RETRO TECH / ANALYSIS
effective frequency of 400MHz. Intel launched the Pentium
4 with the 850 chipset, which only supported RDRAM, but
the Pentium 4 could now take advantage of all that extra
memory bandwidth.
Again, though, RDRAM was expensive, and not many
people already had RDRAM sticks in their old PCs that they
could simply transfer to a new system. Intel started bundling
RDRAM with the CPUs in an effort to get people on board,
but it was a tough ask.
With disappointing sales and reviews for the Pentium 4,
Intel backtracked and launched a new SDRAM-supporting
chipset for the Pentium 4, called 840. The performance
was dreadful without the extra memory bandwidth, but itdid show that Intel’s combination of RDRAM with a quad-
pumped FSB worked – the Pentium 4 really needed to be
paired with fast memory.
The magic bullet finally came with Intel’s 845 chipset,
which supported DDR memory running at 133MHz
(266MHz effective). DDR memory was significantly cheaper
than RDRAM, and it provided a very sensible compromise
over Rambus in terms of bang per buck.
The last gasp for Rambus came with the later 850E
chipset, which could run RDRAM at an effective frequency
of 1066MHz, but had no native support for USB 2 and was
pretty much dead in the water on launch. Intel then put all
its work into supporting DDR, with the E7205 ‘Granite Bay’
chipset supporting dual-channel memory, where two
memory sticks interleave to create more bandwidth.The Pentium 4 was the first CPU to really require the large
heatsink-and-fan assemblies that we still use today, with
some of the first Pentium 4 PCs being equipped with wind
tunnels to link the CPU cooler with the case’s exhaust fan.
The second problem is that it makes a CPU very inefficient
at processing code with unpredictable branches. Intel’s plan
to get around the latency created by the many-stage pipeline
was to make use of advanced branch prediction techniques. If
a CPU was performing a task that repeatedly used predictable
code branches, such as video encoding, then the CPU could
efficiently predict what it needed to do. Loops of code could
be handled quickly too, thanks to Intel’s new L1 Trace Cache
system, which moved the L1 cache to a position after the
decode unit, so any microinstructions held in it would already
be decoded.
For these reasons, Pentium 4 CPUs usually excelled in
software with predictable instructions. The problem was that
a lot of code didn’t behave this predictably, particularly if you
were running lots of legacy applications at once. If the CPU
got it very wrong, the pipeline would have to be flushed and
start again. That’s not a massive problem if a CPU has a short
pipeline, but it quickly makes the CPU inefficient if it has a
long pipeline.
The result was that the Pentium 4 could process fewer
instructions per clock (IPC) than the Pentium III in a lot of
standard software, negating the benefits of those huge
clock frequencies.MISSING THE BUS
Another key difference between the Pentium 4 and its
predecessors was its front side bus. As we mentioned
earlier, there was a large disparity between the 133MHz
Pentium III FSB and the massive bandwidth of RDRAM. Intel
aimed to fix this with the Pentium 4 by introducing a quad-
pumped FSB, where four signals are sent per clock cycle.
The FSB still fundamentally ran at 100MHz, but it had anA relic from Custom
PC’s past – the Beat
the Office CPU was
an overclocked
2.6GHz Northbridge
Pentium 4C attached
to a Vapochill phase
change system at
-22.5°C – we had
it running stably
at 3.54GHz
An Intel Pentium 4C wafer – the ‘C’ denoted that
the processor had an 800MHz front side busIntel backtracked and
launched a new SDRAM-
supporting chipset