ARM/Intel

ARM/Intel

The new A15 SOCs get 14000 DMIPs for duel core at 2 GHz. This is referring to the Samsung SOC that is coming out/is out. I read on XDA that that is the speed but I cannot find any official documents on it. Also have no clue what the TDP is and power consumption. Does anyone know?

Also Intel IB 3517U gets 47300 at 17 TDP and the 3667U gets 52000 at 17TDP but the power consumption is unknown.
Does anyone know how Intel compares to ARM when it comes to DMIPs per watt? TDP and watt consumption?

Also how well does Intel CPUs scale? Does their efficiency go up the slower they go? ie performance per watt?
The Intel MIC cards AKA knights corner gets better performance per watt at 3GHz at 62w TDP compared to faster speeds. It gets ~5GHz at 250w TDP. I'll find the source tomorrow. Does that continue to be the same for the ULVs if you run it at 1 core at 1GHz? How does performance compare? Can anyone test this using HWinfo with an ULV IB and test the scaling and performance in DMIPs? I am interested in seeing how that compares to ARM and how this works out. There appears to be no data on ARM and especially on ARM vs Intel in performance per watt.

If ARM is somehow so much more efficient how do they do that? How does a 32nm SOC have far better performance per watt compared to a 22nm CPU?


I might have more but i am tired and have forgotten so I'll post it later if i remember.

 

I'm assuming we're only comparing straight out Integer thoroughput?

 

i would like to compare everything but ARM doesn't have crap worth documentation

 

ARM CPUs are all RISC, therefore are a lot smaller/ simplier than Intel's x86 CPUs, which are all CISC.

 

read through some of the wikipedia articles on it and i don't really understand how ARM has such high DMIPs? Are they high because RISC chips are good at just simple instructions but when they are given more complex needs they crap out?

 

As I understand it, the ARM CPU is a special purpose device which excels at performing simple integer operations. They have a limited repetoire of simple instructions like multiply, add, subtract etc however, they are also highly optimized to perform those operations. The reason they perform so well in their native environment is that there is a hard limit of possible instructions a piece of code can request of the CPU. While their peak thoroughput is quite impressive, you have to also examine what kind of test are you using. ARM cpus are very light on thread scheduling, lack OOE capabilities and generally have a much less efficient resource management relying more on the high level optimization by the developer. Complex, heavily branching code with heaps of conditionals will very quickly choke an ARM cpu, as will multiple unoptimized code streams which are competing for the core resources.

x86 CPUs also enjoy having a tightly integrated FPU alongside the ALUs so you expect superior Floating Point performance. Additionally, some of the instructions can allow the "bundling" of multiple simple instructions together to execute in a single clock cycle. i.e. an ARM may take 5-6 cycles to complete a particular operation which may only take 2 cycles on an x86. The main one which comes to mind is the FMA instruction.

 

Exynos 5250 is a 3-4.5W TDP design. It's power usage goes up dramatically at higher clock speeds because on HK-MG process you can't scale clock speeds like you can with Qualcomm/TSMC's Poly-Si node. HK-MG is more efficient though. Exynos in the Google Chromebook is 1.7GHz, and the new Nexus 10 is "reportedly" clocked at 1.7GHz. In phones we'll likely see 1-1.2GHz, to keep the voltages lower.

Just think if Intel or AMD could move this fast... In 2007 we were still using ARM11. In just that short time since then we've moved from ARM11 to Cortex A8, A9 and now A15. A15 is 40% faster than A9. A9 was ~20% faster than A8 per-core (FP performance almost doubled). And standard A8 was ~2x faster than the best ARM11 core on the consumer market.

And no, when presented with more complex tasks they don't "crap out". The software/kernel is compiled to the specific instruction set of whatever ARM arch you're working with. Apple's A6 is ARMv7s. A9 is ARMv7... and so on.

As of Cortex A9, ARM is fully out-of-order, where A8 was in-order and Qualcomm's MSM86x0 was speculative OoO. Also as of A9, ARM has pretty much mandated the inclusion of NEON Streaming SIMD's. Tegra 2 was the only A9 SoC that didn't have NEON and it sucked.

On A15, ARM's thread scheduling is MUCH MUCH improved.

 

Do you have a source for that TDP? I saw only one place after googling for hours and it was just a comment by some random person saying it is 4w TDP. So exactly how would a Intel chip stack up against an ARM chip like pros and cons areas where one is better than the other and so on. How is this going to turn out for lets say a 10w TDP haswell chip that is do and the new 20nm Exynos chips that are going to be coming out. Also what about GPU side of stuff?

 

Source: Samsung Semiconductor. They don't publicly state TDP as different OEM's can ship it with different voltages and different clock speeds, both of which will effect TDP. Reference is 3.5-4W. I'm guessing, because I haven't looked at any Exynos 5250 kernel source, that it's also topping out around ~1500mV.

I think Intel is going to run into more and more roadblocks to keeping a certain power envelope. Lowering it to 10W is a massive feat, especially given their IPC with IVB and what's predicted with Haswell. The thing about ARM is that they've hit their ceiling on consumer SoC power consumption, and now ARM's development team is focusing on thread performance and IPC, just like Intel. So far they're doing really good for their TDP ratings. Also, Qualcomm has been a BIG leader in FP performance. When A8 was still high-end, Qualcomm's Scorpion core was putting up 4-5x the FP performance. It still is with Krait. Where A8 had a single 32 or 64-bit wide SIMD, Qualcomm was running dual issue 64-bit NEON SIMD. Krait is also an in-house arch, so you can't really directly compare it to ARM's licensed reference designs like the cores in OMAP, Tegra or Exynos.

Also, Intel is already beating ARM on die size though. Exynos is a damn big die. ~118mm2 if I'm not mistaken. IVB was significantly smaller.

 

BTW I looked all over samsings site and can't find it....where is that refference to TDP. How is their die so big and uses so little power?

 

It's an internal document behind a login. :/ Sorry.

Transistor count. Intel has billions of transistors. ARM's architecture doesn't. Since literally everything is on-die on an SoC, except the audio DAC and PMIC and such, the dies tend to be bigger.

 

you mean radios?

Also you work for samsung or something? Sorry i just don't take....i have intern document privileges seriously without some sort of proof because anyone can just say that...no offense hal

 

I don't work for Samsung, but I have access to data sheets through my employer. Believe it if you want, or not. I can't furnish proof of this information. It violates the agreement I made to have access to this documentation.

Radios aren't on die for Exynos. They can be a package for Qualcomm though. You have the SoC with the CPU cores, GPU cores, memory controllers, bus interfaces, cache and some other stuff. Then off die you have your RAM, usually PoP but lately with these higher TDP SoC's it's not stacked, radio/modem/amps, audio DAC, PMIC, MHL/USB controller, etc...

The modem everyone is using ATM is either Intel's XGold XMM6220 or Qualcomm's multimode MDM9x15 (usually MDM9615).

 

doesn't IB have all of those? I know it is either haswell or boradwell that will have everything on the CPU from north and south bridge to memory controllers and some other stuff so it doesn't seem very different to me

 

ahhhh, im getting rusty at this, so ARM have finally gotten OOE, FP and better scheduling. The question now remains how much of their efficiency can they retain as performance ramps up. x86 CPUs are also on a manufacturing advantage at the moment which muddies the water somewhat.

 

CISC versus RISC doesn't tell you anything about size or complexity. Plus Intel's designs have mostly all been RISC CPUs with front ends that decode x86 instructions since the mid to late 90s (and some other RISC CPUs have had decoders now too).

Obviously ARM so far can only compete on the ultra low end, and it's looking like Intel has the best performance with power consumption somewhere in the middle of the pack with their newest Atom chips (which are all basically the same original Atom design so far but they keep improving die process and what's integrated, etc.)

 

While I see what you're trying to say, I think comparing modern ARM to modern Intel is a poor comparison. Intel's CPU architectures are simply wayyyyy more developed and battleproven than ARM's.

If you're going to make the comparison, you should be comparing ARM (2007-2012) to Intel (2002-2007). Of course, that STILL doesn't account for the benefit that ARM receives from working with mature kernels.

EDIT: what H.A.L. 9000 has been saying is more than reasonable; I don't have any access to insider info, but everything he's said has been highly consistent with AnandTech....and AT is a VERY good resource.