names, versions, generations, and
marketing numbers.
To make a family of chips out of one
design, Intel changes the base clock,
core count, memory speeds, whether
Hyper-Threading is enabled, how much
Level 3 (L3) cache you get, whether it
is overclock locked, and what Turbo
Boost rates it has. Juggling these is how
Intel can sell 24 desktop CPUs out of
one Coffee Lake microarchitecture,
under brands from the Pentium (it’s
still going) to the Core i9.Tick-tock-toe
Intel once followed a production
model called the tick-tock cycle. The
‘tick’ was a die shrink, and refinement
of a design. The ‘tock’ was a new micro-
architecture. The original Core
architecture was quickly followed by
a die shrink, and so on, through a
confusing number of codenames.
Things ticked and tocked along nicely
until problems with the 10nm process
left Intel going ‘tock’ twice. It then
announced it had a new model,
process-architecture-optimisation,
where it follows a die shrink with a new
architecture, which it then optimises.
Then it had to go ‘tock’ again with
Coffee Lake. The first round of the new
model is due to be Cannon Lake (the
shrink to 10nm), followed by Sunny
Cove, then Tiger Lake. The first 10nm
chip is on sale, though you’d be
excused for not noticing. It’s a Core
i3-8121U, and only currently available
in one model of Chinese laptop.
Back in day, it was all about clock
speeds. The faster a chip, the better it
was at everything. Early progress was
rapid, from a 66MHz Pentium in 1993 to
a 1GHz Athlon by 2000. The Pentium 4
was found to be stable at 4GHz in 2004.
Then things stalled. It has taken years
to reach 5GHz – physics gets in the
way. Every time you switch a transistor,
it generates heat. Make them small,
and switch them rapidly, and you can
create more than you can dissipate.
One solution to this is shrinking the die,
making all those transistors smaller, soyou can lower the voltage. This has a
whole other set of problems, too, not
least of which is that the lasers used in
lithography are thicker than the
features you want to cut.
Intel’s Pentium 4 was the turning
point. It had long, convoluted
pipelines, and consumed a lot of
power. It could reach impressive
numbers when liquid cooled, but was
a design dead-end. So, Intel changed
direction, and its Next-Generation
Microarchitecture, or Core, was born.
The answer was short pipelines, higher
efficiency, plus multiple cores. Initially,
clock speeds were lower than the P4,
but soon matched and surpassed it.
The Core chips had up to two cores,
while the Core 2 had up to four (the
use of Core as a brand name can get
confusing, too).
The Pentium 4 did have one feature
that was carried over: Hyper-Threading,
or HT. This is a Simultaneous Multi-
Threading system, or SMT. For each
physical processing core on the chip,
the operating system addresses two
logical ones, and shares tasks between
them. Modern chips are superscalar, sothey can work on multiple
instructions in a single cycle of
the clock. For example, complete
a floating point and an integer
calculation in one click. HT takes
advantage of this. As far as
Windows knows, it is addressing
two separate cores, each one
individually fed data, halted or
directed elsewhere. Thus a four-core
chip with HT can run eight threads, or
tasks, at once. It works extremely well:
The close sharing of resources – both
threads have access to exactly the
same – can work in its benefit, or
occasionally cause contention.
AMD has almost exactly the same
technology, which is simply calls
Simultaneous MultiThreading.Core strength
Pure computational power is how
many instructions you can run in a
given time: the IPC (Instructions Per
Cycle) times the clock speed. Clock
speeds may have stagnated, but IPC
numbers per chip are jumping rapidly
as cores are added. There are
theoretical numbers, simply
multiplying threads and cores, but in
practice these are never reached.
How the application is optimised is
paramount – multi-core designs need
multi-threaded software. To get a
significantly better IPC per core, and
increase the overall efficiency of the
whole family in all configurations, you
need a new microarchitecture.
At this point, we move to AMD. It
had been struggling for years to match
Intel at the high end. It never had the
investment to fully develop its chips,Intel’s Coffee Lake die – eight cores in the middle, GPU to the right,
and the rest spread around the edges.
The whole enchilada, a Coffee Lake
wafer as it comes from the foundry,
ready to be cut up and tested.AMD’s Raven Ridge – that big blue block is an RX Vega graphics engine, which, unlike Intel’s
on-board graphics, makes a fair go of gaming, albeit on a budget.Windows 10
Understanding CPUs
June 2019 | |^57