Memory Speed Terminology Guide

What is RAM?

“What is truth?” – Pontius Pilate, The Gospel According to John

Random access memory, or RAM for short, is a temporary, working storage space in a computer. It is an integral part of the way that all modern computing devices function and has been for decades. However, the technology behind memory has changed a lot over the years – and so has the terminology around it, especially related to the speed at which it operates.

I was reminded of this while editing a recent article from our Puget Lab team, about the impact of RAM speed on Threadripper PRO content creation performance. Following some internal discussions, I wanted to share a public guide to the various ways you can describe memory speeds and how they impact each other. This is intended as a high-level overview, so I’ll try to keep the terminology simple and avoid going too deep into details.

The Many Names of Memory Speed

“Many are my names in many countries…” – Gandalf, The Fellowship of the Ring

Clock Frequency (Hz)

The speed of a lot of components inside a computer or other electronic device is measured in clock cycles. This refers to the number of times per second that the electronic signal in the component oscillates, and the label for this frequency is Hertz (Hz). Most people will be familiar with MHz and GHz – millions and billions of Hz, respectively – as the de facto way to refer to the speed of processors (CPUs) in computers. It only paints one small part of the overall performance picture there, but perhaps that is a story for another time.

In early computer memory, a burst of data would be sent once during each clock cycle – so it was natural to refer to the memory’s frequency as its speed. For example, a memory module with a clock speed of 100 MHz would send data 100,000,000 (one hundred million) times per second. Each of those times, some number of bytes (units of data storage, comprised of eight 0 or 1 bits) could be sent to or from the memory module. This speed was common at the turn of the millennium, but change was on the horizon.

In the early 2000s, CPUs were in dire need of more memory bandwidth. Processor clock speeds were pushing past 1GHz, and technologies for multi-threading, like Intel’s Hyperthreading and advancements in dual-core chips, demanded more data from system memory to avoid bottlenecking. Several technologies were put forth to help with this, including bringing dual-channel memory controllers to the mainstream PC ecosystem, and new memory types were also tried. The one that won out and became the standard was DDR, which stands for “double data rate”, and as the name suggests, it was able to send twice as much data in the same amount of time. This was accomplished by sending two bursts of data per clock cycle instead of just one.

However, that meant the frequency itself didn’t change right away. The first DDR modules for PCs still had a clock speed of 100 MHz, but sent twice as much data as their predecessors, so how did people differentiate? Unfortunately, they just doubled the number but kept using the same units, even though it was now inaccurate – calling it 200 MHz memory. This trend has continued to this day, and it is common to see people talking about 6400 MHz DDR5 memory, and to see speeds referenced that way in motherboard BIOS configuration tools, even though the true speed of those modules is half that (3200 MHz).

Transfer Rate (MT/s)

In order to keep the traditional numbers people were used to seeing for memory, but fix the units, a new label was created: megatransfers per second, or MT/s. This takes the true clock speed and multiplies it by the number of times per clock cycle that data is sent, resulting in exactly the same number but with a unit description that is now accurate to reality. Thus, 3200 MHz multiplied by 2 transmissions per clock cycle equals 6400 MT/s!

Bit Rate (Mbps)

Some memory manufacturers, particularly Samsung and SK Hynix, take a different tack. They arrive at what seems to be the same number but with yet another unit label: megabits per second (Mbps). They have used this approach for a long time, and in all fairness, data-per-second is the way that many other computer-related speeds are measured:

• Storage drives are routinely talked about in MB/s or GB/s (capital B is bytes, lowercase b is bits)
• Wired networking had 10, 100, and 1000 Mbps as common standards, and now we are into the Gbps range
• Wireless networking is also into the Gbps territory now, with WiFi 6 & 7

When referring to memory, though, the bit rate is usually narrowed down to a single data lane. Computer memory modules have multiple lanes, each able to transmit one bit (a 0 or 1) at a time, so the speed of that lane ends up being the same numeric value as the MT/s discussed previously. Thus, a memory module operating at 6400 MT/s can also be characterized as running at 6400 Mbps.

Since this is the per-lane transfer rate, if you multiply it by the number of lanes a given module is working with then you can get a per-module speed as well. In the case of normal, consumer-class DDR5 this is 64 lanes (64 bits, or 8 bytes) for a total theoretical bandwidth of 6400 x 64 = 409600 Mbps, or 409.6 Gbps.

JEDEC Standards Naming

The organization that determines standards for computer memory – along with many other technologies – is the Joint Electron Device Engineering Council Solid State Technology Association, or JEDEC for short. When putting together the documentation for new memory speeds, they assign names to them that traditionally come in two parts. The first designates the type of memory, and the second relates to speed. For example, breaking down DDR5-6400, we can see:

• DDR5 means the 5th generation of double data rate memory.
• 6400 matches the MT/s speed rating, as discussed previously.

Standards Naming Part Deux

Another naming convention for memory looks similar, but is calculated a little differently. The same DDR5-6400 discussed above can also be referred to as PC5-51200. This breaks down as follows:

• PC5 has the same meaning as DDR5: the fifth generation of this type of PC (personal computer) memory.
• 51200 is the theoretical maximum bandwidth of the memory, in MB/s (megabytes per second). This is calculated by taking the number of transfers per second (6400 MT/s) and multiplying by the number of bytes sent per transfer (8, for this type of memory).
  • 8 bytes x 6400 MT/s = 51200 MB/s
  • If you multiply by 8 again, to convert from bytes to bits, you end up with the same 409600 Mbps as above!

Other Aspects of Memory Speed

“There is more to this planet than meets the eye” – Optimus Prime, Transformers: The Last Knight

Latency / Timings

The speed of a given module isn’t the only thing that affects how the RAM in your computer performs. Another specification that gets brought up in tech circles is memory timings, which determine latency. That is a measure of the amount of time from when a request is sent to access memory until the desired data is returned. The settings for this, which can usually be controlled through a motherboard’s BIOS utility, are generally given in clock cycles. There are multiple timing parameters, each referring to the time needed for different steps in accessing memory. Sometimes those all end up being the same number, so very similar, so just a single value may be given when looking at specs sheets… but we don’t need to get too far into the details for this discussion.

What does matter is an understanding that smaller numbers mean fewer clock cycles, which sounds like it would translate to lower latency – but there is a twist. This is just the number of clock cycles, but how long each of those clock cycles is will depend on the memory speed as established earlier. So if the frequency / transfer rate / bit rate goes up, and the memory timings also go up by the same percent, then the actual latency in real, measurable time stays roughly the same. In fact, this is what we usually see when looking at the JEDEC specifications for different speeds of memory within a given product line and generation. For example, here are Kingston ValueRAM DDR5 timings for two different memory speeds:

Speed	Timing	Ratio of Speed:Timing	Actual Latency
5600 MT/s	CL46	~121.7	~16.43 ns
6400 MT/s	CL52	~123.1	~16.25 ns

Overclocking

While virtually all memory modules produced now have built-in speed settings based on JEDEC standards, some memory manufacturers push their products to higher speeds and/or lower timings, using Intel’s XMP or AMD’s EXPO memory profiles. These can potentially squeeze a little bit more performance out of your computer’s memory system, but they come at the cost of potential instability. Achieving those higher speeds also often requires increasing the voltage at which the RAM is running. While popular among enthusiasts, we have found that using these settings leads to higher failure rates of memory modules and increased error rates in professional software.

After much testing, repeated periodically over the years, we have decided to protect our customers from those risks by only utilizing memory that runs within CPU manufacturer specs and at default JEDEC settings in our workstations, laptops, and servers.

Memory Channels

The memory speeds that were the focus of this post have increased over time, generally keeping ahead of advancements in processor clock speeds, but the explosion from a single core per physical CPU twenty years ago to as many as 192 cores in AMD’s EPYC™ 9965 today means that faster memory alone is not enough. Each CPU core needs to be continually fed with data from the RAM in order to run calculations and execute code, so with more cores comes a need for greater overall bandwidth. The solution to this problem is having multiple memory channels.

How many channels of RAM a given computer system supports is determined by the memory controller, which started out as a separate part of the chipset on motherboards but has now been integrated directly into the CPU for most platforms. Generally speaking, the more cores a given family of processors offers the higher the number of memory channels it will support. The table below demonstrates this across the CPUs from AMD and Intel that are used in our workstations (circa 2025):

Manufacturer	CPU Family	Max Core Count	Max Thread Count	Memory Channels
AMD	Ryzen™ 9000	16	32	2
AMD	Threadripper™ 9000	64	128	4
AMD	Threadripper™ PRO 9000WX	96	192	8
AMD	EPYC™ 9005	192	384	12
Intel	Core™ Ultra 200 Series	24 (8P + 16E cores)	24	2
Intel	Xeon® W-2500	26	52	4
Intel	Xeon® W-3500	60	120	8

This is the final piece of information needed in order to complete the picture of memory speed. If you want to calculate the theoretical maximum bandwidth for the whole memory system on a computer, simply take the speed in MT/s (or Mbps of a single lane), multiply it by the number of lanes / bits per transmission (64 for DDR5), and then multiply again by the number of memory channels. For example: 6400 MT/s x 64 bits x 8 channels = 3,276,800 Mbps total theoretical memory bandwidth for AMD’s Threadripper PRO 9000WX. Real-world throughput won’t be quite that high, of course, but that’s the cap.

Conclusion

“The end” – Millions of books and movies

That is the story of memory speed in a nutshell. There are a lot of ways to refer to it, some more accurate or easier to understand than others, but all in use by different people and companies. In general, here at Puget Systems, we are trending toward the use of the official JEDEC naming for products and MT/s when writing about speed – though many of us who have been with the company for a long time grew up with MHz, back when it was actually correct (before DDR came along). If you see other units being referenced, though, now you’ll understand how they relate to each other.

Hopefully this guide has been helpful! If you have any questions or insights, please share them in the comments.

---