4 ms·
Teasing out some inferrences: The comparison to the Intel D525 is really unfortunate, as that's a 2.5 year old, 45nm Pine Trail box with a 13W TDP. The low po
by ajross 14y ago
Teasing out some inferrences:
The comparison to the Intel D525 is really unfortunate, as that's a 2.5 year old, 45nm Pine Trail box with a 13W TDP. The low power 22nm Ivy Bridge cores are only 14W, so I'd really wanted to have seen a comparison to a more recent CPU.
The Exynos wipes the floor on the parallel NAS tests. I'm not sure why exactly.
The Atom core has known issues with floating point relative to desktop CPUs, and the A9 was never particularly good at it. So the FFTE results showing the A15 about 50% faster than Atom and 2x faster than A9 are a surprise. It's almost what I'd expect to see from a desktop box at that.
On the other hand, the 7-zip and x264 tests are integer dominated and largely cache-bound, and the A15 doesn't show itself off particularly well here, mostly matching the A9 and Atom per-clock. Yawn. (This is an area where all these CPUs get toasted by Ivy Bridge, btw -- remember that 14W TDP?)
The "Smallpt Global Illumination Renderer" and "C-Ray" tests (which I don't know anything about per se) shows everyone about the same (and for Smallpt 2x as fast as PandaBoard). So I'm going to guess this means the tests are DRAM-bound, and thus infer that the A15 has a similar memory bandwidth to Atom (probably 2x 32 bit vs. the Pine Trail 1x 64 bit channels). That's not great, honestly, as it's still about half of what you can get with a desktop part. But then DRAM refresh draws power, so may not have been an option for the Chromebook .
- mtgx 14y ago14W TDP is still too high, when a dual core A15 gets 4W at most. I doubt the IVB you're talking about is more than 3x faster than A15. Intel also has an advantage in process right now, and I'm not sure they will have it in the future. I think 14nm will be delayed, while others will catch-up to Intel on 14nm, and they might launch 14nm chips in the same time with Intel's 14nm chips, around 2015. We'll see.
- ajross 14y agoToo high for what? Note that the backlight on these devices is probably 6+ watts already. And TDP isn't a good comparison for "max power" (and it's down right awful for "typical power") estimates. It's a thermal number saying what you should design your packaging for. The point being that the Netbook being compared corresponds in the modern world to a much more performant ("ultrabook", sigh) system at the same power draw. As far as process nodes go, I'm not sure what you're citing as evidence there. Intel has been getting farther ahead with each node in the recent past. They launched 22nm more or less simultaneously with TSMC's 28nm logic (who themselves have a good lead over the rest of the industry!). Six months on, all you need to do is walk into a store to find 22nm digital logic in PCs everywhere. I don't believe there are any sub-32nm ARM SoCs in consumer devices quite yet (the Exynos-5 in the linked article is 32nm I believe).
- brigade 14y agoSnapdragon S4 Plus/Pro SoCs are 28nm and have been in phones for a while now.
- brigade 14y ago7-zip and x264 make almost perfect use of additional cores, so it makes sense that a quadcore beats a dualcore. (aside: I rather wish that Phoronix actually mentioned how much a given test benefits from multiple threads, single-threaded tests are presented next to multithreaded tests when comparing chips with different architectures and core counts without any comment or recognition.) Additionally, x264 is heavily optimized for x86 and nowhere near as much for ARM. 7-zip is probably similar. For floating point, A15 has double the execution resources as A9 (two symmetric floating point pipelines each capable of a FMA per cycle) so that's no surprise to me.
- ajross 14y agoExynos-5, OMAP and Atom D5xx are all dual-core CPUs. And they all get about the same performance per clock in those tests. I don't think parallelism is a determiner -- I went with "cache-bound" as my guess. And again, this is a spot where I'd really like to see a comparison with something other than Pine Trail. One assumes the L3 on Ivy Bridge would be a big help.
- brigade 14y agoThe Calexia Highbank Nodes are quad core. Relative to the OMAP4 tested, Exynos 5 has a 1.4x higher clock but 1.9x better 7-zip performance and 2.3x better x264 performance. For reference, A15 has double the vector execution resources as A9 in addition to adding out of order execution. Integer execution resources remain the same (two symmetric pipelines) but gains triple issue, wider out of order, and the potential of two load/store per cycle. So simple integer heavy workloads aren't expected to be incredibly faster from A9, just the ~40% ARM quotes and these tests back up. Atom D5xx has two hyperthreaded cores, and hyperthreading does help a lot for x264 at least. IIRC when I tested on a single-core Atom, the hyperthreading gave x264 a 1.6x boost.