[Beowulf] NVIDIA buying PGI

mathog mathog at caltech.edu
Tue Aug 6 13:25:32 PDT 2013

On 05-Aug-2013 12:00, beowulf-request at beowulf.org wrote:
> I think it would be hard to
> argue that PG was dominant.  They are good, of that there is no 
> doubt.
> But I don't think they've been spectacularly profitable.  Tools
> companies rarely are.

That's certainly true.

My views on the relationship between software producers and hardware 
producers can probably
best be characterized as radical fringe.  I find that it is frequently 
bad when a hardware company is
the sole provider of software for its own hardware, and can be even 
worse when they have any control
over software that is needed to run another company's hardware.  In my 
ideal world, which I have
no expectations will ever come to pass, there would be strict 
separation of the two functions, which
should result in a free market for software, and a free market for 
hardware, with none of
the "strings attached" limitations between one realm and the other that 
are so common now.  At the
very least hardware companies should be required to publish the 
complete programming specifications for
their devices, which would go a long way towards leveling the playing 

This isn't just for CPUs, the case is actually much more pressing for 
other devices.  (For CPUs they typically
do publish adequate technical documents, so that "anybody" could write 
a compiler for one.  The problem for CPUs
is that writing a good compiler is so much effort that in practice the 
number of compiler providers is quite small.)

Consider, for instance, graphics cards.  Historically the only provider 
of drivers for these devices were their
manufacturers.  The manufacturers kept it that way by not publicly 
releasing adequate technical documents for
their hardware, without which it was impossible to write a driver.  Or 
at least impossible to do so efficiently, without
having to laboriously reverse engineer the device.  For Windows users 
this made no difference whatsoever,
since the manufacturers put huge efforts into developing those drivers, 
but for "minor" platforms support was often spotty
or nil.  Which arguably went a long way towards keeping those platforms 
in the "minor" category.  Electrically there was no reason you could not 
use that ATI card on a Linux box, but good luck finding a good driver.  
There were occasional bright spots, like Xi graphics' fine X11 products 
for some Radeon models, but by and large, choice was limited.

Chipset support, and especially NIC support, was frequently in the same 
boat.  So when the latest and greatest motherboard came out, which you 
generally needed to run the latest and greatest CPU, you could use it 
for Windows, but it took a year or two for adequate drivers for Linux to 
appear.  For these smaller devices I do not know for sure that lack of 
access to technical documentation was the limiting factor, but I suspect 
as much.

It isn't just computers in computer centers that have this issue.  Want 
to know how your car works?  Sure, you can plug an OBD-II reader in and 
pull out some information, but there is a lot more in there that you 
cannot retrieve.  The car manufacturers sell diagnostic equipment that 
can get at those extras, but the cost is often prohibitive for 
independent repair shops.  It would be trivial for other vendors, who 
sell less expensive test equipment, to modify their software to
do the same thing - if they had the technical specs.  This one is 
already in the political domain, see for instance:


Note, there does need to be a safety exception beyond which 
manufacturers would not need to expose information.  "Open" car 
diagnostics would be a plus, having people replacing their car's 
"operating system" would certainly end up killing people.
Similarly, reprogramming your pacemaker would also be a bad idea.  A 
bad NIC driver could _indirectly_ result in somebody's death,  but it 
could not cause the computer to jump out of the rack and physically harm 
someone, so it would not be covered by that exception.


David Mathog
mathog at caltech.edu
Manager, Sequence Analysis Facility, Biology Division, Caltech

More information about the Beowulf mailing list