Memory Tuning: Where XMP Ends and the Fun Begins

I'm Gilgamesh, and yes, I know.

Last time I promised I'd be back every week with another hardware article. It's been a little more than a week since then.

All right. A lot more.

I was quietly hoping my boss would eventually forget the whole idea and let me get back to fixing computers instead of making me sit here writing about them. He didn't.

The truth is, putting together an article like this can easily eat up an entire day for me. Not because I don't know what I'm talking about, but because I still have machines to look at, phones to answer, and somebody always walks in with one of those "it'll only take five minutes" jobs.

It is never five minutes.

By the time I get back to the article, I can't remember whether I was writing about tRFC or diagnosing another dead power supply. So, late but not defeated, here's the next one.

Last time we looked at the differences between DDR4 and DDR5, talked about LPDDR, XMP and EXPO profiles, and went through some of the common problems you can run into with modern memory.

This time we're going deeper. We're getting into the part manufacturers would probably prefer to explain with one simple sentence:

"Just enable XMP."

Sure.

A bit like Windows 98 and "Plug and Play." Those of us who actually worked on computers back then remember that sometimes there was a lot more "pray" than "play."

Memory tuning can be much the same. It can be simple. Or it can become the kind of rabbit hole where it's three in the morning and you no longer remember why you're adjusting tRFC. You only know that yesterday your latency was 58 nanoseconds and now it's 57.4.

Welcome to the club.

What Does Memory Tuning Actually Mean?

A lot of people hear "memory overclocking" and think only about frequency.

DDR5-6000 becomes DDR5-6200. DDR4-3200 becomes DDR4-3600. Done. Faster memory.

Well... not exactly.

Memory performance isn't determined by frequency alone. Timings matter. The memory controller matters. Voltages matter. And perhaps most importantly, it matters how all these things work together.

Memory running at a higher frequency with poor timings can, in some situations, perform worse than a slightly slower kit with properly tuned timings. This is why I always say: don't stare at one number.

Hardware manufacturers love big numbers. It's much easier to print "7200 MT/s" on a box than explain in three paragraphs how that memory actually behaves with different processors, motherboards and memory controllers.

MHz or MT/s?

Let's clear up an old misunderstanding while we're here.

People commonly describe DDR memory speeds in MHz, but technically, when we're talking about the advertised effective data rate, MT/s is the more accurate term.

Take DDR4-3200. That's 3200 MT/s, three thousand two hundred million transfers per second, while the actual memory clock is roughly 1600 MHz.

DDR stands for Double Data Rate. Data is transferred twice per clock cycle.

In everyday conversation, people will still call it "3200 MHz RAM." I'm not going to kick your door down with a soldering iron if you say it that way. But it's useful to understand what's actually happening under the heatsink.

Timings: CL16, CL30 and the Other Hieroglyphics

Look at a memory kit and you'll see numbers like 16-18-18-36, or with DDR5, perhaps 30-38-38-96.

These are not lottery numbers.

The one most people know is CAS Latency, or CL. Very simply, it represents the number of clock cycles involved in a particular memory operation.

But here's the trap. A lower CL number does not automatically mean faster memory.

DDR4-3200 CL16 and DDR5-6000 CL30 might make the CL16 kit look much faster at first glance. But they're operating at completely different frequencies. Timings always need to be considered together with frequency.

Then we have tCL, tRCD, tRP, tRAS, tRFC, tREFI and enough other settings to keep you occupied until retirement. Open the advanced memory section in some BIOSes and it looks like you've accidentally entered the maintenance menu of a Boeing.

Start with the primary timings. Secondary and tertiary timings are where you can spend enormous amounts of time chasing relatively small improvements.

For some people, that's pointless. For those of us who enjoy tinkering, that's half the fun.

XMP and EXPO: Factory-Packaged Tuning

XMP is Intel's memory profile system, while EXPO is AMD's solution. The profile stored with the memory configuration allows the motherboard to apply the manufacturer's intended frequency, timings and voltage automatically.

It's convenient. Enable it in BIOS. Reboot. And, if you're lucky, you're done.

The important part is: if you're lucky.

XMP and EXPO are not magic. Just because a particular speed is printed on the memory box doesn't mean every CPU, motherboard and BIOS version will run that profile in exactly the same way.

The memory controller on modern systems is generally inside the processor. And two CPUs of exactly the same model can behave differently. One chip may happily run your memory settings all day. Another one may throw errors with exactly the same settings.

Silicon lottery. Old story. New hardware.

Fine-Tuning XMP and EXPO

I normally recommend starting with the factory profile. Enable it. Boot. Test. If it's stable, you've got a working baseline.

Only then start changing things. You can try tightening certain timings, or, with suitable hardware, carefully push the memory frequency higher.

But change one thing at a time. This matters. If you change ten settings at once and the machine stops booting, you've got no idea which one caused the problem.

I still take notes. On paper, sometimes. Laugh if you want. Paper doesn't crash, doesn't need a firmware update, and thirty years from now it'll probably still be compatible with a pencil.

Gear Modes: When the Memory and Controller Start Negotiating

On Intel systems you'll come across settings such as Gear 1 and Gear 2, with other Gear configurations existing on certain platforms as well. Put simply, these settings affect the relationship between the memory clock and the memory controller.

Gear 1 keeps the controller working more closely with the memory, which generally gives you better latency. The problem is that this becomes increasingly difficult to maintain as memory frequency rises.

That's where Gear 2 comes in. The memory controller runs at a lower effective rate relative to the RAM, making higher memory speeds easier to achieve, but usually at the cost of additional latency.

And once again, we arrive at the same lesson: the bigger number isn't automatically better.

You can have an impressive memory frequency, but if the controller configuration adds enough latency, the real performance gain may be much smaller than the number on the box suggests.

Command Rate: 1T or 2T?

Command Rate, in simplified terms, affects the timing used by the memory controller when sending commands to the memory modules.

1T generally offers slightly better performance and lower latency. 2T is more relaxed, but often easier to stabilise.

With two modules and sensible frequencies, the more aggressive setting may work perfectly well. But if you fill all four DIMM slots, push the frequency, tighten the timings and expect everything to behave, don't be surprised if the system simply says:

Nyet.

Sometimes relaxing the Command Rate is worth far more than another half-day of swearing inside the BIOS.

Two Modules or Four?

This one matters.

Some people look at four RAM slots on a motherboard and assume all four are supposed to be filled.

No. Four slots are an option, not a mission objective.

At high memory frequencies, two modules generally put less stress on the CPU's memory controller, making it easier to achieve a stable configuration. With DDR5, this can become particularly noticeable.

A motherboard and CPU combination that happily runs high memory speeds with two DIMMs may need a considerably lower frequency when all four slots are populated. So if you need 64 GB, for example, 2×32 GB can often be a better choice than 4×16 GB.

Of course there are exceptions. There are always exceptions.

PC hardware is like that. If there were one rule that worked for every configuration, I'd have been unemployed years ago.

DDR5 Memory Training: Don't Pull the Plug Yet

I mentioned memory training in the previous article, but it's particularly important when tuning DDR5.

After changing certain BIOS settings, the system may perform memory training again. During this process you can be left staring at a black screen for quite a while. The machine may reboot. Sometimes several times.

This doesn't necessarily mean anything is broken. The motherboard is trying to find settings that allow the memory to operate correctly.

Now, if ten minutes later the machine still has all the signs of life of a dead Pentium III, then yes, it's probably time to investigate. But give it a chance first.

And When It Doesn't Boot at All?

Sooner or later, this happens.

Black screen. Fans spinning. RGB lighting working perfectly, because naturally that's the one thing that never fails. No picture.

First, wait and make sure the system isn't still memory training. If nothing happens, it's probably time for a Clear CMOS.

Some motherboards have a dedicated button. Others use a jumper. On older or simpler boards, you may be dealing with the battery and the appropriate reset procedure.

This is why I recommend reading the motherboard manual before you start memory tuning. Not afterwards. Before.

The manual isn't just there to collect dust at the bottom of the box.

It Looks Stable. That Doesn't Mean It Is.

This is probably one of the most important rules in the entire article.

Just because Windows boots doesn't mean your memory is stable. And neither does playing a game for ten minutes.

Memory instability can be sneaky. The computer may run perfectly for hours, then suddenly a program closes. Blue screen. Game crash. Corrupted file. Broken archive. Or Windows simply starts behaving strangely.

The worst memory error isn't necessarily the one that crashes your computer immediately. At least that one is honest.

The nasty ones are the errors that quietly corrupt something once a week.

How Do You Test It?

Always test after memory tuning.

There are plenty of tools for this. MemTest86 is a useful starting point, while Windows users also have access to more demanding memory and stability testing software. People who tune memory seriously often use tools such as TestMem5 with suitable configurations, Karhu RAM Test, HCI MemTest or OCCT.

I don't like trusting a single test blindly. If the machine matters, I use different types of workloads. Memory can pass one test and fail another.

And don't test for five minutes. In five minutes even a bad power supply can pretend to be a good boy.

Voltage Isn't Free Performance

When something becomes unstable, the first instinct of many beginners is: give it more voltage.

Sometimes that works. But higher voltage also means more heat and more electrical stress, and not every stability problem can be solved by feeding the hardware more power.

With DDR4, 1.35 V is completely normal for many XMP kits. You'll see higher voltages when tuning, but what makes sense for long-term use depends on the particular memory ICs, motherboard and intended use.

DDR5 brings even more voltage parameters into the game. VDD. VDDQ. CPU-side voltages related to the memory controller.

And this is where blindly copying somebody else's settings from the internet becomes a bad idea.

"It works for him at 1.45 volts."

Wonderful. His CPU isn't your CPU. His motherboard isn't your motherboard. And his memory chips may not even be the same as yours, despite the same brand name being printed on the box.

Don't blindly copy voltages.

Memory Temperature Matters Too

Years ago, manufacturers put a piece of aluminium on a RAM module, called it a heatspreader and everybody went home happy.

DDR5 modules are more complicated pieces of electronics. With the PMIC on the module and higher operating speeds, temperature becomes something worth watching when you're tuning, especially once you start raising voltages.

Memory sitting above a large graphics card inside a poorly ventilated case is living a very different life from memory running on an open test bench.

This is why I like testing under real conditions. I don't care how impressive a RAM kit looks in a five-minute benchmark with the side panel removed.

I want to know whether it's still stable in August, in a warm room, after three hours of gaming. That matters.

When Is Manual Tuning Worth It?

If you enjoy understanding how your machine works, you've got the time to test properly, and optimisation interests you, memory tuning can be great fun.

On some systems you can gain measurable performance. Integrated graphics can benefit particularly from memory bandwidth because the GPU shares system memory. CPU-limited games can also show improvements. Certain compression, rendering and computational workloads may benefit as well.

But don't expect miracles. A mid-range PC isn't going to turn into a space station because you spent two evenings adjusting tRFC.

And When Should You Leave It Alone?

If the machine is used for work and every minute of downtime costs money, I'd be much more conservative. A server. A critical workstation. A machine handling important data.

Or maybe you simply don't care about tuning. Enable the appropriate XMP or EXPO profile, test it properly and use the computer.

You don't have to tune everything just because the option exists. There comes a point where an extra one or two percent of performance isn't worth sacrificing stability.

When I was younger, I could spend the whole night fighting for one more benchmark point. These days I'd rather sleep.

A man develops with age. Or maybe he just gets old.

A Few Rules From the Workshop

• Start with a stable baseline.
• Save a working BIOS profile if your motherboard supports it.
• Change only one or two parameters at a time.
• Write down what you changed.
• Test after every significant adjustment.
• If you get errors, don't immediately start increasing voltages.
• When you've reached the point where you don't even remember why the machine stopped booting, go back to the last known stable configuration.

That last one is much faster than changing another twenty random BIOS settings to fix the ten random BIOS settings you changed before.

I didn't learn that from a book. Unfortunately.

Benchmarks Aren't Everything

There's one more thing worth remembering. It's very easy to get dragged into a numbers war on overclocking forums.

"Mine does 55 ns." "Mine does 52." "I'm running 8000 MT/s."

Lovely.

The real question is what you gain from it in actual use. If your machine scores eight percent better in a memory benchmark but the software you use all day runs half a percent faster, you have to decide whether spending three evenings tuning it was worthwhile.

If you did it as a hobby? Of course it was.

Hobbies don't have to make financial sense. If they did, nobody would restore forty-year-old cars, and I wouldn't still have computer parts in my workshop that probably belong in a museum.

But if performance is the only goal, benchmark the applications you actually use as well.

Finally

Memory tuning is one of the more interesting parts of PC optimisation because it quickly teaches you just how interconnected a modern computer really is.

It isn't only the RAM. Processor. Memory controller. Motherboard. BIOS. Memory ICs. Voltage. Temperature. Even the number of modules and which DIMM slots they're sitting in can matter.

Everything talks to everything else. Sometimes politely. Sometimes like two drunk neighbours at four in the morning.

That's why there is no universal "best RAM setting" you can copy into the BIOS and forget about.

Good memory tuning isn't about finding the biggest number on the internet. It's about finding the point where your particular hardware is fast, stable and reliable enough to stay that way.

And once you find it? Save that BIOS profile. Seriously.

Because two months from now you'll update the BIOS, everything will disappear, and you'll be sitting in front of the monitor trying to remember whether your tRFC was 480 or 580.

I've done it. More than once.

Next time

Another subject I promised to cover earlier: what the hell happened to memory prices?

We'll look at why RAM can be ridiculously cheap during one period and then, only a few months later, the same capacity costs considerably more. We'll talk about DRAM production, the games of supply and demand, how AI and the enormous memory appetite of data centres affect the market, and why DDR4 suddenly became interesting again just when everyone was preparing its funeral.

More importantly, I'll give you some practical advice on buying memory without personally financing half of the memory industry's next quarterly report.

Because buying new technology is easy. Buying it at the right price is a skill.

Until then, remember one rule:

C'est la vie.