RAM Is Expensive — Can an SSD Stand In? Checking the Latency and the Lifespan

2026年8月30日

19 August 2026

RAM prices are three to six times off the bottom.

Local AI runs larger models the more memory you have. But go to buy more and 64GB of DDR5 is over $600. Which led to a thought.

Could an SSD stand in for memory? A 1TB SSD is a fraction of that. Group a few and they should get faster, and the price per gigabyte is a small fraction of RAM.

Would that actually be enough? I looked into it.

This article is compiled from published information. It is not the result of building it here (checked 19 August 2026).

SSD endurance and grouping were covered in the previous article.

What I looked at, and where from

The scope, up front.

• What: the latency gap between RAM and an NVMe SSD; what using an SSD as memory does to its lifespan; and what to do instead
• Sources: an NVMe latency reference, a peer-reviewed study measuring lifespan when flash is used to extend memory, and a manufacturer’s endurance documentation (linked at the end)
• Checked: 19 August 2026
• What is from here: the GPU measurements below, taken on 19 August 2026

Using an SSD as memory is an old idea

Every operating system already includes a mechanism for using a disk as overflow when memory runs short.

No special preparation is needed. It is already running. Add an SSD, give it a large area, and your apparent memory goes up.

The question is what happens when you do.

RAM and SSDs are three orders of magnitude apart in latency

This was the clearest thing in everything I read.

The important part is that what matters here is not “how many GB per second" but “how long from asking to receiving".

Generating text, the model is traversed once per token. Every one of those waits. Moving more data per trip does not shorten any individual wait.

Grouping SSDs in RAID does not shorten the wait

This is the part I most wanted to write.

Four SSDs in parallel move more data per second. But the time from asking to receiving is barely different from one drive.

Four people carrying boxes move four times as many boxes, and any single box still takes as long to arrive.

On whether RAID improves latency at all, what I read was divided. Some report it feeling quicker; others point out that the RAID layer itself adds delay. I did not find material solid enough to settle it either way.

But whichever it is, it does not close a gap of three orders of magnitude. That latency gap is what decides this.

Used as memory, it consumes the SSD’s life continuously

The other issue is lifespan.

As covered last time, an SSD has a fixed number of write cycles. Around 1,000 for the current mainstream TLC, and fewer for the cheap high-capacity QLC drives.

And from what I read, normal use essentially never exhausts that. It becomes a problem under continuous heavy writing.

There is a study that measured it. Using a specialised SSD known for endurance, under everyday workloads, it reports a lifespan of 2.4 to 4.5 years.

That 2.4–4.5 year figure is for a drive in a far more durable class. Doing the same thing on a cheap QLC drive would reasonably be expected to be shorter.

The same thing happened on my own GPU the moment capacity ran out

This article is mostly research, but there is a related figure from here.

The setup first.

Run a model that does not fit in the card’s memory and the overflow is pushed onto the PC’s own RAM. That is the same shape as the question here: using somewhere else in place of memory.

Going 1.3GB over dropped it to a thirteenth. That is what “push the overflow somewhere else" costs.

And the overflow here went to the PC’s RAM. An SSD is another order slower again. What happens if you push it there is not hard to picture.

So what do you do when memory is short?

Ending on “I do not recommend it" would be unhelpful, so here are the alternatives.

1. Choosing a different model often solves it

This was the biggest surprise in my own measurements.

The 101GB model is more than five times faster than the 39.6GB one. Size is not what decides it.

The reason is how they are built. The 101GB one is large in total but only uses a small part of itself per token — an architecture called MoE. Models built that way run at usable speed despite their size.

Before adding memory, it can be quicker to reconsider which model you run.

2. Buy a complete machine with the memory already in it

Right now there are cases where a mini PC with memory fitted costs less than buying the memory alone, because manufacturers are still shipping stock bought before the price rise.

One condition, though. It has to be a design where the CPU and graphics share the whole pool of memory, or large models will not load. Get that judgement wrong and the cheap purchase does not do the job.

An example with a lot of memory fitted. It is all in the machine, so there is nothing to add later. With standalone memory as expensive as it currently is, the complete machine can work out cheaper.

3. As a place to keep models, an SSD is exactly right

Everything above says do not use it as memory. As a place to store models, an SSD is doing what it is for.

Read once at load time, and there is no sustained writing, so nothing is consumed. My own models sit on an SSD.

4. Rent it when you need it

If you touch a large model a few times a month, renting by the hour costs less. Buying means paying for electricity during all the hours you are not using it.

An example of renting. You pay by the hour, so for occasional work with large models it can come out below the cost of hardware.

I did not build it: what this article cannot tell you

This article is compiled from published information. I did not actually configure an SSD as memory and try it.

• “2.4 to 4.5 years" is a research figure for a specialised endurance-focused SSD under everyday workloads. It is not a number for running local AI on a cheap drive
• The two models compared are different products. It is not the same model resized
• The 3–4x write amplification also varies with use and with the product

I still wrote “I do not recommend it" because the three-order latency gap is not a margin that product differences close. That much can be said without building it.

In summary: it is possible, and I do not recommend it

• The mechanism ships with every OS. Add an SSD, give it a large area, and apparent memory goes up
• But the time from asking to receiving differs by three orders between RAM and an SSD (nanoseconds against microseconds)
• Grouping in RAID does not change that. Throughput rises; the individual wait does not shrink
• It consumes the drive continuously. A study on a durable-class SSD put the life at 2.4–4.5 years
• The alternatives are choosing a different model, buying a machine with memory fitted, or renting

Looking into it, I found that while trying to find the cheap way, I had nearly picked the expensive one. An SSD is a consumable. Wear one out in a few years and what you spend over that period starts to approach the price of RAM.

If you do try it, use a drive you are willing to lose, for something you can stand being slow. Better than not running at all is a real situation. But I would not make it a permanent arrangement — that is my judgement after reading.

Sources (SSD endurance and memory latency)

Checked 19 August 2026. Measurements here were taken on a GMKtec EVO-X2 (Ryzen AI Max+ 395, 128GB unified memory) / Ubuntu 26.04 / Radeon RX 9060 XT 16GB over OCuLink / integrated Radeon 8060S / llama.cpp Vulkan, on 19 August 2026.