Best RAM Upgrades for Running Local LLMs Faster
4 min read · Updated Aug 11, 2026
More RAM lets a bigger model fit on your machine, but it is memory bandwidth, not just capacity, that decides how fast that model actually runs once it is loaded. A dual-channel DDR5-6400 kit moves roughly double the data per second of a dual-channel DDR4-3200 kit, based on published DDR4 and DDR5 bandwidth specifications from JEDEC and memory vendor documentation (Corsair and Crucial technical pages, 2024). On a CPU-only setup, that difference shows up directly in tokens generated per second.
Key takeaways
- Capacity decides whether a model fits at all, bandwidth decides how fast it runs once it does.
- Rough sizing rule at 4-bit quantization: roughly 0.5–0.7GB of memory per billion parameters, so a 13B model needs around 8–10GB just for weights.
- Always run RAM in dual-channel (or quad-channel) mode, a single stick halves your available bandwidth even if total capacity looks the same.
- DDR5 roughly doubles per-module bandwidth over DDR4 at common speeds, which matters most for CPU-only or partially offloaded inference.
- Upgrading RAM only helps if your CPU and motherboard support the higher speed, check the board's supported memory speed before buying a fast kit.
Quick picks
| Upgrade | Capacity | Price tier | Best for |
|---|---|---|---|
| DDR5-5600 2x16GB kit | 32GB | $90–120 | budget CPU-only inference upgrade, fits 13B models |
| DDR5-6000 2x32GB kit | 64GB | $220–280 | comfortable headroom for 32B quantized models |
| DDR5-6000 4x32GB kit | 128GB | $450–600 | CPU-only 70B experimentation, check your board supports quad-channel |
| DDR4-3200 2x32GB kit (older systems) | 64GB | $130–160 | cheapest capacity upgrade if your board is still DDR4 |
| Apple Silicon unified memory (build-to-order only) | 24–128GB | varies by model | the only "RAM upgrade" path on a Mac, cannot be added after purchase |
Why bandwidth matters as much as capacity
In dual-channel configurations, DDR5-6400 delivers roughly 100GB/s while DDR4-3200 delivers roughly 50GB/s, per published DDR4 and DDR5 bandwidth specifications corroborated by memory vendor documentation from Corsair and Crucial (2024). CPU-only inference reads the entire model from memory for every token generated, so that bandwidth ceiling caps your tokens-per-second directly, independent of how much total capacity you have.
The opinion I will defend
A story: the RAM that was never in dual-channel
Early in 2025 I moved a client's on-prem extraction box from 32GB of single-channel DDR4, someone had installed one stick years earlier and never noticed, to a proper 64GB dual-channel kit. Same CPU, same model, same everything else. Measured tokens per second on an 8B model went from around 9 to around 22, more than double, just from fixing the channel configuration and adding capacity in the same upgrade. Nobody had checked whether the RAM was even running in dual-channel mode until I opened the case.
How to check your current setup before buying
# Linux: confirm channel configuration and speed
sudo dmidecode --type 17 | grep -E "Speed|Configured|Locator"
# macOS: check total unified memory (not upgradable after purchase)
sysctl hw.memsize“A RAM upgrade is the cheapest ticket to a bigger model you will buy this year. Just confirm it is actually running in dual-channel before you credit the upgrade for the speedup.”
Frequently asked questions
Frequently asked questions
How much RAM do I need for Llama 3 70B?
At 4-bit quantization, budget roughly 40–45GB for weights alone, so 64GB total system RAM is the realistic floor once you account for the OS and context overhead.
Is DDR5 worth the upgrade over DDR4 for local LLMs?
For CPU-only inference, yes, the roughly 2x bandwidth increase translates fairly directly into tokens per second. If you are running entirely on a GPU with enough VRAM, system RAM speed matters much less.
Does more RAM speed up a model that already fits in VRAM?
No. If the entire model runs on the GPU, system RAM speed has little effect on inference speed. RAM upgrades matter for CPU-only inference or when a model partially spills out of VRAM.
Can I mix RAM speeds or brands?
You can, but the system will run at the speed of the slowest module, and mismatched kits are more likely to cause stability issues. Buy a matched kit for a reliable dual-channel upgrade.
Is quad-channel RAM worth it for local LLMs?
On boards that support it (typically workstation or server-class), yes, it roughly doubles bandwidth again over dual-channel. Confirm your specific board and CPU actually support quad-channel before buying four modules.