I currently have a Ryzen 7 9700X, an RTX 5070 Ti, and a four-DIMM AM5 motherboard with a matched 2x16GB TeamGroup DDR5-6000 CL38 kit running at 6000 MT/s with EXPO at 1.35V. The sticks are installed in A2 and B2.
I'm considering adding a single 32GB T-Force Delta DDR5-6000 CL38 module, resulting in this layout:
A1: 32GB
A2: 16GB
B1: empty
B2: 16GB
The extra module is from a different kit and may use different memory ICs or ranks, even though its advertised speed and timings match. The upgrade would cost around $500 where I live, so replacing everything with a matched 2x32GB kit is difficult to justify.
I mainly need the additional capacity for running large language models locally, particularly Qwen 3.8 27B. My current heavily quantized model already uses nearly all of my available memory, and the newer model exhausted it completely.
How does a 2x16GB plus 1x32GB configuration behave on AM5 in practice? Would the system use dual-channel or flex mode for part of the memory, and how likely is DDR5-6000 EXPO to remain stable? If 6000 MT/s fails, would 5600 or 5200 be a realistic fallback? I'm interested in long-term experiences with three mismatched DDR5 modules, especially for workloads where having more total RAM may be more important than maximum memory bandwidth.
4 Answers
If you try it, keep the matching 16GB modules together in the recommended A2/B2 slots and place the 32GB module in the remaining slot on the other channel. Make sure the system is stable at default settings first, then experiment with 5600 or 5200 before attempting 6000 EXPO. A matched 2x32GB kit would be the cleaner and more predictable solution, but the three-DIMM setup can still be worth testing if capacity is the priority.
Even if the memory has to run at 4800 or 5200 MT/s, having 64GB is generally better than running out of RAM entirely. For large models, avoiding swapping or failing to load the model can matter more than the bandwidth reduction. However, expect lower performance in the unmatched region and test thoroughly with memory diagnostics and your actual workload before relying on it.
The system may operate in a flex or asymmetric-channel arrangement: the portion that can be paired across channels gets normal dual-channel bandwidth, while the unmatched portion of the 32GB module may run with single-channel bandwidth. The exact behavior depends on the motherboard and how it maps the modules, so I wouldn’t assume the entire 64GB will perform like a matched kit.
I’d be very cautious with this combination. AM5 memory controllers can already be sensitive with multiple DIMMs, and mixing a 32GB module with a 2x16GB kit makes the electrical load and memory layout less predictable. DDR5-6000 EXPO is unlikely to be guaranteed; you may need to fall back to 5200 or the motherboard’s JEDEC defaults, and stability could still depend on the specific CPU and modules.
That’s disappointing, but I mainly need the capacity for local models. Would the extra memory still be worthwhile if I have to run it at a lower speed?