Ubuntu on ARM faster than some Atoms?
phoronix.com
phoronix.com
I use an Atom-based netbook as my home machine and it always felt "fast enough" (up to the point it hits its disk, which is really slow). Now I want a new computer.
BTW, the Atom N270 they are benchmarking was launched in 2008. They picked more or less the slowest Intel processor available.
http://ark.intel.com/products/36331/Intel-Atom-Processor-N27...
There are a lot of valid comments here that these are ancient ATOM processors. But it is also valid to say that ARM SOCs have gotten to the point where they are encroaching on the space, and their makers (TI, BroadCom, Marvell, and others) are being forced by the demands of the SMARTPHONE wars to make them actually useful as general purpose platforms.
As a systems guy I see the things that ARM is lacking, a decent IO bus equivalent to PCI, a memory channel architecture that allows for decent off chip memory performance with memory modules.
Of course the 'market' for general purpose computers is much smaller than the market for computers as accessories (tablets, smartphones, laptops) so perhaps we'll hit a convergence point in the next couple of years.
The surprise to me was how well it did in MP3 and Vorbis encoding, which makes me think there might be some DSP help in the TI board for that.
If you're doing video encoding on any architecture, you need to use one that has SIMD for your architecture, or else it will be horribly slow.
x264 is the only free software encoder I know that has NEON assembly code, though there might be others (certainly none in libavcodec).
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.
These tests are really about showing exactly how much you can get out of under a watt of power when opting out of x86.
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?
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.
[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.
Intel won't die. These things take many years or even decades. You could argue that the PC didn't "kill" the mainframe for 30 years since the PC first appeared, but what happened is that mainframes became increasingly less relevant, and PC architectures increasingly more relevant and popular in more than one market.
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.
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.
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.
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.