Apple A7 is now 64-bit

[Exophase]

Also, compile GB3 with a proper compiler instead of GCC for the Intel chip (if Intel spends all this $$ developing a compiler to give its chips an edge, it should be used) and you'll see FPU and INT scores skyrocket.

That's an assumption on your part, you don't really know what the scores would be like with ICC. It's not this compiler that's a huge deal faster in some majority percentage of programs.. on SPEC it's mostly dramatically faster on just a few benchmarks and very same-ish on a lot of other ones. Gives a great end score for SPEC, but that doesn't mean it'll necessary do the same with GB3.

 

[Arachnotronic]

That's an assumption on your part, you don't really know what the scores would be like with ICC. It's not this compiler that's a huge deal faster in some majority percentage of programs.. on SPEC it's mostly dramatically faster on just a few benchmarks and very same-ish on a lot of other ones. Gives a great end score for SPEC, but that doesn't mean it'll necessary do the same with GB3.

Fair enough

 

[liahos1]

Rooooh want to look at them :biggrin:

BTW did you notice this: http://browser.primatelabs.com/geekbench3/compare/93819?baseline=93813

(we're getting off topic, probably should move...)

what is going on in this benchmark? I dont get it?? What is the first column being reported off of?

 

[krumme]

This is probably the most misguided thing that I've seen in this thread.

On a per-core basis, excluding the utterly BS crypto benchmarks, Silvermont holds its own against Cyclone in integer performance (don't give me the BS IPC argument, either, perf/watt is all that matters). Also, compile GB3 with a proper compiler instead of GCC for the Intel chip (if Intel spends all this $$ developing a compiler to give its chips an edge, it should be used) and you'll see FPU and INT scores skyrocket.

But let's just say that A7 is better than Bay Trail and that Geekbench is the be-all, end-all of chip benchmarking tools. Who gives a $#!%? Is Apple selling chips to Intel's customers now? When I buy a Windows 8.1 tablet or Android tablet, will I have an Apple A7 option? No.

I dont think BT gets faster when bt phone variants arive for Android late h1 2014...

I dont doubt tablet bt is faster than current gen phone soc from apple or qq. No wonder. But its pretty aparent its game over. What a waste of money and totally deprived of any business creativity.

You/me just buy our windows tablet. Unless consumer behavior changes we will be a minority.

 

[Arachnotronic]

I dont think BT gets faster when bt phone variants arive for Android late h1 2014...

I dont doubt tablet bt is faster than current gen phone soc from apple or qq. No wonder. But its pretty aparent its game over. What a waste of money and totally deprived of any business creativity.

You/me just buy our windows tablet. Unless consumer behavior changes we will be a minority.

Maybe, but remember that Intel's Merrifield/Moorefield will be using ImgTec's Rogue GPUs and not Gen7, and it can use its process lead to clock 'em higher than anybody else.

 

[Exophase]

Maybe, but remember that Intel's Merrifield/Moorefield will be using ImgTec's Rogue GPUs and not Gen7, and it can use its process lead to clock 'em higher than anybody else.

Let's hope they use a suitably wide one then. They clocked their SGX540s and 545s higher than anyone else too but that didn't stop them from getting crushed by the competition using much wider designs. Until they brought out SGX544MP2s anyway.

It's going to be really funny if Merrifield ends up with significantly more GPU power than BayTrail-T. Not at all unlikely though. At least it can't be as dramatic as CT vs CT+.

 

[NTMBK]

Maybe, but remember that Intel's Merrifield/Moorefield will be using ImgTec's Rogue GPUs and not Gen7, and it can use its process lead to clock 'em higher than anybody else.

Does 22nm mobile offer significantly higher clocks? We've seen 2.4GHz turbo, which is the same ballpark as the competitors (2.3GHz "benchmark mode" on Samsung parts ). I thought that the advantage was in power consumption and die size?

 

[ancientarcher]

I wish Intel sticks with its own GPU, a variant of HD whatever, rather than going on board the IMG wagon. (even though they do own ~15% of the company). It will be much better than competition in the GPU space.

But then Vivante has been gaining share of late and Nvidia has said that others can outsource its mobile GeForce GPU. So, we have at least 5 GPU suppliers (IMG, ARM-Mali, Nvidia, Vivante and Adreno) even without Intel's prop solution. That's pretty good!

 

[Nothingness]

It's going to be really funny if Merrifield ends up with significantly more GPU power than BayTrail-T. Not at all unlikely though. At least it can't be as dramatic as CT vs CT+.

That'd be ironic indeed. Intel GPU history is so funny :biggrin: The only great thing about Intel GPU is that at least they put real effort into open source drivers.

 

[mavere]

Futuremark guy offers explanation/speculation on the 5s's lack of improvements in 3DMarks Physics:

We've looked at this at great length, and the result is quite interesting. We are working with our relevant partners on this to verify, but at current it seems that:

The compiler does do SIMD and Neon. These do not offer any effect. We even tried doing Neon optimizing by hand to see if the compiler is not doing what it should, but no real effect

The Physics test is uses Bullet, and practically the whole CPU time is spent in the soft body solver, PSolve_links. If you pull this function out of Bullet and bench it separately, you do see a 2x speed increase. However, once it's inside the physics engine, you see nothing.

As this seemed to make no sense, we spent a few days trying to understand what is happening. The result seems to be that if the soft bodies are arranged in memory so that the CPU can access them in a sequential fashion, you get a 2x to 3x increase in speed. This is higher if it can run up the memory, a bit lower if it runs down. The way bullet places the bodies in the memory is a lot more random and they are accesses in a jump-back-and-forth manner. When memory is accessed in this way, all speed gains are lost.

iPhone 5 shows none of this behaviour. It is realistic to assume that in the new 5s we see the new prefetch in action, but it cannot gain traction with a random memory access pattern.

It's good to understand that this is not a flaw in Bullet. Arranging complex memory structures in memory to be in a sequential fashion is non-trivial to say the least. Our hacked solution only worked for us as we knew exactly the data that we would be using. Worrying about where your memory segments lie is not something the programmer should have to worry about anyway.

In terms of our Physics test at large, it does not appear that we have an inherent flaw - our use-case is simply such that no gain is seen. Any game/app using Bullet would see the same, and there are naturally other apps that will see the same (in GeekBench, you can see a few tests where the same thing is happening).

 

[Khato]

While it still doesn't provide a full picture, hooray for another data point in the form of notebookcheck's review of the iPhone 5s. Note that it'll probably be a bit before the English version is available, but German's good enough, right? Especially since the primary point of interest is the power consumption/battery life data and a comparison of it to their testing of the iPhone 5c and the iPhone 5.

Maximum idle/load goes from 1.4/2.9W on the iPhone 5 to 1.3/3.3W on the iPhone 5c to 1.6/3.8W on the iPhone 5s. The only peculiar number in there being the maximum idle for the iPhone 5c - it should be comparable to the iPhone 5s and a bit above the iPhone 5, not below. After all, the majority of components are shared between the iPhone 5c and iPhone 5s, no? As well, there's not much reason for the maximum idle to load delta to go from 1.5W on the iPhone 5 to 2W on the iPhone 5c, whereas 1.5W to 1.7W sounds reasonable.

Regardless, it looks like the estimates for A7 power consumption being in the 2-2.5W range are correct. Which makes for a pretty decent bump from the A6 in the 1.3-1.8W range. (Note that I'm giving a pretty decent range there and each should be taken on either the low or high side of it, not low for one and high for the other.) Either way it's quite reasonable and certainly rains on Snapdragon 800's parade - the delta power consumption between maximum idle and maximum load in the three smartphone's that they've tested using it end up around 3.5W. Now part of that difference is certainly the fact that the A7 is only a dual core, but I'd guess the majority of that maximum load power is going to the GPU anyway.

 

[Nothingness]

Anand iPad Air review is up.

I don't buy the 6-wide issue. At least there's close to 0 chance Cyclone decodes 6 instructions per cycle, that'd be crazy. I wonder if Anand did not see something like 100% clock boost with a 3-wide machine.

 

[Khato]

I'd been suspecting something odd with reported clock frequency ever since the initial reviews, but if Anand is doing low level benchmarks like I'd expect then he could rule that out. A 3-wide machine with a 100% clock boost would only show a 3x boost in performance if you change a loop from a single add that's dependent upon the previous iteration to six adds that are dependent upon their previous iteration while a 6-wide could well be capable of a roughly 6x boost - please note that I know that's not exactly the correct way to describe it, but it's close enough/easier to understand. That would also explain why the core sees a much larger gain in micro-benchmarks than actual programs.

Regardless, best thing about the review was that Anand finally took at least a rough look at power consumption... and it looks like the dual core A7 has a TDP of approximately 7W on pure CPU load. I'm quite hopeful that Anand will be so kind as to give us similar numbers for Baytrail-T and Haswell-Y tablets at some point - I'd not be surprised at all if Haswell-Y is a tad bit higher while Baytrail-T is a fair amount lower.

 

[Exophase]

Regardless, best thing about the review was that Anand finally took at least a rough look at power consumption... and it looks like the dual core A7 has a TDP of approximately 7W on pure CPU load.

There's about a 6W differential from idle to peak.. for the entire platform, running something unknown. You can't account all of this to the SoC alone. At the very least there's the RAM and PMIC, and we don't know what was being done with wifi.

 

[Khato]

There's about a 6W differential from idle to peak.. for the entire platform, running something unknown. You can't account all of this to the SoC alone. At the very least there's the RAM and PMIC, and we don't know what was being done with wifi.

No, it's pretty clear from the first graph that the platform idle power consumption with this method of measurement is about 4 watts. Meanwhile the maximum power consumption visibly crosses the half-way mark between the 10 and 12 watt markers, so it's safe to say that it hits 11 watts - the delta is at least 7 watts.

Now you are correct that we don't have any way of knowing how much of the power delta is due to the SoC versus other components that also become active during this power virus test. It easily could be the case that up to a watt of that delta is not attributable to the SoC (I doubt it given that the most effective CPU core power virus test would reside entirely in cache though.) Regardless, you're still talking about going from a 3 watt delta on the A6X (which would have a greater non-SoC power delta if there was one, especially if the workload is hitting memory) to a 7 watt delta on A7. Which looks even worse if you subtract 1 watt of that from each as you're then talking triple the peak power consumption for the SoC.

Edit: Actually, doing a quick bit of pixel comparison on the full size graphs yields an idle of roughly 3.8W and load of 11.2W, so the delta is closer to 7.4W. (Not that there's much of a point in trying to be that precise with relatively rough data - looks like the resolution used was roughly 0.2W.)

 

[Ajay]

What, not 2 GB LPDDR3 upgrade? No A7X with a 128 bit memory bus. That's it, I'm not buying this overpriced, sub par piece of technology.

Sorry, it just had to be said

Oh, does anyone have half a dozen bennies burning a hole in the pocket calling my name?

 

[tential]

What, not 2 GB LPDDR3 upgrade? No A7X with a 128 bit memory bus. That's it, I'm not buying this overpriced, sub par piece of technology.

Sorry, it just had to be said

Oh, does anyone have half a dozen bennies burning a hole in the pocket calling my name?

For the price Apple is demanding, I don't think upgrading to 2GB was a big deal, especially when 2 and even 3 GB has been industry standard this generation. Having to reload pages ALWAYS annoyed me on the old Ipads, and the fact that 64bit apps also take up more memory just means it's even less suited to having 1GB of ram.
For such a spectacularly designed product I'm surprised they didn't just make the adjustment for the what? 15 dollars it might add tops to their production costs (I think that may be exaggerating even).

Overall a great product, and it makes me wonder when Android phone/tablet makers will get a CPU this good. Thing is though, as good as it is, people don't jump ship that easily. Even though I like this, I can't justify $500, and then having to buy all my apps again on top of that. Nexus 7 is like 80-90% of the performance for 1/2 the price but if I was on the Apple ecosystem already, and I had more disposable income I'd definitely get it.

 

[lopri]

Both the iPhone 5S and the new iPad seem like too transitional products to me. iPhone will inevitably receive larger screen treatment next year and that will make the previous iPhones instantly obsolete. iPads are obviously not as balanced resource-wise as iPhones, according to the AT's review, and will probably be updated to a better configuration.

So for the phones Apple cemented the internals and for tablets it set the form factor for the next several years. No doubt iOS 7 will last a few years with refinements along the way. I just am not sure about today's models and their longevities.

 

[lopri]

Imagine Intel fabbed A7 and then what AT's review of iPhone/iPad might look like. ^^ Too bad it is Samsung.

 

[mavere]

S

Actually, doing a quick bit of pixel comparison on the full size graphs yields an idle of roughly 3.8W and load of 11.2W, so the delta is closer to 7.4W. (Not that there's much of a point in trying to be that precise with relatively rough data - looks like the resolution used was roughly 0.2W.)

Until Anand provides more results, I'm not sure if using a a mysterious stress test as a measuring stick is particularly useful.

However, it appears that Kraken creates a 3W power delta and GFXBench creates a 2.5W one. Are there similar data points for Bay Trail and Snapdragon 800?

 

[Khato]

Eh, it's useful to get an idea of just how much power Cyclone sucks down under load. But if you want to stick to Kraken it's still basically the same story - the stress test would yield 3.75W per core while Kraken shows 3W, it's not a huge difference.

Unfortunately I'm not aware of any 'clear' measurements of Snapdragon 800 - notebookcheck reviews are the closest and they don't really specify what's used for 'maximum load' measurements on these devices. Meanwhile the closest you can get with Baytrail that I'm aware of are the measurements taken on Intel equipment at launch, don't believe anyone has good data on the actual devices.

 

[fteoath64]

Anand iPad Air review is up.

I don't buy the 6-wide issue. At least there's close to 0 chance Cyclone decodes 6 instructions per cycle, that'd be crazy. I wonder if Anand did not see something like 100% clock boost with a 3-wide machine.

Well, the clock boost would be easily detected. However, I would like to suggest that the 4MB of "L3" cache might be doing some crazy pre-staging of some sort ?.

 

[fteoath64]

There's about a 6W differential from idle to peak.. for the entire platform, running something unknown. You can't account all of this to the SoC alone. At the very least there's the RAM and PMIC, and we don't know what was being done with wifi.

This might indicate some sort of power management trickery at play (rather clever). Could the 4MB Ram cache be turned off ? Or memory be downlocked and voltage lowered ? RAM controllers and caches are known power hogs.

 

[scannall]

It's hard to know what's going on inside Apple. But it seems they have either 2 or 3 CPU teams (Probably 2, but you never know), with a LOT of people working on this stuff.

Is their CPU design department bigger than AMD's now?

The A7 really is a beast, and tells me they have some real talent there.

 

[Jurge92]

Everything is tempting about the iPad Air, except the fact that it only has 1GB of RAM. It's a tough decision because tomorrow, and tomorrow only, I can get the iPad Air 20% off. So if I'm going to get it, it would be wise to act fast...

Is the "low" amount of RAM a reason not to buy it? I mean, I have never noticed any problems on my iPhone 5, but I reckon the 64-bit CPU will consume larger amounts of RAM.

Any thoughts?