3 ms·

While I am not entirely surprised, this still makes me so happy to see, and makes me even more positive about a more efficient future of computing through the A

While I am not entirely surprised, this still makes me so happy to see, and makes me even more positive about a more efficient future of computing through the ARM core and its ISA. That Atom chip was even clocked 33% higher, and if my memory serves me right the Atom Diamondville also has a higher IPC than the current ARMv7.

ajross 15y ago
It's a little cooked. The OMAP4430 is a latest-and-greatest SoC design. The Atom N270 was released 4 years ago, for goodness sake. It's true that there's a hole in the Atom line right now -- Medfield would be the best comparison, and it's not in the market yet. Cedar Trail (also not quite released) is a higher power part than the N270, but should honestly wipe the floor with the OMAP. That's not to say that ARM hasn't come a long way over the past half decade. They clearly have, and are getting to the point where there almost at parity with Intel's SoC offerings. But this comparison isn't really fair.
- hackermom 15y ago
  This isn't just about computational performance. It's just as much about power consumption. Diamondville sits at about 2.5 watts per core, while the A9 sits at under 300 mW per core, totalling to some orders of magnitude better performance per watt. Should we say that the comparison is unfair because of this as well? We could maintain that stance up until the point where we compare two CPUs with identical processing power and power consumption, but somehow that would defeat the purpose of comparing at all. And for the record, they could just as well have brought the Pineview and Cedar Trail cores into the picture, as fresh as the A9, which might've been more "fair", but very little still would've changed in computational performance. These tests are really about showing exactly how much you can get out of under a watt of power when opting out of x86.
  - ajross 15y ago
    Sure, there are apples/oranges problems all over the place in this comparison. But you completely missed my point. I'm saying that comparing TI's flagship SoC product with a 4-year old CPU that is three (soon to be four in a month or two) process nodes behind Intel's state of the art and claiming it is somehow representative of "ARM" vs. "Atom" is straight up crazy. How would you like it if Phoronix decided to show off a Medfield vs. ARM11, which are separated by roughtly the same amount of engineering time?
    - hackermom 15y ago
      No I got your point exactly. But Diamondville, Pineview et al. represents the current available state of the art in mobile x86 cores from Intel. Medfield isn't released yet, but when it shows up I would be pleased seeing it pitted against the state of the art ARM cores.
    - nextparadigms 15y ago
      Intel hasn't improved performance of Atom all that much in the past 4 years. They've been mostly working on lowering TDP, while in the same time ARM tried to maintain their very low TDP, while increasing performance. Either way, these are still underclocked Cortex A9 CPU's. By the time Intel Atom hits the market in a phone or tablet, it will compete with same clock dual core or quad core Cortex A9, dual core Krait chips, and possibly even dual core Cortex A15 chips by summer. There's no way Intel will remain competitive - not even in performance, let alone power consumption and price.
      - ajross 15y ago
        I'm with you up until the last sentence, where you veer off into fantasy land. When will the ARM competitor to a 4-core 3.7GHz Sandy Bridge arrive? :) [edit, rather than needlessly extend a fanboi thread: the comparison with SB certainly is relevant to your statement "There's no way Intel will remain competitive - not even in performance", which was laughably wrong. Intel won't be losing the lead in performance any time this decade, I assure you.] Look, this is all just fine. ARM has made some amazing progress in the last five years, and they are to be commended. But the cheerleading has gotten out of hand. Intel's death has been predicted so many times over the years that it's become a running gag in the industry. ARM Ltd. is just the latest competitor to show some spunk. Will they succeed where Motorola, MIPS, DEC, and AMD failed? ... Maybe! But I'm not putting a lot of money on it. And I'm certainly not changing my opinion based on a comparison between a brand new Pandaboard/GNexus and a so-old-it-smells-bad netbook. [1] 68000, then R2000, then Alpha, then the K6/Athlon/Athlon64 -- every one of those designs was going to "kill Intel" according to "serious people in the Industry" -- check the Google groups archives for all the fun flame wars.
rbanffy 15y ago
I think the comparison would be better if made with currently available similar devices. There are Atom processors with comparable power consumption and, I assume, better performance than the N270. And then there is the OMAP 5 family, which looks very promising. And the 4470 is a more current representative of the OMAP 4 family. The 4430 they benchmarked is the low-end guy. I think it's safe to assume someone is building OMAP 4470 or OMAP 5 devices that could be benchmarked against the current crop of netbooks and low power x86 notebooks.
- hackermom 15y ago
  That would be Pineview and Cedar Trail. I am not sure if the CPU portion of these cores have the same power consumption as the N270; their TDP includes the consumption of the on-die GPU as well. As far as computational power goes, they are just marginally better performers than the Diamondville core.
  - rbanffy 15y ago
    I think the N2800 should be significantly faster than the N270: http://ark.intel.com/compare/36331,58917 http://ark.intel.com/compare/36331,58917 Not only it has two cores instead of one, it has 64-bit instructions (and twice as many registers), an L2 cache twice as large and 4 times the memory bandwidth (if I read the specs correctly). The TDP is much larger, but that only means it has an integrated number cruncher that can, conceivably, do some numeric heavy-lifting.
Symmetry 15y ago
Atom Diamondville also has a higher IPC than the current ARMv7 I'm not sure what you mean here. ARMv7 is an ISA, not an architecture, so different ARM cores that run ARMv7 code, everything from the A5 to the A15, will have radically different IPCs even though they can have the same set of instructions for a given program. And you can't even really compare the IPC of processor architectures with different ISAs directly, its often easier to solve a problem in one ISA with less instructions than in a different ISA. RISC architectures often do less work per instruction than CISC architectures, but ARM has some features like instruction predication and argument shifting that mean that its individual instruction are often more powerful than, say, MIPS. And it is actually fair to compare processor that have different clock rates, because often the architecture determines how fast a chip you can make. Itanium servers are only clocked at around half the speed of other architectures because the ISA is very complex and that's as fast as they can make them, and so comparing e.g. Power and Itanium processors of equal clock speed would be horribly unfair to the decisions that Power made that allow them to clock so high. So to get a fair comparison you really want to just build the best processor of a given architecture you can on a given process, and let them fight it out like they do here. This can be pretty hard to do, since the performance of different company's processes on a given node can be rather different. I seem to recall that Intel could put maybe 20% more current through a N-gate on the 65nm node than their competitors, for instance, but using the same process would be the ideal.
- hackermom 15y ago
  How can you produce such a wall of text over a little misnomer? You obviously understood that I am referring to the current incarnation of the ARMv7 ISA, a.k.a the Cortex family.
  - Symmetry 15y ago
    Mostly I put up a wall of text because I think the subject is interesting and I wanted to talk about it, but really there is no one IPC for the Cortex Family. An A5 will tend to have an IPC that is much lower than an Atom. An A15 will tend to have an IPC that is much higher than an Atom. And for the cores being compared here, on tight microbenchmarks the superior execution resources[1] and OoO execution of an A9 will put it ahead of an Atom, but on pointer chasing code the superior branch predictor of the Atom will probably give it a higher IPC. EDIT: [1] Well, really the 16 vs 8 registers, 3 operand instructions, and free barrel shift in each instruction. I think they have the same number of ALUs and AGUs.

ajross 15y ago

It's a little cooked. The OMAP4430 is a latest-and-greatest SoC design. The Atom N270 was released 4 years ago, for goodness sake. It's true that there's a hole in the Atom line right now -- Medfield would be the best comparison, and it's not in the market yet. Cedar Trail (also not quite released) is a higher power part than the N270, but should honestly wipe the floor with the OMAP. That's not to say that ARM hasn't come a long way over the past half decade. They clearly have, and are getting to the point where there almost at parity with Intel's SoC offerings. But this comparison isn't really fair.

hackermom 15y ago
This isn't just about computational performance. It's just as much about power consumption. Diamondville sits at about 2.5 watts per core, while the A9 sits at under 300 mW per core, totalling to some orders of magnitude better performance per watt. Should we say that the comparison is unfair because of this as well? We could maintain that stance up until the point where we compare two CPUs with identical processing power and power consumption, but somehow that would defeat the purpose of comparing at all. And for the record, they could just as well have brought the Pineview and Cedar Trail cores into the picture, as fresh as the A9, which might've been more "fair", but very little still would've changed in computational performance. These tests are really about showing exactly how much you can get out of under a watt of power when opting out of x86.
- ajross 15y ago
  Sure, there are apples/oranges problems all over the place in this comparison. But you completely missed my point. I'm saying that comparing TI's flagship SoC product with a 4-year old CPU that is three (soon to be four in a month or two) process nodes behind Intel's state of the art and claiming it is somehow representative of "ARM" vs. "Atom" is straight up crazy. How would you like it if Phoronix decided to show off a Medfield vs. ARM11, which are separated by roughtly the same amount of engineering time?
  - hackermom 15y ago
    No I got your point exactly. But Diamondville, Pineview et al. represents the current available state of the art in mobile x86 cores from Intel. Medfield isn't released yet, but when it shows up I would be pleased seeing it pitted against the state of the art ARM cores.
  - nextparadigms 15y ago
    Intel hasn't improved performance of Atom all that much in the past 4 years. They've been mostly working on lowering TDP, while in the same time ARM tried to maintain their very low TDP, while increasing performance. Either way, these are still underclocked Cortex A9 CPU's. By the time Intel Atom hits the market in a phone or tablet, it will compete with same clock dual core or quad core Cortex A9, dual core Krait chips, and possibly even dual core Cortex A15 chips by summer. There's no way Intel will remain competitive - not even in performance, let alone power consumption and price.
    - ajross 15y ago
      I'm with you up until the last sentence, where you veer off into fantasy land. When will the ARM competitor to a 4-core 3.7GHz Sandy Bridge arrive? :) [edit, rather than needlessly extend a fanboi thread: the comparison with SB certainly is relevant to your statement "There's no way Intel will remain competitive - not even in performance", which was laughably wrong. Intel won't be losing the lead in performance any time this decade, I assure you.] Look, this is all just fine. ARM has made some amazing progress in the last five years, and they are to be commended. But the cheerleading has gotten out of hand. Intel's death has been predicted so many times over the years that it's become a running gag in the industry. ARM Ltd. is just the latest competitor to show some spunk. Will they succeed where Motorola, MIPS, DEC, and AMD failed? ... Maybe! But I'm not putting a lot of money on it. And I'm certainly not changing my opinion based on a comparison between a brand new Pandaboard/GNexus and a so-old-it-smells-bad netbook. [1] 68000, then R2000, then Alpha, then the K6/Athlon/Athlon64 -- every one of those designs was going to "kill Intel" according to "serious people in the Industry" -- check the Google groups archives for all the fun flame wars.

rbanffy 15y ago

I think the comparison would be better if made with currently available similar devices. There are Atom processors with comparable power consumption and, I assume, better performance than the N270. And then there is the OMAP 5 family, which looks very promising. And the 4470 is a more current representative of the OMAP 4 family. The 4430 they benchmarked is the low-end guy. I think it's safe to assume someone is building OMAP 4470 or OMAP 5 devices that could be benchmarked against the current crop of netbooks and low power x86 notebooks.

hackermom 15y ago
That would be Pineview and Cedar Trail. I am not sure if the CPU portion of these cores have the same power consumption as the N270; their TDP includes the consumption of the on-die GPU as well. As far as computational power goes, they are just marginally better performers than the Diamondville core.
- rbanffy 15y ago
  I think the N2800 should be significantly faster than the N270: http://ark.intel.com/compare/36331,58917 http://ark.intel.com/compare/36331,58917 Not only it has two cores instead of one, it has 64-bit instructions (and twice as many registers), an L2 cache twice as large and 4 times the memory bandwidth (if I read the specs correctly). The TDP is much larger, but that only means it has an integrated number cruncher that can, conceivably, do some numeric heavy-lifting.

Symmetry 15y ago

Atom Diamondville also has a higher IPC than the current ARMv7 I'm not sure what you mean here. ARMv7 is an ISA, not an architecture, so different ARM cores that run ARMv7 code, everything from the A5 to the A15, will have radically different IPCs even though they can have the same set of instructions for a given program. And you can't even really compare the IPC of processor architectures with different ISAs directly, its often easier to solve a problem in one ISA with less instructions than in a different ISA. RISC architectures often do less work per instruction than CISC architectures, but ARM has some features like instruction predication and argument shifting that mean that its individual instruction are often more powerful than, say, MIPS. And it is actually fair to compare processor that have different clock rates, because often the architecture determines how fast a chip you can make. Itanium servers are only clocked at around half the speed of other architectures because the ISA is very complex and that's as fast as they can make them, and so comparing e.g. Power and Itanium processors of equal clock speed would be horribly unfair to the decisions that Power made that allow them to clock so high. So to get a fair comparison you really want to just build the best processor of a given architecture you can on a given process, and let them fight it out like they do here. This can be pretty hard to do, since the performance of different company's processes on a given node can be rather different. I seem to recall that Intel could put maybe 20% more current through a N-gate on the 65nm node than their competitors, for instance, but using the same process would be the ideal.

hackermom 15y ago
How can you produce such a wall of text over a little misnomer? You obviously understood that I am referring to the current incarnation of the ARMv7 ISA, a.k.a the Cortex family.
- Symmetry 15y ago
  Mostly I put up a wall of text because I think the subject is interesting and I wanted to talk about it, but really there is no one IPC for the Cortex Family. An A5 will tend to have an IPC that is much lower than an Atom. An A15 will tend to have an IPC that is much higher than an Atom. And for the cores being compared here, on tight microbenchmarks the superior execution resources[1] and OoO execution of an A9 will put it ahead of an Atom, but on pointer chasing code the superior branch predictor of the Atom will probably give it a higher IPC. EDIT: [1] Well, really the 16 vs 8 registers, 3 operand instructions, and free barrel shift in each instruction. I think they have the same number of ALUs and AGUs.