58 |^ |^ June 2019
and always seemed to be the budget
alternative. It had followed Intel’s path
to producing complex designs, and
similarly started with a fresh multi-core
approach. Its Zen architecture is
cleaner, more efficient, and had up to
eight cores. This was rapidly followed
by Threadripper, a ludicrously big chip
for a desktop, with 16 cores. If it’s all
about the core count, AMD is winning.
Simply adding cores is no guarantee
of performance, of course. In the worst
case scenario, your fancy eight-core
machine is running software that
simply hammers one core. However, if
you have idle cores, it also means you
have spare thermal dissipation. It’s safe
to run some cores above specification,
because they’re next to idle, and
therefore cold, cores. Turbo Boost does
this (or Turbo Core, if you are AMD).
When software demands the highest
performance possible, via the ACPI
(Advanced Configuration and Power
Interface), the processor overclocks
selected cores, and lets
others idle. This is how Intel’s
Core i9-9900K can reach the
magic 5GHz: It is only
running on two cores.
The upshot is that base
clock speeds are no longer
a reliable indication of
performance. After years
selling us better chips
because they were ‘faster’,
with a number to prove it, the industry
has yet to give us an easy metric to
measure a chip’s relative power.
Where single-core speed can still
count is in games, which can be
notorious for hitting single cores hard.
Intel has traditionally had the top GHz
numbers, which has resulted in Intel
machines running most (but not all)
games faster. This is changing. Clock
speeds aren’t going to increase
anywhere like as dramatically as they
have done in the past. It has taken
more than a dozen years to go from4GHz to 5GHz. There’s no sign that
6GHz will be any easier. What we do
have is more cores, and this is where
programmers are going for more
power. Running two cores at 5GHz
on your i9-9900 leaves six cores (12
threads) doing nothing, potentially
nearly three times the computational
power. Boost rates give a nice headline
number, but as applications spread
out properly, they will start to
become redundant.Cache phrases
All chips have a small amount of
memory to store the data it is working
on, this is the Level 1 (L1) cache. This
sits right in the core, next to the
registers: the final destination of data to
be processed. A chip can only work at
top speed if it is fed data as fast as it
needs it. If the L1 cache can’t be filled
fast enough, it limits the speed of the
processor. Fetching data from the main
memory is far too slow. There are
separate L1 caches for instructions and
data, 32KB of each for Coffee Lake;
Zen+ gets double that for the
instructions. Cache sizes are quoted in
KiB and MiB, which are the binary
versions, just to avoid confusion, as KB
can mean 1,000 bytes in some circles
(which is why your hard drive is smaller
than you expected).
Above this is a Level 2 (L2) cache. It
used to sit on the motherboard, but
is now on the same silicon as the
processor. This is the main interface
between CPU and memory; data copied
to the L1 cache is also copied to the
larger L2 cache, ready in case. Coffee
Lake chips hav e 256KiB of L2 per core,
Zen+ stretches this to 512. Above this
sits the L3 cache, slowest and largest.
It is the only cache that all cores can
access freely, or blocks of core, in the
case of AMD. This has also migrated to
the CPU, but lives just outside the
cores, and varies depending on which
version of a chip you have – Coffee
Lake has 1–3MB per core; Zen+ ranges
from 2MB to 4MB.
Feeding the caches with the right
data is a dark art. It’s all about
maximising L2 hits, and avoiding
going beyond L3. This gets needlesslyAn Intel Core i7-6700K, the original Skylake halo
chip from 2015. This is still a capable CPU, despite
being only a 6700 – its base clock is 400MHz faster
than an i9-9900.Rise of the Tomb Raider is one of new breed of games that can make use of multi-core processors. As ever,
software lags behind hardware, but it is catching up now.
Battlefield V chews through eight threads, if you have them; above that, it starts to level out. Good, but
we need more, as 16-thread and better chips are not going to be uncommon soon.
Image credits : AMD, AmericanMegatrends, EA, Intel, Microsoft Studios, Natascha Eibl/Meltdown Attack, SquareEnix