|
Sun Microsystems
has hit another milestone on its way toward the delivery of servers
based on its future "Rock" Sparc RK processor, which sure is a lot
better than having a millstone around its neck as often happens with
advanced chips. Sun announced that only 15 weeks after taping out and
only six weeks after getting back first silicon versions of the Rock
chip, the company has its Solaris 10 Unix variant running on the
processor in alpha editions of its upcoming "Supernova" servers.
Sun Microsystems
has hit another milestone on its way toward the delivery of servers
based on its future "Rock" Sparc RK processor, which sure is a lot
better than having a millstone around its neck as often happens with
advanced chips. Sun announced that only 15 weeks after taping out and
only six weeks after getting back first silicon versions of the Rock
chip, the company has its Solaris 10 Unix variant running on the
processor in alpha editions of its upcoming "Supernova" servers.
Because it is clearly the nerdiest of the IT suppliers regardless of the advertisements from Apple Computer that try to portray the amalgam of IBM and Microsoft
cultures that represents the PC as the hopeless nerd, Sun is actually
trying to show off in its press release by referring to the Rock chip
as a "hexadeca core" UltraSparc implementation. This nomenclature,
which is long for 16-core, is unintelligible to people who don't know
Greek and Latin, of course, but if you want to impress your friends,
there is a great site called The Phrontistery
that explains the numerical prefixes and adjectives in Greek, Latin,
and other languages, including hexadeca as well as one of the most
peculiar words I have ever seen in English: zenzizenzizenzic,
which means two to the power of 256. If only Scrabble had six Zs in the
bag, you could really crush your opponent with that one on a triple
word spot. This site also explains that we should have been programming
in sexadecimal, but apparently programmers were too shy or prudish to admit it. (That's a joke. Don't send that email.)
Jeff Thomas, senior vice president of engineering at Sun's brand
spanking new Microelectronics group, did not give out a lot of detail
about the Rock chip that was being tested. But he did say that Sun has
obtained a number of the Rock chips "in the low double digits" from its
foundry partner, Texas Instruments,
and that it will have a number of Rock parts in the "low triple digits"
soon. "We have lots of parts, and we could get more if we needed them,"
he said. The Rock chips are being tested in real Supernova systems, not
in a mock-up test board, since Sun tests and debugs the chip and the
server design at the same time. He said that Sun is testing the Rock
chip in entry Supernova configurations now, but will eventually test it
inside larger machines with multiple processor sockets. At some point,
Sun will deploy the Supernova systems it builds internally, running
real workloads, and will then open up a system beta program for early
adopter customers to take the Supernova out for a spin.
Solaris was booted on that entry, single-socket Supernova system last
Friday, according to Thomas. The version of Solaris 10 is not the 11/06
update that shipped in March (yes, that was a little bit late), but a
future and as-yet unnamed update to Solaris 10 that has been tweaked to
support the Rock processors and their unique hardware architecture. The
engineers in the lab are setting up Supernova boxes with multiple
sockets now and will soon have this patched Solaris 10 running on them
as well.
Sun is not running application-level benchmarks on these Supernova
machines right now, but rather lower-level code that stresses different
parts of the Rock chip and the underlying systems. The idea is to try
to break the system with whacky code because out there, somewhere,
there is a data center that just might put together a string of
improbably complex or strange code that can send a processor and its
operating system out of its mind. Thomas said that Sun simulated
several trillion processor cycles for the Rock chip on its big wonking
chip simulator, but that with a real piece of silicon, the company can
accomplish what took many months in a matter of seconds. The tests also
allow Sun to push the clock speeds to see where TI's manufacturing
processes might need a little tweaking.
Like other chip makers, Sun does not want to say exactly what kind of
bugs that it is looking for, and it sure doesn't want to say what it
has found so far, if anything. "You either have stuff that is really
stupid and simple that is easy to detect and fix, or complex stuff that
is electrically strange, not repeatable, and takes lots of conditions
to make happen," explains Thomas. "You can always find ways to get
around the latter one, too. Luckily, you don't have a lot of problems
in the middle, since the simulations tend to find that stuff early on."
Sun is not being specific right now about when the Supernova systems
will ship, but Fadi Azhari, director of marketing for Sun's "Niagara"
Sparc T1 and Rock Sparc RK processors, reiterated that Sun was on track
to deliver Supernova systems in the second half of 2008.
While Sun is booting Solaris 10 on the Rock systems now, it is also
possible that Sun works with the Linux community to get a Linux
distribution ported to the processors, much as Canonical
has done with the 8-core, 32-thread Sparc T1 chip. "It is not currently
in the plan, but technically, there is no reason it cannot be done,"
says Azhari. And Thomas backed him up, saying that there is nothing
that would technically prevent Linux from running on Rock systems. But
there is a question about the scalability of the Linux kernel and how
it will make use of so many compute and scout instruction threads.
The Rock machines are expected to have one, two, four, and eight
sockets, and will probably range from four cores (that means
three-quarters of one dud chip) in an entry configuration, from eight
to 32 cores in a midrange box (using two sockets and probably some
partially defective cores), and up to 128 cores in the largest
Supernovas (all possible cores going full-out).
Read the original article: http://www.itjungle.com/breaking/bn050207-story01.html
|