically, in fact I'd say look better than anything available on iOS thus far.

I don't think a fourfold increase in GPU power is necessary to handle the resolution, the MP2 was already pretty beefy.

 

[SickBeast]

I'm guessing that a quad core proved to be useless to a tablet in apple's testing. Even so, they could have at least bumped up the cpu speed to 1.5ghz. It's not as though arm has made zero progress over the past year.

 

[smartpatrol]

I'm guessing that a quad core proved to be useless to a tablet in apple's testing. Even so, they could have at least bumped up the cpu speed to 1.5ghz. It's not as though arm has made zero progress over the past year.

The A5's die size is HUGE for a mobile chip, 50% bigger than Tegra 3 despite being only dual-core. A5X will be even bigger, assuming it's on a 45nm process. Even if it's on 32nm or 28nm it'll still be rather large.

Apple simply chose to use its die area for more GPU cores rather than more CPU cores, and chose to dedicate more of its power budget for driving the high-res display and bigger GPU rather than cranking up the CPU clockspeed. They seem like pretty sensible design choices to me.