> One reason I believe that over the next 10 years there will be a
> paradigm shift is that normal business metrics will be thrown out by
> Google's behavior. When you speak of building a compiler or porting
> apps, you base this argument on the fact that it should be
> economically advantageous for a company to do so... I maintain that it
> is economically advantageous for Google to create an environment where
> the hardware and software are cheap to allow a far larger
> audience. Google has the will and it has the financial ability.
Consider a company like Oracle and say that they have 10,000 engineers
(I really don't know how many they have; this is just an estimate).
Their loaded cost is say $200K so that's an expense of $2b/year and
they are making loads of money. There are lots of other software
companies out there making money, selling software and paying their
engineers. Same deal at Google. I read that they received 75,000
resumes in the last couple of weeks for 6,000 jobs.
Would Google want to create an environment where you could build your
own Google Earth? If they truly wanted to make an environment where
software is cheap, they could just open source their code. Why haven't
they done this?
As far as I know, Intel has the best-perforing general purpose x86
compilers. Microsofts are fairly good but they only run on Windows.
With GNU, performance depends on community interest and hardware
company dollars thrown at specific platforms - Intel isn't going to
donate money for the AMD64 port - AMD would if they want a better
compiler.
Let's look at a simple program:
#include "stdio.h"
#include "emmintrin.h"
int main() {
__attribute__ ((aligned(16))) float diddle[4]={1.0, 2.0, 3.0, 4.0};
__attribute__ ((aligned(16))) float diddle2[4]={1.0, 2.0, 3.0, 4.0};
__m128 a;
int i;
scanf("%d", &i);
_mm_storer_ps((float *) diddle+i, a);
_mm_store_ps((float *) diddle2, a);
}
If I compile it (using GCC on Mac OS X) at optimization level 0,
the main guts looks like this:
movaps -48(%rbp), %xmm0
movaps %xmm0, -64(%rbp)
movaps -64(%rbp), %xmm0
movaps %xmm0, -80(%rbp)
movaps -80(%rbp), %xmm0
shufps $27, -80(%rbp), %xmm0
movaps %xmm0, -96(%rbp)
movq -16(%rbp), %rax
movq %rax, -24(%rbp)
movaps -96(%rbp), %xmm0
movaps %xmm0, -112(%rbp)
movq -24(%rbp), %rax
movaps -112(%rbp), %xmm0
movaps %xmm0, (%rax)
leaq _diddle2(%rip), %rax
movq %rax, -32(%rbp)
movaps -48(%rbp), %xmm0
movaps %xmm0, -128(%rbp)
movq -32(%rbp), %rax
movaps -128(%rbp), %xmm0
movaps %xmm0, (%rax)
If I compile at optimization level 3, I get this:
movaps -32(%rbp), %xmm0
shufps $27, %xmm0, %xmm0
leaq _diddle(%rip), %rax
movaps %xmm0, (%rax,%rdx,4)
movaps -32(%rbp), %xmm0
movaps %xmm0, _diddle2(%rip)
You don't have to be an expert at programming assembler code to tell
which is faster. The move instructions to and from memory can take
quite a while and there are a ton more in the -O0 generated
code. Intel's compiler can do automatic parallelization, automatic
vectorization, processor specific optimizations and they have
high-performance libraries for a variety of mathematical tasks.
In the old days when there were lots of computer companies with their
own architectures, you had teams that worked on the OS, teams that
worked on compilers, teams that worked on middleware, maybe teams that
developed applications. So you had experts in many areas of relatively
low-level programming. x86 came along and decimated that model and now
you have far fewer companies doing operating systems, compilers,
databases, etc.
In x86, you have the big 3: Intel, Microsoft and GNU. There is
Pathscale and probably some other smaller players.
What do you have for ARM? Is there anything comparable to Intel's
compiler for ARM? There might not be a need as the architecture is
much simpler. But if companies want the best performance from
software, they'd have to roll their own or pay someone else to build
and maintain one for them. As ARM chips take on x86 features, the
complexity of the compiler will go up.
> How do you compete with a company whose motivation is to make products
> free in one space in order to make more money in another space...
Competing with free is always tough. In the old days, there was an
often-used paradigm of creating a company that did consulting work
and software development work. The idea was to take the profits of
the consulting work to fund software development work which would
have higher margins but it requires a large capital outlay upfront.
Microsoft and Apple are competing with Linux which is free. Apple
and Microsoft are doing rather well, correct?
Microsoft Office competes with Oracle Open Office which is free.
Microsoft still gets a ton of money from Microsoft Office.
DB2, SQL/Server and Oracle all compete with MySQL and PostGres
which are free (Oracle recently bought Sun which bought MySql).
IBM's stock price is at an all-time high today. Microsoft and
Oracle seem to be making lots of money too.
Hulu provides free television content. Netflix seems like it is
doing fine competing against free.
Your typical large restaurant has soup kitchens. Does that mean
that there are no restaurants around?
Market Data 
Markets 
AI
