quickstart THE BEGINNING OF THE MAGAZINE, WHERE ARTICLES ARE SMALL
CPU for
Dream
Machine
2050?
A team of scientists from
IBM and Georgia Institute of
Technology recently astounded
the PC overclocking crowd by
announcing they’d pushed a
silicon-germanium chip, which
normally runs at 350GHz, to a
ludicrous 500GHz. The insanely
clocked CPU was cooled to
about -451 F using liquid heli-
um—just slightly warmer than
absolute zero.
Though the ultra-high clock
speed is certainly an astound-
ing achievement, what’s even
more impressive is that it was
achieved using low-cost manu-
facturing techniques and com-
mercial chip technology. And
though the power user in us
salivates at running a 500GHz
liquid-helium-cooled CPU, the
ultimate goal of this project
is to provide faster CPUs for
mobile phones. Still, a guy can
dream, can’t he?SEPTEMBER 2006 MA XIMUMPC 09
AMD Axes Large
Cache CPUs
Just two months after launching its AM2 line
of CPUs, AMD suddenly axed three out of the
eight new dual-core procs—all of the con-
sumer-level processors that feature 1MB of L
cache, except the FX-62.
The company says it killed the 1MB parts to
reduce a confusing overlap between processor
models, and simplify its roadmap. The com-
pany’s 2.4GHz Athlon 64 X2 was offered with
both 512KB and 1MB of L2. The company said
customer pressure is the main reason for the
demise of the 1MB-cache CPUs. This should
save AMD some money, as CPUs with greater
L2 cache take up more space on a wafer than
CPUs with smaller cache. By killing the procs
with 1MB cache, AMD will be able to crank out
more CPUs per wafer, increasing production
capacity. AMD denies that cost savings was the
primary motivator for the change.We’ve been singing the BTX
formfactor’s swan song for
months, but according to Intel,
not only is BTX not dead, it’s
thriving. Intel recently released
its BTX forecast, pegging the
adoption rate at 36 percent by
the end of 2007. It also claims
that 5 percent of all DIY builds
will be BTX within the same
time frame. The numbers may
seem fi shy, but it makes sense
when you consider that 92 per-
cent of Dell desktops and 98
percent of Gateway desktops
are BTX.BTX Is Back
Intel insists its ATX formfactor
replacement is alive and wellwhen you consider that 92 per- The BTX formfactor’s massive
CPU cooler won’t be necessary
with the cooler Conroe-based
CPUs shipping soon.W
ith both AMD and Intel working on quad-core
processors for next year, people are wonder-
ing if we’re in another shortsighted CPU arms race,
like the earlier battle for astronomical clock speeds
that ended in a cease-fire due to temperatures spi-
raling out of control. Will multicore processors hit
a ceiling as well? Sure they will. Every technology
has limitations. But someday, I think we’ll see PC
processors with dozens or even hundreds of cores
per chip. And writing software for them won’t be a
major headache, as programmers fear today.
Chip-fabrication technology is the biggest
technical limitation. Broadly speaking, the fabrica-
tion process determines the transistor budget
(the number of transistors available to implement
a design) and the chip’s power consumption
and manufacturing cost. Those are strict limits.
Engineers can’t integrate more cores than the
transistor budget allows. Nor can a chip require so
much power or cost so much money that it would
be impractical. Within those limits, the potential
number of cores that can fit on a die depends on
the cores’ size.
While AMD and Intel work on future quad-
core processors, other companies already make
processors with hundreds or even thousands of
cores. Cisco Systems designed a chip for Internet
routers that has 192 cores. Connex Technology
has a video processor with 1,024 cores and a
prototype with 4,096 cores. These cores, how-
ever, are much smaller and simpler than the x
cores on PC processors.
Future PC processors, I believe, will inte-
grate several complex cores with many simpler
cores, and some cores will be highly specialized.
Programmers won’t have to write difficult mul-
tithreaded programs that distribute workloads
across multiple cores. Instead, the operating
system will assign particular tasks to individual
cores. The network stack might run on a simple
core optimized for packet processing. Bigger tasks
will run on the complex general-purpose cores. If
some cores are idle for a while, so what? We won’t
worry about wasting CPU cycles, just as today we
don’t worry about not using every available byte of
RAM or megabyte of disk space.
I expect AMD and Intel will begin designing
these superchips when engineers reach the inte-
gration limit using conventional x86 cores. That
limit is less than 10 years away.Tom Halfhill was formerly a senior editor for Byte magazine
and is now an analyst for Microprocessor Report.Super-Size
MeFAST FORWARD
TOM
HALFHILL