Processor and graphics card specifications are live

Problems · How to check

What is the difference between memory bandwidth and memory speed?

Memory speed is a rate: how many transfers per second the memory runs at, written as a number like DDR5-6000. Memory bandwidth is a quantity: how much data actually moves between memory and the processor per second, written in gigabytes per second. Speed is one input to bandwidth, and on its own it predicts real performance poorly.

The reason they diverge is that bandwidth also depends on how many channels are populated, whether the advertised profile is switched on, and how much the memory controller can keep in flight at once. A machine with fast modules in a single channel routinely delivers less bandwidth than slower modules correctly installed in two.

First, notice when it happens

From the first minute

Memory limits show up immediately and constantly rather than building over time. If a machine is slow from the first second, stays slow, and feels worse the more applications are open, memory is a reasonable suspect. Bandwidth and latency do not warm up, so they read the same in second two as in minute ten.

Only after minutes of use

Performance that is fine at first and degrades over minutes is almost never memory. That pattern belongs to heat and power limits. Memory being the ceiling looks like a flat, uniformly lower line rather than a falling one, which is why the two are worth separating before spending money on either.

The causes, ranked

1. One channel populated instead of two

most common

How to confirm it

Measured multi-thread bandwidth is close to single-thread bandwidth rather than well above it. Physically, one module in a machine with two or four slots, or two modules in the wrong pair of slots.

The fix

Add a matching second module, or move the existing pair into the slots the manual specifies, which are usually not the two adjacent ones. This is often the single largest and cheapest performance change available on a budget desktop.

2. The advertised profile was never switched on

common

How to confirm it

The modules are rated well above the speed the system reports. Memory defaults to a conservative baseline until the XMP or EXPO profile is enabled in firmware, so fast modules quietly run slow.

The fix

Enable the XMP or EXPO profile in the firmware setup. It is a single toggle on most boards. Test afterwards, because an unstable profile can cause crashes that are easy to blame on something else.

3. The memory controller, not the modules, is the ceiling

common

How to confirm it

Bandwidth barely improves when more threads read at once, even with two channels populated and the profile enabled. Common on low-power laptop processors and on integrated graphics, which share the same memory.

The fix

There is no fix short of a different machine, but there is a decision: on a shared-memory laptop, faster modules genuinely help because the graphics is competing for the same bandwidth, whereas on a desktop with a discrete card the gain is usually small.

4. Latency rather than bandwidth is the actual problem

occasional

How to confirm it

Bandwidth measures healthy but the machine still feels unresponsive. Latency is the round trip to fetch one value, and workloads that chase pointers rather than stream data are limited by it instead.

The fix

Lower timings help a little, but the honest answer is that latency has improved slowly for years and is rarely worth buying for. Confirm it is the bottleneck before spending anything.

Why the sticker number overstates what you get

The figure printed on a module is a theoretical maximum: transfers per second multiplied by the width of the bus. Real bandwidth is always lower, because a memory controller cannot keep the bus perfectly busy, refresh cycles take time, and a single thread cannot ask for data fast enough to saturate a modern channel on its own.

That last point is why one-thread and all-thread bandwidth are worth measuring separately. On a memory system with headroom, adding threads adds bandwidth. On one already saturated by a single thread, it does not, and the ratio between the two is the diagnosis.

Below roughly 1.35 times, the memory system rather than the cores looks like the ceiling, which frequently means one channel is doing the work of two.

Which number should decide a purchase

For a desktop with a discrete graphics card, channel configuration first, then capacity, then speed. Going from one channel to two is often worth more than going from DDR5-4800 to DDR5-6400, and it usually costs less.

For a laptop or a machine using integrated graphics, speed matters more than it does elsewhere, because the graphics is reading from the same memory as the processor and every megabyte per second is shared between them.

For everyone, capacity beats both once you are actually running out. No amount of bandwidth compensates for a machine that has started swapping to disk.

Measure it instead of guessing

SystemCheck measures memory in the browser rather than reading a specification: bandwidth from a single thread and from every thread at once, and the round-trip latency to cache and to main memory using a random permutation walk the processor cannot prefetch. The ratio between the one-thread and all-thread figures is reported directly, because that ratio is what indicates a single populated channel.

What software cannot tell you here

  • How many channels are populated, the module capacity, the clock or the timings. No browser exposes any of it, so the report infers a likely single-channel configuration from the bandwidth ratio and says so as an inference.
  • Whether the XMP or EXPO profile is enabled. It shows up only as bandwidth lower than the modules should deliver, which requires you to know what they are rated at.
  • Whether a module is faulty. Bandwidth and latency measurement is not an error test.

Asked, in people’s own words

Is higher memory bandwidth or higher memory speed more important?
Bandwidth, because it is the thing that actually affects your machine and speed is only one of its inputs. A system with slower modules in dual channel usually delivers more bandwidth than faster modules in single channel, which is why the configuration matters more than the number on the sticker.
How do I know if I am running single channel?
Compare bandwidth measured from one thread with bandwidth measured from all threads. If the all-thread figure is barely higher, the memory system is already saturated by one thread, which usually means a single populated channel. Physically, check whether more than one slot is filled and whether the filled ones are the pair the manual specifies.
Does RAM speed matter for gaming?
It matters most when the graphics is integrated, because the graphics and the processor are then sharing the same memory. With a discrete graphics card the effect is real but modest, and getting the channel configuration right will usually do more for frame rates than buying a faster kit.
What is a good memory bandwidth figure?
It depends entirely on the platform, so an absolute number is less useful than a comparison. The more informative check is the ratio between single-thread and all-thread bandwidth on your own machine, because a low ratio indicates a configuration problem you can fix rather than a hardware limit you cannot.

Name the cause. Three and a half minutes, in the tab you have open.