Re: LuaJIT SciMark Intel/ARM comparison

  • From: Dimiter 'malkia' Stanev <malkia@xxxxxxxxx>
  • To: luajit@xxxxxxxxxxxxx
  • Date: Sun, 12 Aug 2012 14:50:24 -0700

On 8/12/2012 12:52 PM, Mike Pall wrote:
Thanks to everyone for their data points! This allows an
interesting cross-architectural comparison ...

Sure, SciMark is floating-point heavy (as the name implies), but
it still has a significant integer component, due to all of the
array indexing and traversals. Of course, you'll have to decide
whether this is a typical work load for your application and
whether this benchmark is relevant for you at all.

Absolute LuaJIT SciMark scores (higher is better):

785.4  Intel Core 2 E8400, 3.0 GHz
124.7  Intel Atom N270, 1.6 GHz
118.4  Asus Transformer Prime, Nvidia Tegra 3, ARM Cortex-A9, 1.3 GHz
  76.7  HTC EVO 3D, Qualcomm Snapdragon S3, ARM Cortex-A8+, 1.2 GHz
  32.4  Beagleboard xM, TI DM3730, ARM Cortex-A8, 1.0 GHz
  27.5  Raspberry Pi, Broadcom BCM2835, ARM 1176JZF-S, 0.7 GHz

LuaJIT SciMark scores weighted by GHz:

261.8  Intel Core 2 E8400
  91.1  ARM Cortex-A9
  77.9  Intel Atom N270
  63.9  ARM Cortex-A8+
  39.3  ARM 1176JZF-S
  32.4  ARM Cortex-A8

A couple observations:

* A slightly older Intel Atom loses on a floating-point intensive
   benchmark against a recent ARM core.

* The hearsay that a (plain) ARM Cortex-A8 has rather unimpressive
   VFP performance is apparently true -- it's beaten by a much older
   ARM 1176JZF-S design when scaled by GHz.

   [Ok, so the the simpler design probably wouldn't scale without
   extensive modifications to the pipelines. But the original
   Cortex-A8 design with VFP on handbrakes _does_ show.]

* The other hearsay that Qualcomm has taken the original Cortex-A8
   design, but tuned it heavily for their Snapdragon line seems to
   be true. The S3 nearly doubles the FP performance over a plain A8.

A comparison with the Snapdragon S4 or the Exynos line would be
really interesting.

Even more interesting would be to show the SciMark scores weighted
by power consumption (one core under full load). Alas, that data
is not readily available, so I'll leave the speculation to others
(but I think it would be off-topic for this list).


That's very cool!! (I still don't understand why LuaJIT was taken off the
language shootout page)

Btw, I'm not sure how easy and practical it would be, but the SBCL
(Common Lisp) project had this tracking of performance per version
released, or commit:

Would it be legal if the shootout benchmarks are pulled, and used also?
(For what is worth I don't care about Mr. Gouy arguments that ffi was used).

Other related posts: