Explainer

DDR5 Speed vs Capacity for AI: What Matters More

In this article
  1. What each one does for AI
  2. How much capacity a model needs
  3. What speed buys you
  4. Why channels and modules matter
  5. When faster memory pays off
  6. Quantisation and mixture-of-experts models
  7. How memory is split between GPU and system RAM
  8. Reading a kit label
  9. A simple rule for choosing
  10. What this means when buying
  11. Frequently asked questions
  12. Sources & credits

Short answer

Capacity matters more. If a model does not fit in memory, it will not run at all, or it will crawl. Speed then decides how quickly tokens come out on the CPU. A published CPU-only test gained roughly 17 to 19 percent from DDR5-4800 to DDR5-6000, and llama.cpp notes tie generation speed to memory bandwidth.

  • Buy capacity first, then pick the fastest speed your processor and board can hold.
  • Memory speed matters for CPU inference and for the part of a model that sits in system RAM; it matters little when the model fits in GPU memory.
  • Two modules beat four: AMD lists DDR5-5600 for two DIMMs and DDR5-3600 for four on the Ryzen 9 9950X.
  • Channel count has a bigger effect than a few hundred MT/s of rated speed.

What each one does for AI#

Capacity is the amount of memory. Speed is how fast data moves in and out of it, measured in MT/s, or megatransfers per second. A language model is a very large file of numbers called weights. To produce each word, a CPU reads through most of those weights again.

That is why the two settings do different jobs. Capacity decides whether the whole file can be held in memory. Speed decides how quickly the processor can read it. The llama.cpp developer Johannes Gaessler describes CPU generation as limited by memory bandwidth, and says speed is almost proportional to memory frequency.

There is a third factor, which is the number of channels. A standard desktop has two channels. Each channel moves data in parallel, so the total bandwidth is the product of speed and channels. Capacity does not change that ceiling.

How much capacity a model needs#

The Ollama library publishes download sizes for each model and quantisation level. Quantisation shrinks the weights by storing them with fewer bits, so a 4-bit version is much smaller than a 16-bit one. The table lists sizes for a range of popular models.

Table
ModelQuantisationFile size
Qwen3 8BQ4_K_M5.2 GB
Qwen3 14BQ4_K_M9.3 GB
gpt-oss 20BOllama default14 GB
Qwen3 30B-A3B (MoE)Q4_K_M19 GB
Qwen3 32BQ4_K_M20 GB
Qwen3 32BQ8_035 GB
Llama 3.3 70BQ4_K_M43 GB
Llama 3.3 70BQ8_075 GB
gpt-oss 120BOllama default65 GB
Llama 3.3 70BFP16141 GB
Qwen3 235B-A22B (MoE)Q4_K_M142 GB

Add working room on top of the file size. The operating system, your applications and the context cache all need memory. Corsair reports that a 70B DeepSeek-R1 distill did not load in 32 GB or 48 GB on the CPU alone. Its advice is 64 GB as a floor and 96 GB for a longer-lived system.

Reading the table: a 64 GB kit loads Llama 70B at Q4_K_M. A 96 GB kit adds room for gpt-oss 120B. A 128 GB kit reaches 70B at Q8_0. None of the desktop sizes holds the largest models at 4-bit. See how much RAM for a local LLM for a longer walk through sizes.

What speed buys you#

Speed helps when the CPU does the reading. Maxim Saplin measured CPU-only generation in LM Studio on a Core i5-13600KF with two 32 GB sticks. His results are community-reported, not a lab test.

Mistral 7B Q6_K (5.94 GB) CPU-only generation, prompt: Tell me about Mars: Token generation speed (tokens/s, higher is better)

Mistral 7B Q6_K (5.94 GB) CPU-only generation, prompt: Tell me about Mars: Token generation speedDDR5-6000: 11.3 tokens/sDDR5-600011.3 tokens/sDDR5-4800: 9.66 tokens/sDDR5-48009.66 tokens/s

2 products. Highest: DDR5-6000, 11.34 tokens/s. Lowest: DDR5-4800, 9.66 tokens/s.

Mistral 7B Q6_K (5.94 GB) CPU-only generation, prompt: Tell me about Mars: Token generation speed
Producttokens/s
DDR5-600011.34
DDR5-48009.66
Source: Maxim Saplin (DEV Community) (opens in a new tab). LM Studio 0.3.4, 12 threads, 2 x 32 GB DIMMs. Intel Core i5-13600KF (overclocked), Windows 11, LM Studio 0.3.4, 12 threads, 100% CPU offload; post dated 12 Oct 2024, edited 17 Feb 2025; community blog post, not a publication lab.

About this data: figures are copied from the named source, not measured by us, and the test system and settings are those the source states above. See all benchmark sources and credits.

Mistral 7B at Q6_K generated 9.66 tokens/s on DDR5-4800 and 11.34 on DDR5-6000. A token is roughly a short word or word fragment. The rated speed rose by 25 percent and the output rose by about 17 percent.

Llama 3.1 8B f16 (16.07 GB) CPU-only generation: Token generation speed (tokens/s, higher is better)

Llama 3.1 8B f16 (16.07 GB) CPU-only generation: Token generation speedDDR5-6000: 4.74 tokens/sDDR5-60004.74 tokens/sDDR5-4800: 4 tokens/sDDR5-48004 tokens/s

2 products. Highest: DDR5-6000, 4.74 tokens/s. Lowest: DDR5-4800, 4 tokens/s.

Llama 3.1 8B f16 (16.07 GB) CPU-only generation: Token generation speed
Producttokens/s
DDR5-60004.74
DDR5-48004
Source: Maxim Saplin (DEV Community) (opens in a new tab). LM Studio 0.3.4, 12 threads, 2 x 32 GB DIMMs. Intel Core i5-13600KF (overclocked), Windows 11, LM Studio 0.3.4, 12 threads, 100% CPU offload; post dated 12 Oct 2024, edited 17 Feb 2025; community blog post, not a publication lab.

About this data: figures are copied from the named source, not measured by us, and the test system and settings are those the source states above. See all benchmark sources and credits.

Llama 3.1 8B at f16, a 16.07 GB file, went from 4.00 to 4.74 tokens per second between the same two speeds. Both results show a clear but sub-linear gain. A Popular AI roundup puts the theoretical dual-channel peak at about 76.8 GB/s for DDR5-4800 and 96 GB/s for DDR5-6000, a rise of 25 percent, while measured output rose about 17 percent. The peak rate is a ceiling that real processors do not reach.

A rough check shows how close the test came. Each token requires reading about the whole file once. At 4.00 tokens per second, a 16.07 GB file implies roughly 64 GB/s of reads, which is about 84 percent of the 76.8 GB/s peak. At 4.74 tokens per second it implies roughly 76 GB/s, about 79 percent of 96 GB/s. This is our arithmetic from the published figures, not a measurement, and it assumes one full pass over the weights per token. It shows why speed tracks bandwidth so closely and why a larger file produces fewer tokens per second.

A bigger model also reads more data per token, which is why the 16 GB Llama file generates fewer tokens per second than the 5.94 GB Mistral file at the same memory speed. Capacity decides which model you can load. Speed then sets the pace for that choice.

Why channels and modules matter#

Channel count often matters more than rated speed. InsiderLLM measured Llama 3.2 3B at Q4_K_M with llama.cpp on an older DDR4 mini PC. It produced 7.36 tokens per second with one stick and 11.59 with two. The hardware is dated, so read the figures as proof of the effect and not as DDR5 numbers.

Llama 3.2 3B Q4_K_M, llama.cpp tg128: Token generation speed (tokens/s, higher is better)

Llama 3.2 3B Q4_K_M, llama.cpp tg128: Token generation speedDual channel: 11.6 tokens/sDual channel11.6 tokens/sSingle channel: 7.36 tokens/sSingle channel7.36 tokens/s

2 products. Highest: Dual channel, 11.59 tokens/s. Lowest: Single channel, 7.36 tokens/s.

Llama 3.2 3B Q4_K_M, llama.cpp tg128: Token generation speed
Producttokens/s
Dual channel11.59
Single channel7.36
Source: InsiderLLM (opens in a new tab). DDR4-2133 SODIMM; llama.cpp b10326 CPU backend, tg128, 4 threads. Intel i7-6700T (4C/8T) Lenovo ThinkCentre M710q, DDR4-2133 SODIMM, llama.cpp b10326 CPU backend, tg128, 4 threads; page dated 11 Aug 2026; self-measured blog on old low-end hardware, shows the channel effect only, not DDR5 desktop values.

About this data: figures are copied from the named source, not measured by us, and the test system and settings are those the source states above. See all benchmark sources and credits.

Phoronix saw the same pattern on a current Core Ultra 7 270K Plus, where llama.cpp ran significantly faster with two DDR5-6400 sticks than with one. Puget Systems tested Threadripper PRO and found that DDR5 speed made little difference to llama.cpp, while cutting the 96-core 9995WX from eight channels to four cost about 14 percent. Bandwidth comes from channels as much as from speed.

The number of modules also changes the speed you can use. The table lists what the makers state.

Table
ProcessorMemory support by processor
AMD Ryzen 9 9950XDDR5-5600 with 2 DIMMs; DDR5-3600 with 4 DIMMs; 256 GB maximum
Intel Core Ultra 9 285KUp to DDR5-6400; 2 channels; 256 GB maximum

Saplin's four-stick, 96 GB setup could not keep its XMP 6000 profile and was stable only at 4800. At that matched speed, two sticks gave 9.66 tokens per second on Mistral 7B and four gave 9.42. Use two large modules when you can.

When faster memory pays off#

Faster is not always better. Phoronix tested a Ryzen 7 9700X with DDR5-8000 against DDR5-6000. The faster kit won the llama.cpp test, but some compile tests did better at DDR5-6000 because the memory controller changes ratio above that speed on the platform.

Phoronix compared DDR5-6000 with DDR5-4800 on a Ryzen 7 9850X3D across more than 300 tests, and PC Games Hardware summarised the average gain as about 1 to 1.5 percent. Most of those tests were not language models.

Saplin adds a related point about threads. With three CPU threads his speed reached 73 percent of the six-thread figure, which suggests that once memory bandwidth is the limit, extra cores add little. Buying a faster processor to speed up a model that lives in RAM will not help much, while buying memory that fits the model will.

The pattern is consistent. Speed pays off when a processor reads a large model from system RAM, and it pays off little in everything else. If your model sits in the memory of a graphics card, system RAM speed barely changes token output. The guide to how much VRAM an LLM needs explains where that line falls.

Quantisation and mixture-of-experts models#

Capacity needs are not fixed, because the same model comes in several sizes. Dropping from Q8_0 to Q4_K_M cuts the Llama 3.3 70B file from 75 GB to 43 GB in the Ollama library.

Mixture-of-experts files are sized in the same way. The Ollama library lists Qwen3 30B-A3B at 19 GB for Q4_K_M and Qwen3 235B-A22B at 142 GB, both in Q4_K_M form, and either one has to be loaded in full.

Choosing a smaller quantisation is therefore the first way to stretch a memory budget. Buyers who plan to run the largest mixture-of-experts models, such as Qwen3 235B at 142 GB, need to look beyond desktop capacities.

How memory is split between GPU and system RAM#

Most local AI setups use both a graphics card and system RAM. Runtimes such as llama.cpp can place some layers of a model on the graphics card and keep the rest in system memory. The layers in VRAM run at graphics-card speed. The layers in RAM run at system memory speed, and they set the pace for the whole token.

Gaessler's notes tie generation speed to memory bandwidth, so the layers that stay in system RAM set the pace for the whole model, while layers in VRAM run at graphics-card speed. Capacity is the safety net in both cases, because it lets the model load at all.

Reading a kit label#

The speed on the box is not the speed you get on first boot. Makers state this plainly. G.Skill says its kits start at the SPD speed, DDR5-5600 for the Trident Z5 and Ripjaws S5 models, until XMP is enabled in the BIOS. Kingston gives the JEDEC default of its FURY kits as 4800 MT/s. TeamGroup says that enabling XMP 3.0 or EXPO counts as overclocking and that the final speed depends on the board, processor and BIOS.

Three terms are worth knowing. JEDEC is the industry baseline that every module can run without extra settings. XMP and EXPO are stored profiles that Intel and AMD boards read to apply the faster speed and timings the kit was tested at. CAS latency, written CL, is the wait in clock cycles before data begins to flow. A lower number is quicker, but bandwidth sets the pace for model reading.

For an AI buyer, the lesson is to check three things before paying for a speed rating. The profile must suit the board. The rated speed should sit near what the processor lists. And the kit should come as one set from one maker, because TeamGroup and G.Skill both advise against mixing kits.

A simple rule for choosing#

Work through the questions in order.

  1. What is the largest model you want to run? Take its file size from the table above and add working room.
  2. Does it fit in GPU memory? If yes, system RAM speed matters little. Choose capacity for your other work.
  3. Will part of it sit in system RAM? Then choose the highest capacity your budget allows in two modules.
  4. Then choose speed. Pick the highest rated speed your processor lists, with a profile your board supports.
  5. Do not trade capacity for speed. A slower 128 GB kit runs a model that a faster 64 GB kit cannot load.

What this means when buying#

Start with the capacity that matches your largest model, then compare kits at that capacity. Our head guide to RAM for AI work compares a 64 GB, a 96 GB and a 128 GB kit. For one capacity in depth, see the best 64GB DDR5 RAM kit and the best 128GB RAM kit for local LLM. The RAM and memory page lists every guide.

Check what Indian retailers actually stock before you decide how much to pay for extra speed, since the kits on sale locally may not match the ones overseas reviews cover.

If 32 GB is your target, the best 32GB DDR5 RAM kit for AI and ML guide compares three kits.

Indian shoppers will find that kits with higher speeds can cost more per gigabyte, so decide the capacity first and then see what speed fits the budget in India.

Frequently asked questions#

Is faster RAM better for AI?

It helps CPU inference, where token speed follows memory bandwidth. It helps little when the model fits in GPU memory. If you run models on the CPU, faster memory is worth having, but capacity still comes first.

Is 128 GB at a lower speed better than 64 GB at a higher speed?

If your model needs more than 64 GB, yes. A model that does not fit cannot run at the faster speed. Capacity decides what you can load at all, while speed only changes how quickly it responds.

Does DDR5-6000 beat DDR5-5600 for LLMs?

Slightly. In Saplin's test, going from DDR5-4800 to DDR5-6000 raised Mistral 7B from 9.66 to 11.34 tokens per second. The gain is modest, so capacity usually matters more than a small speed step.

Why not fill all four memory slots?

AMD lists DDR5-3600 for four DIMMs on the Ryzen 9 9950X, against DDR5-5600 for two. Saplin's four-stick setup also dropped to 4800. A pair of larger sticks reaches the same total at a faster, steadier supported speed.

Does CAS latency matter for AI?

Less than bandwidth. The llama.cpp notes tie CPU generation speed to memory bandwidth, so a lower CAS number rarely changes results by much. Choose capacity and rated speed first, and treat latency as a tiebreaker.

Do I need ECC RAM?

For home use, no. DDR5 includes on-die ECC, which is separate from full error correction on supported workstation boards. Full ECC needs compatible memory and a compatible platform, so it is a workstation choice.

Can I run a model larger than my RAM?

Software can swap data to storage, but the result is very slow. Choose capacity that holds the model. A model that only just fits will run, so leave headroom for the system and for the context you need.

Does RAM speed matter if I have a fast graphics card?

Much less. Token generation is limited by memory bandwidth, according to the llama.cpp developer notes, so a model that sits entirely in VRAM reads at graphics-card speed. System RAM speed then affects loading time and any layers that spill over. Capacity still matters, because the operating system, your tools and any partial offload all draw on system memory.

Sources & credits#

We have not tested these products ourselves. Specs come from the makers; benchmark figures come from the sources we name.

Facts and figures in this article come from these sources. Articles are compiled from the makers' published specifications and the benchmarks named above, and checked by the NextAIGrid Editorial Team.

Disclaimer

This explainer is for information only. Specifications change, so check the maker page. Found an error? Tell us.

How was this explainer put together?

It draws on maker specification pages and published third-party benchmarks, named in the article. We have not tested these products ourselves. Specs come from the makers; benchmark figures come from the sources we name. Specifications and sources were last checked on 17 November 2026. Read how we research and our editorial policy.

Spotted an error? Tell us. Confirmed fixes are listed on the corrections page.

What questions do readers ask next?

Back to top