Frequency, CAS latency, dual channel: how to read a RAM spec sheet

A RAM technical spec sheet fits on a single line: "DDR5-6000 CL30 2×16 GB." Barely four pieces of information, and yet it is the component for which we see the most configuration errors in the workshop.

A fast kit installed in the wrong slots, an XMP profile never activated in the BIOS, a single stick added to one from a different model: in each of these cases, the memory runs significantly below what was paid for. And unlike an undersized graphics card, the problem is not visible: the machine starts, games launch, and everything looks normal.

This guide dissects the three figures that really matter—frequency, latency, and the number of channels—and, most importantly, explains how to read them together rather than separately. By the end, you will be able to tell in a few seconds if a kit is consistent with your platform.

Frequency: what the number after the hyphen really measures

MHz, MT/s, and the "double data rate" misunderstanding

DDR stands for Double Data Rate. A stick labeled DDR4-3200 does not run at 3,200 MHz: its actual clock runs at 1,600 MHz, but it transfers data on both the rising edge and the falling edge of each cycle. This results in 3,200 million transfers per second, or 3,200 MT/s.

The correct unit is megatransfers per second, not megahertz. Manufacturers and most BIOSes, however, display "3200 MHz" out of commercial habit. This is not a problem in itself, but it explains why some monitoring utilities show 1600: they read the actual clock, not the throughput.

The theoretical bandwidth is derived from this value. A module's bus is 64 bits, or 8 bytes. Multiply the MT/s by 8 and you get MB/s.

Memory	Actual clock	Throughput	Bandwidth per channel	In dual channel
DDR4-2666	1,333 MHz	2,666 MT/s	21.3 GB/s	42.6 GB/s
DDR4-3200	1,600 MHz	3,200 MT/s	25.6 GB/s	51.2 GB/s
DDR4-3600	1,800 MHz	3,600 MT/s	28.8 GB/s	57.6 GB/s
DDR5-5600	2,800 MHz	5,600 MT/s	44.8 GB/s	89.6 GB/s
DDR5-6000	3,000 MHz	6,000 MT/s	48.0 GB/s	96.0 GB/s

These figures are theoretical ceilings: no application reaches them in practice. They are used to compare kits with each other, not to predict a gain in gaming.

DDR5 doesn't just go faster

The leap from DDR4 to DDR5 is not just a matter of frequency. On a DDR5 module, the 64-bit bus is split into two independent 32-bit sub-channels. A single DDR5 stick therefore already powers two sub-channels, which improves bus occupancy rates for short, scattered accesses—exactly the load profile of a game engine.

DDR5 also integrates its own power management integrated circuit (PMIC) directly on the stick, whereas DDR4 relies on the motherboard's power delivery stage. This is one of the reasons why DDR5 kits remain more expensive at the same capacity.

JEDEC, XMP, and EXPO: the displayed frequency is not the running one

Every stick contains an SPD chip that describes its operating profiles. The JEDEC profile is the standard mode, guaranteed without adjustments: 2,133 to 3,200 MT/s for DDR4, 4,800 to 5,600 MT/s for DDR5, depending on the generation. This is the frequency at which your machine starts by default.

The profile advertised on the box—DDR4-3600 or DDR5-6000—is actually a factory-validated overclocking profile: XMP for Intel, EXPO for AMD. You must activate it manually in the BIOS, usually via a single option in the memory settings tab.

This is the most important point of this article: a DDR5-6000 CL30 kit left in factory configuration runs at 4,800 MT/s. You have paid for performance that is lying dormant. On machines coming out of our workshop, this profile is activated and validated before shipping; for a personal build, it is the first thing to check after installing Windows.

CL Latency: the figure everyone misreads

A number of cycles, not a duration

CL stands for CAS Latency: the number of clock cycles that elapse between the moment the controller requests a column of data and the moment the data leaves the stick. It is a number of cycles, not a time. And a cycle does not last the same amount of time at 1,600 MHz as it does at 3,000 MHz.

Conclusion: comparing CL16 to CL30 without looking at the frequency makes no sense. The formula that settles it for everyone is simple:

Actual latency (ns) = (CL × 2,000) ÷ throughput in MT/s

Kit	Calculation	Actual latency
DDR4-2666 CL19	(19 × 2,000) ÷ 2,666	14.3 ns
DDR4-3200 CL16	(16 × 2,000) ÷ 3,200	10.0 ns
DDR4-3600 CL18	(18 × 2,000) ÷ 3,600	10.0 ns
DDR5-5600 CL46	(46 × 2,000) ÷ 5,600	16.4 ns
DDR5-6000 CL30	(30 × 2,000) ÷ 6,000	10.0 ns
DDR5-6400 CL32	(32 × 2,000) ÷ 6,400	10.0 ns

The table speaks for itself: a DDR5-6000 CL30 and a DDR4-3200 CL16 show exactly the same actual latency. However, the former provides nearly double the bandwidth. Conversely, an entry-level DDR5-5600 CL46 kit is slower in response time than a good DDR4 kit—while costing more.

Secondary timings, and how much to worry about them

A complete spec sheet often displays four numbers, for example 16-18-18-36:

• CL — the CAS latency described above
• tRCD — delay between activating a row and accessing a column
• tRP — delay to close a row before opening another
• tRAS — minimum time a row stays open

For gaming and office use, CL and frequency are more than enough to distinguish between two kits. Secondary timings come into play when chasing the last few percentage points in manual overclocking, with the associated risk of instability. Our workshop advice: choose a kit with a consistent factory profile, activate it, and leave it at that.

The special case of AM5

On Ryzen 7000 and 9000 series processors, the memory controller and the memory itself can run in a 1:1 or 2:1 ratio. The 1:1 mode keeps latency low; the 2:1 mode allows for higher frequencies but adds a latency penalty that often cancels out the benefit.

This is why DDR5-6000 CL30 has established itself as the sweet spot on AM5: it is the last step comfortably handled in 1:1 on the vast majority of chips. Aiming for DDR5-8000 on an entry-level B650 motherboard is more of a contest than a daily use case.

Dual channel: the free performance boost that gets forgotten

Why 2×8 GB is better than a single 16 GB stick

AMD and Intel consumer processors have a dual-channel memory controller. With only one stick installed, one channel remains empty: you lose half of the available bandwidth, regardless of the kit's frequency.

The impact depends heavily on the context. On a machine equipped with a dedicated graphics card, the gap in gaming is measurable but moderate, and it widens mostly in intensive scenes where the processor feeds the GPU continuously—open worlds, simulations, major multiplayer titles. On a machine without a dedicated graphics card, however, the gap becomes massive: the integrated graphics chip draws directly from system memory, and depriving it of a channel is equivalent to choking it.

For a given budget, two sticks of half the capacity are therefore almost always the right choice. The only case where a single stick makes sense is when you plan to add an identical one very quickly.

The right slots: A2 and B2

On a motherboard with four slots, two sticks shouldn't just be placed anywhere. The almost universal layout dictates filling the second and fourth slots starting from the processor, marked A2 and B2 in the manual. This layout limits the length of the active traces on the board and preserves signal integrity at high frequencies.

Installing the two sticks in A1 and B1, or worse in A1 and A2, leads either to single-channel operation or instability that will manifest as random blue screens weeks later. The motherboard manual always contains the population table: it is thirty seconds of reading well spent.

Four sticks: rarely a good idea

Filling all four slots does not increase the number of channels—it remains at two. However, it doubles the electrical load on the processor's memory controller. The result: stable frequencies drop, sometimes significantly, and this is even more pronounced with DDR5.

If you are aiming for 32 GB, a 2×16 GB kit will be easier to run at full speed than a 4×8 GB assembly. And if you start with 2×8 GB hoping to upgrade later, be aware that adding two sticks bought separately, even with the same part number, guarantees nothing: kits are paired and tested together at the factory, whereas sticks sold individually are not.

Single rank or dual rank

A stick is called "single rank" or "dual rank" depending on how its chips are organized. Dual rank allows for additional interleaving and provides a small gain, at the cost of higher frequencies being harder to achieve. It is a secondary criterion: at this stage, frequency, latency, and the number of channels have already decided the essentials.

What to aim for according to your platform

The type of memory is not a free choice: it is dictated by the processor/motherboard combination. A motherboard accepts either DDR4 or DDR5, never both.

Platform	Mandatory type	Sweet spot	Note
AM4 — Ryzen 5000 (5500, 5700X…)	DDR4	3200 CL16 or 3600 CL18	Mature platform, affordable kits
AM5 — Ryzen 7000 / 9000	DDR5 required	6000 CL30	1:1 ratio to be preserved
Intel LGA1700 — DDR4 motherboard	DDR4	3200 CL16	Configuration used in our Intel PCs
Intel LGA1700 — DDR5 motherboard	DDR5	6000 CL30 to CL36	Extra cost to be balanced based on use

At DSGate, our Intel LGA1700 configurations are assembled with DDR4: on the Core i3 and i5 processors we use, the actual performance gap in gaming compared to DDR5 does not justify the extra cost, and the saved budget is better spent on the GPU or SSD. Some models are also offered in a DDR5 version for those who prefer the other trade-off.

And what about capacity?

Speed figures do not replace capacity. 16 GB remains a comfortable floor for gaming in 2026; 32 GB becomes relevant as soon as you start streaming, editing video, keeping a browser loaded with tabs in the background, or dabbling in simulation. Running out of memory costs infinitely more in fluidity than a few nanoseconds of latency.

The five errors we see most often

• Not activating XMP or EXPO: the kit runs at the JEDEC profile, resulting in up to 20% less throughput.
• Mixing an existing stick with a different part number: the system aligns everything with the slowest profile, if it boots at all.
• Populating A1 and B1 instead of A2 and B2: instability or single-channel operation.
• Buying a single large stick to "upgrade later": half the bandwidth is lost from day one.
• Paying for a frequency the platform won't handle: a DDR5-8000 kit on an entry-level board will drop back to a lower profile.

In summary

Three reflexes are enough. Convert latency to nanoseconds using the formula (CL × 2,000) ÷ MT/s to compare kits honestly. Always install two sticks, in the correct slots. Activate the XMP or EXPO profile in the BIOS, then check the frequency displayed in Task Manager or a monitoring utility.

If you prefer not to deal with this, that is exactly the work we do upfront: every machine assembled in France in our workshop leaves with its memory profile activated, validated under load, and documented. You can browse our RAM sticks and SSDs for an upgrade, or start from a complete build in our gaming PC catalog. For a precise configuration, our custom configuration tool allows you to choose memory capacity and type line by line.