Custom PC - UK (2020-05)

FEATURE / DEEP DIVE

In theory, this most obviously affects

workloads such as gaming, where the

constantly changing nature of your

interactions with the virtual world make

for an ever-changing dataset, but it also

applies to some more mundane workloads

too. Conversely, a more fixed dataset that

can be read from memory in a controlled,

predictable manner – such as encoding

video – cares very little about latency.

In our tests, however, tighter latency

timings have made a miniscule difference

to performance compared with the clock

speed. Lower-latency memory undoubtedly

makes for a faster system than higher-latency

memory, but it’s also more expensive, and

that money would arguably be better spent

on memory with a higher clock speed instead.

Regardless of your memory needs, the

manufacturers of memory have thankfully

made it relatively easy for you to know

exactly what you’re buying, even if you’re

still not entirely sure what speed you need.

You’ll generally get the four latencies

advertised alongside the clock speed of the

memory, with both figures nearly always

printed on each memory module. And if

these figures aren’t shown, you’ll at least

generally see the PC4 number.

The PC4 numbering scheme tells us the

type of memory (PC4 refers to DDR4, PC3

to DDR3 and so on). The next part of the

number is the memory bandwidth, which is

calculated by multiplying the clock speed by

the memory bus width (that’s the number of

parallel traces connecting the memory to the

CPU, which is 64-bit) and dividing by eight (to

convert from bits to bytes). Then, at the end

of the number will be the tCL timing in the

format CLxx. So, for 3000MHz memory with

tCL 14, you’ll get a figure of PC4-24000CL14.

COMPLETING THE PICTURE

Getting back to the internals of our memory

modules, as you’d expect, a single memory

bank alone doesn’t make up an entire

memory module. Instead, multiple banks of

cells, sense amps and data buffers are stacked

together into bank groups. Several chips are

then arranged together to form what’s known

as a rank of chips, just as you see on the side of

a memory module. Slap another set of chips

on the PCB and you have the classic dual inline

memory module (DIMM).

Notably, your computer needs

to account for all these steps when

addressing the memory. As such, a

memory request will specify the

DIMM, the rank and the bank before

getting to the rows and columns.

However, along with the

memory banks themselves,

there’s very little to a

memory module. A

small chip called the

Serial Presence Detect

(SPD) is responsible

for communicating the

operating parameters of the RAM to the

system, such as voltage, clock speed and

memory timings, but otherwise memory

modules are quite bare boards.

Instead, the grunt work of taking full

advantage of all the data parallelism

created by having multiple memory banks

in multiple ranks and on multiple DIMMS,

is performed by the memory controller in

your CPU. It’s here that much of the work is

performed to optimise the order in which

data is sent to or requested from memory.

This is also partly why the likes of the

first generation of Ryzen processors were

so troublesome when it came to memory

compatibility. With the controller being built

right into the CPU, any gremlins with the

new architecture and platform couldn’t

be cushioned by the RAM at least being

controlled by a known quantity.

AUTOMATIC PROFILES

One of the easiest ways to unlock the best

performance from your memory is to

hop into your BIOS and opt for either the

Extreme Memory Profile (XMP) or AMD

DOCP option. These options automatically

set the modules to the settings to which

they’re rated on the packaging, overriding

A SMALL CHIP CALLED THE SPD IS

RESPONSIBLE FOR COMMUNICATING

THE OPERATING PARAMETERS OF

THE RAM TO THE SYSTEM

The tiny SPD chip is the only other major integrated circuit

on a DIMM, other than the memory chips themselves

FEATURE/ DEEPDIVE

Intheory,thismostobviouslyaffects

workloadssuchasgaming,wherethe

constantlychangingnatureofyour

interactionswiththevirtualworldmake

foranever-changingdataset,butit also

appliestosomemoremundaneworkloads

too.Conversely,a morefixeddatasetthat

canbereadfrommemoryina controlled,

predictablemanner– suchasencoding

video– caresverylittleaboutlatency.

In ourtests,however,tighterlatency

timingshavemadea minisculedifference

toperformancecomparedwiththeclock

speed.Lower-latencymemoryundoubtedly

makesfora fastersystemthanhigher-latency

memory,butit’salsomoreexpensive,and

thatmoneywouldarguablybebetterspent

onmemory with a higher clock speed instead.

Regardless of your memory needs, the

manufacturers of memory have thankfully

made it relatively easy for you to know

exactly what you’re buying, even if you’re

still not entirely sure what speed you need.

You’ll generally get the four latencies

advertised alongside the clock speed of the

memory, with both figures nearly always

printed on each memory module. And if

these figures aren’t shown, you’ll at least

generally see the PC4 number.

The PC4 numbering scheme tells us the

type of memory (PC4 refers to DDR4, PC3

to DDR3 and so on). The next part of the

number is the memory bandwidth, which is

calculated by multiplying the clock speed by

thememorybuswidth(that’sthenumberof

paralleltracesconnectingthememorytothe

CPU,whichis 64-bit)anddividingbyeight(to

convertfrombitstobytes).Then,attheend

ofthenumberwillbethetCLtiminginthe

formatCLxx.So,for3000MHzmemorywith

tCL14,you’llgeta figureofPC4-24000CL14.

COMPLETINGTHEPICTURE

Gettingbacktotheinternalsofourmemory

modules,asyou’dexpect,a singlememory

bankalonedoesn’tmakeupanentire

memorymodule.Instead,multiplebanksof

cells,senseampsanddatabuffers are stacked

together into bank groups. Several chips are

then arranged together to form what’s known

as a rank of chips, just as you see on the side of

a memory module. Slap another set of chips

on the PCB and you have the classic dual inline

memory module (DIMM).

Notably, your computer needs

to account for all these steps when

addressing the memory. As such, a

memory request will specify the

DIMM, the rank and the bank before

getting to the rows and columns.

However, along with the

memory banks themselves,

there’s very little to a

memory module. A

small chip called the

Serial Presence Detect

(SPD) is responsible

for communicating the

operatingparametersoftheRAMtothe

system,suchasvoltage,clockspeedand

memorytimings,butotherwisememory

modulesarequitebareboards.

Instead,thegruntworkoftakingfull

advantageofallthedataparallelism

createdbyhavingmultiplememorybanks

inmultipleranksandonmultipleDIMMS,

is performedbythememorycontrollerin

yourCPU.It’sherethatmuchoftheworkis

performedtooptimisetheorderinwhich

datais senttoorrequestedfrommemory.

Thisis alsopartlywhythelikesofthe

firstgenerationofRyzenprocessorswere

so troublesome when it came to memory

compatibility. With the controller being built

right into the CPU, any gremlins with the

new architecture and platform couldn’t

be cushioned by the RAM at least being

controlled by a known quantity.

AUTOMATIC PROFILES

One of the easiest ways to unlock the best

performance from your memory is to

hop into your BIOS and opt for either the

Extreme Memory Profile (XMP) or AMD

DOCP option. These options automatically

set the modules to the settings to which

they’re rated on the packaging, overriding

A SMALL CHIP CALLED THE SPD IS

RESPONSIBLE FOR COMMUNICATING

THE OPERATING PARAMETERS OF

THE RAM TO THE SYSTEM

The tiny SPD chip is the only other major integrated circuit

on a DIMM, other than the memory chips themselves