Graphics and memory
RAM and memory: buying guides for India
DDR5 and DDR4 desktop kits, high-capacity kits for AI work, laptop memory and ECC modules, compared on capacity, speed and platform support.
By capacity3 articles
By use3 articles
What should you know about RAM and memory for AI?
System memory (RAM) holds the model weights, data and working files that do not fit on the GPU. For AI work, capacity matters first, then the platform, speed and whether the kit is certified for the board. Larger models and CPU-only inference need more RAM than everyday use.
New here? Start with Best RAM for AI and Machine Learning in India.
Where should you learn the basics first?
Which other categories are related?
Common questions from buyers in India
How much RAM do you need for AI work in India?
It depends on whether the model runs on the GPU or the CPU. CPU-only inference needs room for the full model in RAM, while GPU work mostly needs RAM for data and the operating system. The guides explain each case.
DDR5 speed or capacity: which matters more for AI?
Capacity usually comes first, because a model that does not fit cannot run well at any speed. Speed helps once capacity is enough. See the explainer on speed versus capacity for the published reasoning.
RAM and memory buying guide for India
Short answer
Buy capacity first, then speed. A two-module kit of 64 GB suits most AI desktops. Start with best RAM for AI and machine learning, then see the 32 GB DDR5 kits, the 64 GB DDR5 kits and the 128 GB kits for local LLMs.
Features that matter
System RAM holds the model when it does not fit on the graphics card. It also feeds the CPU during inference, data preparation and everyday work. A model larger than memory will not run properly, so capacity comes first. Our guide to how much RAM a local LLM needs lists sizes by model. For the graphics card side, read how much VRAM an LLM needs.
| Feature | What to check | Why it matters |
|---|---|---|
| Capacity per kit | 32 GB, 64 GB, 96 GB or 128 GB | Corsair suggests 64 GB as the minimum for serious AI systems and 96 GB for future-proofing. |
| DDR5 type | Desktop U-DIMM for a desktop board | DDR5 is the current desktop standard. Laptops and servers use other module types. |
| Speed and latency | Rated speed in MT/s and CAS latency (CL) | Speed helps CPU inference. Latency matters less, because token generation waits on bandwidth. |
| Kit layout | Two modules, not four | AMD lists DDR5-5600 for two DIMMs and DDR5-3600 for four on the Ryzen 9 9950X. |
| Platform support | EXPO for AMD, XMP for Intel, board QVL | A rated speed is an overclock that depends on the board, processor and BIOS. |
| ECC | On-die ECC versus full ECC | DDR5 has on-die ECC built in. Some kits are listed as non-ECC modules. |
| Heat spreader | Module height in mm | Makers list heights from 32.7 mm to 42.23 mm, which can clash with large air coolers. |
Capacity and speed
Capacity decides whether a model loads at all. Speed decides how fast it generates once it loads. In Maxim Saplin's CPU-only test, Mistral 7B produced 9.66 tokens/s on DDR5-4800 and 11.34 tokens/s on DDR5-6000. That is a clear gain, but a smaller one than the jump in rated speed. DDR5 speed vs capacity for AI explains the trade-off in full.
Modules, channels and platform limits
A desktop has two memory channels. InsiderLLM measured a Llama 3.2 3B model at 7.36 tokens/s with one stick and 11.59 with two on an older DDR4 mini PC. The hardware is dated, so read it as a demonstration of the effect, not as a DDR5 figure.
Four modules often force a lower speed. In Saplin's test, four sticks totalling 96 GB could not hold the 6000 profile and ran stably only at 4800. Intel lists up to DDR5-6400 for the Core Ultra 9 285K, and both AMD and Intel list 256 GB as the maximum on these desktop chips.
Compatibility and heat
TeamGroup points buyers to the motherboard maker's QVL, the list of memory the board maker has tested, and advises against mixing kits. G.Skill gives the same advice on mixing. Check the QVL for 64 GB modules in particular. Look at the module height before you buy if you use a tall air cooler.
Common mistakes
- Paying for speed before capacity. A faster 32 GB kit cannot load a model that needs 43 GB.
- Filling all four slots. Four modules lower the speed the platform supports. Buy two larger modules and keep the other slots free.
- Choosing the wrong profile. Some kits list only XMP for Intel. Others list EXPO and XMP. Match the profile to your processor.
- Mixing kits. Makers advise against adding a second kit to an old one. Plan the final capacity at the start.
- Ignoring module height. A tall heat spreader can block the cooler. Check the maker's listed height.
Frequently asked questions
How much RAM do I need for AI work?
Take the model's file size and add room for the system. Corsair suggests 64 GB as a minimum for serious AI systems.
Does faster RAM speed up local LLMs?
It helps when the CPU does the work, because generation follows memory bandwidth. Gains are smaller than the rise in rated speed. Capacity still matters first.
Do I need ECC memory?
Most desktop buyers do not. DDR5 includes on-die ECC, and G.Skill lists its desktop kit as non-ECC. Full ECC needs support from the board and processor.
Should I buy two modules or four?
Two. AMD lists DDR5-5600 for two DIMMs and DDR5-3600 for four on the Ryzen 9 9950X. Two larger modules give the same capacity at a higher speed.