Qualcomm Announces 64-bit Snapdragon Processors

[schmidtbag]

64bits isn't just about accessing more ram, that's just ONE of the many advantages.

How about the 31 extra general purpose registers? The dedicated SP? The new instructions? NEON has access to double the amount of registers, now supports double precision floating point, has new instructions for encryption/decryption and SHA hashing.

Assuming that accessing more RAM is the only advantage of 64bit processors is one big mistake, and shows a big lack of understanding.

I wasn't assuming that accessing more RAM was the only advantage, and I wasn't implying it was either. I previously commented that I know there is a performance gain using 64 bit, which the memory itself has very little to do with. Also, many of the features you mentioned aren't dependent on a 64-bit architecture. Key word here is many, not all, so don't get all anal on me about that.

My point was that if you were to strictly just get the extra registers, I personally wouldn't consider the performance gain significant enough considering the broken compatibility and immense headache of porting everything. As far as I'm aware, AArch64 doesn't have compatibility with running a 32-bit OS, though I welcome being proven wrong.

I do understand the idea of doing doing this transition to 64 bit now before it becomes mandatory, but seriously, AArch64 should've been mainstreamed 4 years ago. I don't know enough about android to know how much it will be affected by this transition, but linux and Windows RT could suffer broken compatibility issues, particularly with closed-source software. It's already hard enough as it is trying to entice people to use ARM.

[droidhacker]

Adreno GPUs have been historically non-competitive, until Adreno 320 and Adreno 330. But if the latest one, that's coming in H1 2015 is only 80 percent better, then it won't even beat Tegra K1 from this year.

Also, why does Qualcomm bother to license the ARM architecture if they're doing to use stock ARM anyway?

Historically, Adreno has always been (and remains) the only mobile GPU to offer even the POTENTIAL for open source drivers. Further, the current adreno has ALWAYS been more than competitive against alternatives.

As far as comparing anything against "tegra"... thats just nvidia trash. The only thing they have in their favor is stronger marketing.

[Tundra]

Also, many of the features you mentioned aren't dependent on a 64-bit architecture.

As far as I'm aware, AArch64 doesn't have compatibility with running a 32-bit OS, though I welcome being proven wrong.

That's the thing, with each maker building its own ARM, they can add support for DIFFERENT extra things over the base, and as a software developer you can't support a lot of different CPU models. The 64bits ARM adds many useful things to the BASE model. They can still add things, but the common shared base is BETTER (there is an option to REMOVE things for custom designs, but that's less common).

AArch64 provides user-space compatibility with ARMv7-A ISA, the 32-bit architecture, referred to as "AArch32" and the old 32-bit instruction set. It's similar to how AMD64 can run 32bit software on a 64bit kernel.

[droidhacker]

I wasn't assuming that accessing more RAM was the only advantage, and I wasn't implying it was either. I previously commented that I know there is a performance gain using 64 bit, which the memory itself has very little to do with. Also, many of the features you mentioned aren't dependent on a 64-bit architecture. Key word here is many, not all, so don't get all anal on me about that.

My point was that if you were to strictly just get the extra registers, I personally wouldn't consider the performance gain significant enough considering the broken compatibility and immense headache of porting everything. As far as I'm aware, AArch64 doesn't have compatibility with running a 32-bit OS, though I welcome being proven wrong.

I do understand the idea of doing doing this transition to 64 bit now before it becomes mandatory, but seriously, AArch64 should've been mainstreamed 4 years ago. I don't know enough about android to know how much it will be affected by this transition, but linux and Windows RT could suffer broken compatibility issues, particularly with closed-source software. It's already hard enough as it is trying to entice people to use ARM.

I'm not familiar enough with the differences between arm and 64bit arm to know if backwards compatibility will be an issue or not, but I can say for certain that Android *itself* will not be a problem (being open source), nor the MAJORITY of Android applications, which don't actually have any native code -- therein lies *the* advantage of Java over C. Android is probably the LEAST stressful platform to build for alternative CPU architectures.

Edit... post above me says that arm64 can run arm32 code straight up. Makes the transition trivial and absolute.

HOWEVER, obviously 64bit arm code won't run on a 32bit arm chip. The sooner this transition happens, the less impact it will have on older devices.

[robclark]

Lately, qualcomm has been using a lot of actual ARM cores on their chips. Up to now, they have been limited to low-end chips though.
Thought krait was very nice, the main selling point for the snapdragon chips has and will continue to be the adreno GPUs.

mostly marketing for certain markets.. especially w/ the cortex-a53 stuff (which isn't going to be as fast as current 32b krait)

a57, maybe we start seeing some of the benefits of some of the improvements thanks to armv8 instruction set.

[LightBit]

has new instructions for encryption/decryption and SHA hashing.

"Reduced Instruction Set Computing"

I think building crypto algorithm into CPU is very bad, because:
- it requires writing ASM to use them
- algorithms change
- they are quite big

[bnolsen]

ugh. 64 bit addressing allows files to be directly mapped into memory and also allows developers to not have to worry about heap + stack space overflowing. The bookkeepimg required to deal with 32bit address limitations is utterly beyond irritating. If you have the ram, and ram is insanely cheap, use it and let the OS manage the paging stuff for you. I just don't want to deal with the lame artficial what if cases that 32bit forces you to deal with.

[liam]

Not that I have a problem with 64 bit ARM, but what exactly is the point for non-server purposes? The highest RAM content I've seen on a non-server ARM system was 2GB. Considering the simplicity of the architecture, I doubt there's going to be a significant performance improvement switching to 64 bit. I'd much rather see more plug'n'play features.

AIUI, for armv8 it offers both the legacy 32bit mode, and 64bit mode. The big advantage, however, is that the new isa is supposed to be much more efficient (this is, apparently, the source of the majority of performance gains that are expected with the move from A7/A15 to A53/A57).

[liam]

Adreno GPUs have been historically non-competitive, until Adreno 320 and Adreno 330. But if the latest one, that's coming in H1 2015 is only 80 percent better, then it won't even beat Tegra K1 from this year. Also, why does Qualcomm bother to license the ARM architecture if they're doing to use stock ARM anyway?

The 220 and 225 weren't bad, actually, and they were energy efficient, unlike nvidia. Simliarly, checkout this (http://kyokojap.myweb.hinet.net/gpu_gflops/). The adreno 430 is supposed to get be about 82% faster than the 330 which puts it somewhere north of 235 gflops @ 450MHz. Compare that to nvidia's kepler which needs to run at 950Mhz to get their alleged 365gflops. Consider, also, that the adreno 430 has a faster bus (25GB/s vs around 17GB/s for nvidia) and that power draw difference (~1W for 330 vs ~3W for K1) AND if you qualcomm wanted to blow their power budget they might be able to clock their gpu as high as nvidia which would put it north of 470gflops @ 900MHz. IOW, nvidia didn't design a very efficient gpu, like usual.

[robclark]

The 220 and 225 weren't bad, actually, and they were energy efficient, unlike nvidia. Simliarly, checkout this (http://kyokojap.myweb.hinet.net/gpu_gflops/). The adreno 430 is supposed to get be about 82% faster than the 330 which puts it somewhere north of 235 gflops @ 450MHz. Compare that to nvidia's kepler which needs to run at 950Mhz to get their alleged 365gflops. Consider, also, that the adreno 430 has a faster bus (25GB/s vs around 17GB/s for nvidia) and that power draw difference (~1W for 330 vs ~3W for K1) AND if you qualcomm wanted to blow their power budget they might be able to clock their gpu as high as nvidia which would put it north of 470gflops @ 900MHz. IOW, nvidia didn't design a very efficient gpu, like usual.

fwiw, on the ifc6410/snapdragon-600, it seems pretty stable to overclock the gpu to 487MHz.. also, the android kgsl kernel driver seems to (by default) trade off quite a bit of performance for power. So even the current (already old) stuff has some headroom.

(btw, thanks for that link.. I didn't actually realize the different a320 variants had different # of ALU)