One review concludes that DDR5-6000 is the sweet spot and everything above it is wasted. Another shows a clear win for 8000. A third finds the whole question worth about two per cent and moves on.
All three ran real tests and reported them accurately. The disagreement is not about the results — it is that almost nobody testing memory is changing one thing at a time.
Nobody is testing frequency
Memory is sold as kits, and a kit is a bundle of settings. When a reviewer compares a 5600 CL46 kit against a 6400 CL32 kit, two variables moved together: the transfer rate went up by 14 per cent and the CAS latency came down by 30.
Whatever the benchmark says about that pair, it cannot tell you which change produced it. And this is not sloppiness — it is what the market sells. Nobody manufactures a 6400 CL46 kit for the sake of an experiment, so a reviewer comparing products is necessarily comparing bundles.
The consequence is that a headline like "6000 beats 8000" is almost always shorthand for "this 6000 kit beat that 8000 kit on this platform in these tests", and the four qualifiers are doing more work than the numbers.
True latency, and why three different kits are the same kit
There is one arithmetic that cuts through most of the confusion, and it takes a line.
CAS latency is quoted in clock cycles, not in time. To get time you need the clock, and a DDR transfer rate is double the clock — so:
latency in nanoseconds = CL × 2000 ÷ transfer rate
Run it on three kits that look completely different on a shelf:
- DDR5-5600 CL28 → 28 × 2000 ÷ 5600 = 10.0 ns
- DDR5-6000 CL30 → 30 × 2000 ÷ 6000 = 10.0 ns
- DDR5-6400 CL32 → 32 × 2000 ÷ 6400 = 10.0 ns
Identical. Every one of those kits answers a request in the same ten nanoseconds. What differs is throughput: the 6400 kit moves 14 per cent more data per second than the 5600 one, at exactly the same responsiveness.
Which is the key to reading any memory review. Those three kits are not three points on a speed scale. They are the same latency at three bandwidths — and whether that matters depends entirely on what the test was waiting for.
Which benchmark you ran decides who wins
Workloads are not uniformly hungry for memory, and the two things memory offers are close to independent.
Latency-bound work is waiting for individual requests. A game's main thread chasing pointers through a scene graph is a long chain of dependent lookups, each of which must complete before the next can be issued. More bandwidth does nothing for a queue of one. This is why games respond to tight timings far more than to headline transfer rates, and why the response shows up in the 1% lows before the average.
Bandwidth-bound work is streaming: compression, encoding, large linear passes over arrays, integrated graphics. These issue many independent requests at once and are limited by how fast data can arrive in bulk, not by how long any single request takes.
So a review whose suite leans on rendering and compression will crown the high-frequency kit. A review whose suite leans on games will crown the low-latency one. Both are right about their suite, and the reader who wants to know about games should discount the first one entirely rather than averaging the two.
The ratios that nobody puts on the box
The second reason reviews diverge is that a memory transfer rate is not a free parameter — it is coupled to other clocks inside the processor, and where the couplings break is platform-specific.
On current AMD desktop parts, the memory controller runs best when its clock and the memory clock stay in step. That relationship holds up to somewhere around 6000 to 6400 MT/s on most samples, and past it the controller has to drop to a divided ratio. The moment it does, the extra transfer rate is paying for a penalty elsewhere, and a 7200 kit can measure slower than a 6000 one on the same board while every number on its label is bigger.
Intel platforms have their own version, where the controller runs in a divided mode above a threshold and the divisor changes what a given transfer rate is worth.
Neither of these is on the packaging, both vary between individual processors, and a reviewer who tested one sample has measured one sample. This is the single largest reason two credible outlets report different conclusions about high-frequency kits: they are describing different silicon lotteries and neither says so.
Subtimings do more work than the number on the front
CAS latency is one of dozens of timings, and it is the only one most kits advertise. Several of the others move results more.
The refresh timings are the clearest case. Memory has to be refreshed periodically, and during a refresh it is unavailable; how long each refresh takes and how often it happens are set by parameters that no consumer kit quotes. Tightening them is one of the more effective things an enthusiast can do, and it is invisible on every product page.
Which means two kits with identical headline specifications — same rate, same CAS, same capacity — can perform measurably differently because their remaining timings were programmed differently, and because the memory dies inside them came from different production. A review of "DDR5-6000 CL30" is a review of the kit that arrived, not of the specification.
Capacity hides a variable inside it too
The same confusion appears in capacity comparisons, and it produces one of the most repeated wrong conclusions in this subject.
Memory modules are organised in ranks — sets of chips the controller addresses as a group. A single-rank module presents one; a dual-rank module presents two, and the controller can interleave between them, working with one while the other is busy. That interleaving is worth a few per cent at the same transfer rate and the same timings, for free.
The trouble is that rank tends to travel with capacity. Smaller kits are frequently single-rank and larger ones frequently dual-rank, because of how many chips it takes to reach the capacity with the dies available. So a review that compares a 2 × 16 GB kit against a 2 × 32 GB kit and finds the larger one faster in games has measured something — but attributing it to capacity is a guess, and where the working set fitted in both, the rank is the better explanation.
This is worth knowing when you buy, because it inverts the usual advice in one specific case: at the same money, the larger dual-rank kit can be faster than the smaller single-rank one even for a workload that never uses the extra capacity. It is also the reason four sticks sometimes measure well in isolated tests — four single-rank modules give the controller the same interleaving — right up until the four-slot frequency penalty takes it all back.
Every kit above the base speed is an overclock
One framing makes the whole subject easier to reason about, and it is rarely stated plainly.
The industry standard defines a base speed that every module must run at. Everything above it — every profile with a brand name on it — is a factory-tested overclock, stored in the module and applied by your board on request. The kit is a promise that these modules reached those settings on the vendor's test platform.
It is not a promise about your memory controller, which lives in your processor and is a different sample from theirs. This is why a kit on the board's compatibility list can still refuse to run at its rated speed, why the same kit works in one machine and not another, and why the answer to "will 8000 work" is genuinely unknowable before you try it.
Read that way, a memory review is a report on how one controller sample handled one kit. The conclusion is real and the confidence interval is one machine wide.
What a useful memory test looks like
The tests worth reading are the ones that hold something still, and they are rarer than they should be:
- One kit, several settings. Take a single high-binned kit and run it at 5600, 6000, 6400 with the timings scaled to hold true latency constant. Now frequency is the only variable and the answer means something.
- Or hold the frequency and vary the timings, which isolates latency instead.
- State the controller ratios. A result at a divided ratio and a result in step are not the same experiment.
- Report both derived quantities — true latency in nanoseconds and bandwidth — rather than the label on the box.
- Use 1% lows for games. Memory's effect on averages is small and its effect on the distribution is not.
What to buy
- Compute the true latency before comparing anything. CL × 2000 ÷ rate. Two kits within a few tenths of a nanosecond are the same kit for your purposes, and the cheaper one wins.
- Stay inside the platform's in-step range unless you intend to tune. On current AMD parts that means the 6000–6400 neighbourhood, and going past it is a project rather than a purchase.
- Buy two sticks, not four — the four-slot penalty is larger than any difference discussed here.
- Verify the profile actually engaged. A kit that silently fell back to its base speed is the most common cause of "memory made no difference".
Two things we would not buy
A high-frequency kit chosen on the number alone for a platform that cannot run it in step. You are paying a premium for a transfer rate the controller will convert into a penalty, and the machine will measure slower than a cheaper kit while looking faster on paper.
And a kit chosen from a review that compared bundles. If the article moved frequency and timings together and concluded that frequency won, it has not shown that — and the conclusion inverts on a suite of different tests. Look for the review that held one of them still, or compute the true latency yourself and ignore the conclusion.
How this was put together
The scaling picture comes from the outlets that publish memory tests as scaling studies rather than product round-ups — Hardware Unboxed and TechSpot's frequency-and-timing comparisons, TechPowerUp's per-title memory runs, and Gamers Nexus's insistence on reporting the platform state alongside the result. The controller-ratio behaviour and the point where the couplings break is documented by the platform vendors in their own tuning guidance, and traced board-by-board in Buildzoid's timing analyses, which is also where the subtiming effects are quantified in a way no product page attempts.
The derived figures are ours: the true-latency arithmetic, and the demonstration that DDR5-5600 CL28, 6000 CL30 and 6400 CL32 are all exactly 10.0 ns — three kits that a shelf presents as a speed ladder and that are in fact one latency at three different bandwidths, which is why the benchmark you happened to run decides which of them wins.
