Thanks for these details. This makes sense. So I'm going back to read v8 code now! :)
Alexandre On Sep 10, 12:52 am, Erik Corry <[email protected]> wrote: > 2009/9/10 Rames <[email protected]>: > > > > > The -march=armv7-a indeed solved all my problems. Thank you > > very much for all you answers. > > > So I eventually get a score around 94 for the v8 (version 3) tests > > with v8 on an OMAP3530 (Cortex A8 600MHz). > > I previously got a score around 800 on a 800Mhz Intel processor. > > > The score gap is amazing for only 200MHz difference. I have not > > compared v8 arm and ia32 code. Do you know if this gap results from a > > better code optimization and/or better native code usage (I see a lot > > of posts relative to ARM work on the v8-dev group so I guess the ARM > > work is not yet to the level of the ia32's) or if the x86 architecture > > merely have a more efficient Instruction Set or architecture? > > The code generator for ARM is not as advanced as the one for Intel. > My very rough guess is that the score on the V8 benchmark suite could > be almost doubled if the code generator were as advanced on ARM. The > tricky bit is to speed up the running code without slowing down > compilation, which can be a big factor in page load time on slow > devices. > > Looking at raw MHz can be misleading. The Cortex A8 design in the > BeagleBoard is an in-order design, whereas the Cortex A9 ARM chip will > be a "multi-issue superscalar, out-of-order, speculating 8-stage > pipeline" according to ARM's white paper. This can make quite a > difference. The in-order Atom is said by Wikipedia to have half the > performance of the out-of-order low end Celerons with around the same > Mhz. > > Apart from the MHz, the speed of the memory subsystem can be very > important. This is partly a property of the CPU core itself, but to a > larger extent a property of the second level cache size and speed as > well as the speed of the external RAM interface. There are many > tradeoffs here in system design to do with cost, space and power use. > It's fair to say that V8 (and browser) performance is very dependent > on the memory subsystem. > > -- > Erik Corry, Software Engineer > Google Denmark ApS. CVR nr. 28 86 69 84 > c/o Philip & Partners, 7 Vognmagergade, P.O. Box 2227, DK-1018 > Copenhagen K, Denmark. --~--~---------~--~----~------------~-------~--~----~ v8-users mailing list [email protected] http://groups.google.com/group/v8-users -~----------~----~----~----~------~----~------~--~---
