A dated card comparison with residency kept explicit
RTX 3060 Ti vs RTX 2060 for Ollama: we benchmarked each card alone
Recorded RTX 3060 Ti and RTX 2060 Ollama results, with fully resident runs separated from GPU-plus-CPU offload configurations and their interpretation limits.
Separate fully resident results from CPU-offload results.
qwen: both resident75.8 vs 57.7 tok/s at the recorded caps.
llama: unequal residency100% on GPU for the Ti; 86% for the 2060.
14B: both offload71% and 48% on GPU; CPU work is part of both results.
Your plan, at a glance.
THE LAB BRIEF
The short answer.
The fully resident qwen2.5:7b row recorded 75.8 tok/s on the 3060 Ti and 57.7 tok/s on the 2060. The llama and 14B rows involve unequal CPU offload, so their gaps are configuration results, not isolated GPU speed comparisons.
- qwen, Ti / 206075.8 / 57.7 tok/s
- qwen residency100% GPU on both
- llama residency100% / 86% GPU
Readers comparing these recorded configurations for the named models and runtime settings.
A GPU-only performance ranking derived from CPU-offload rows, a slot-width experiment, or a current used-price recommendation.
On the recorded qwen2.5:7b workload, both cards kept the model fully on GPU: 75.8 tok/s on the RTX 3060 Ti and 57.7 tok/s on the RTX 2060. That is the closest available per-card comparison in this record. The other named workloads involve CPU offload and answer a different question: how the complete configurations performed.
All measurements here are from the September bench notes, not a new test or current-price recommendation. The results do not isolate the effect of GPU cores, memory bus, slot width or any single specification.
Read the common method before the numbers
The card-vs-card session was measured on September 25, 2026, late evening. Each card ran alone in its own Ollama 0.34 instance, with nothing else loaded on either GPU. Models were Q4_K_M, flash attention was enabled and the KV cache used q8_0.
Each row used a warm-up followed by three 512-token generation runs, seed 1 and temperature 0. Generation speed is the mean of the runs. The 3060 Ti cap was 150 W and the 2060 cap was 125 W for the rows below. A 512-token output target does not document the context-window setting; that setting is not separately supplied with these rows.
Fully resident result: the closest per-card comparison
| Card | Software cap | Generation | Model on GPU |
|---|---|---|---|
| RTX 3060 Ti, qwen2.5:7b | 150 W | 75.8 tok/s | 100% |
| RTX 2060, qwen2.5:7b | 125 W | 57.7 tok/s | 100% |
The 2060 result is 76% of the Ti result after rounding. Both configurations kept this model resident, which removes the particular unequal-residency issue present below. The cards still differ in hardware and cap; this was not a controlled same-card slot or power-normalised comparison.
Ollama’s “on GPU” report describes model placement. It is not GPU utilisation, and it does not mean that the rest of the machine does no work. Keep that definition alongside every residency percentage.
Offload results: compare the complete configurations
| Model | RTX 3060 Ti, 150 W cap | RTX 2060, 125 W cap | Interpretation |
|---|---|---|---|
| llama3.1:8b | 73.6 tok/s; 100% GPU | 44.8 tok/s; 86% GPU | CPU participates in the 2060 configuration |
| qwen2.5:14b | 14.5 tok/s; 71% GPU | 8.6 tok/s; 48% GPU | CPU participates in both configurations |
For llama, the remaining 14% of the model was on CPU in the 2060 configuration. For qwen14B, the CPU shares were 29% on the Ti configuration and 52% on the 2060 configuration. These complements describe allocation, not CPU utilisation or a fraction of elapsed execution time.
Those rows are useful if the reader wants to know what these configurations achieved with those settings. They cannot establish that the Ti GPU alone is faster by the corresponding end-to-end ratio. The unequal placement changes the work assigned to the GPU and CPU, so a hardware-specification explanation alone is insufficient.
Keep power observations in the same test context
On the fully resident qwen workload, increasing the cap changed Ti generation from 75.8 tok/s at 150 W to 77.9 tok/s at 200 W, and the 2060 from 57.7 tok/s at 125 W to 59.1 tok/s at 170 W. The lower caps retained about 97% and 98% of those respective recorded speeds.
For llama on the 2060, the corresponding recorded speeds were 44.8 and 45.0 tok/s. For the solo qwen14B rows, speeds were unchanged at the two recorded caps. These observations do not support a universal rule that a cap can never affect an offloaded workload.
The same session recorded idle board power of 16.5 W on the Ti and 6.9 W on the 2060 with no model loaded. These are GPU driver readings. They cannot rank total system energy cost when CPU participation or other components differ.
Keep separate runs and buying decisions separate
The notes also report about 38 tok/s for qwen2.5:14b split across both cards earlier that week. That observation does not carry the same documented three-run protocol, so it is not a controlled baseline for the solo-card table. No causal speedup is calculated from it here.
For a new decision, choose the actual model, quantisation, context and concurrency, then verify residency and completed-work speed on the proposed configuration. A parameter-count label does not establish memory fit across every model or setting. Obtain current purchase quotes and condition/return terms separately; this comparison retains no old listing-price premium or value-per-dollar ranking.
What remains unmeasured includes a same-card slot swap, a controlled equal-residency rerun of the offloaded workloads, whole-system outlet power and the throughput of untested GPU variants. The tables support the named recorded configurations and no broader GPU-only conclusion.
For a current purchase, the budget GPU buying guide compares exact 12 GB, 16 GB and 24 GB cards with documented specifications and physical/power requirements. Its buying candidates are separate from the measured Ti and 2060 results above.
Common questions
Quick answers.
Is the 2060 76% as fast?
Its fully resident qwen2.5:7b result was about 76% of the Ti result at the recorded, different power caps. That ratio does not apply to every model or isolate slot width.
Why separate the llama row?
The Ti reported the model 100% on GPU; the 2060 reported 86%, with the remainder on CPU. The difference therefore includes residency and CPU participation.
Do the GPU percentages show utilisation?
No. They are Ollama’s reported share of the model on the GPU. They do not report GPU busy time or CPU utilisation.
Does the earlier split prove a multi-GPU speedup?
The earlier about-38 tok/s observation lacks the later matched three-run protocol. It is a separate observation, not a controlled split-versus-solo speedup.
Sources & methodology
Where these facts come from.
- September 25 card-vs-card measurements and method · checked October 19, 2026
Source dates are listed above.
Prepared by Racks & Watts.