Qwen3-30B-A3B (MoE) on 华为昇腾 Ascend 910B 64GB (HBM2e): VRAM Requirement & Speed Estimate
Yes. At Q4_K_M 4-bit quantization with a 16K context window, Qwen3-30B-A3B (MoE) requires 20.92 GB of VRAM (18.49 GB weights + 1.61 GB KV cache + 0.82 GB overhead), fitting within 昇腾 910B 64G's 61.4 GB usable memory at ~307.9 tokens/s.
Physical 64 GB · Load 34.1%
MoE: reads only 2 GB active weights/step
Enable Q8/Q4 KV quantization to scale batch concurrency
OLLAMA_CONTEXT_LENGTH=16384 ollama run qwen3:30b-a3b昇腾 910B 64G Fits Qwen3-30B-A3B (MoE) · Cost & TCO
Approx. $13,200 (~$206/GB VRAM) · 310W TDP
📊 All Quantization Levels for Qwen3-30B-A3B (MoE) on 昇腾 910B 64G
| Quantization | Total VRAM (16K) | Verdict | Est. Speed | Max Context |
|---|---|---|---|---|
| FP16 / BF16 | 64.28 GB | Partial RAM Offload (Slow) | ~76.9 tok/s | — |
| FP8 / Q8_0 | 35.11 GB | Smooth Fit | ~200.4 tok/s | 128K |
| GGUF Q6_K | 27.57 GB | Smooth Fit | ~246.1 tok/s | 128K |
| GGUF Q5_K_M | 24.19 GB | Smooth Fit | ~274.1 tok/s | 128K |
| Q4_K_M / AWQ | 20.92 GB | Smooth Fit | ~307.9 tok/s | 128K |
| GGUF IQ4_XS | 18.59 GB | Smooth Fit | ~337.7 tok/s | 128K |
| GGUF Q3_K_M | 17.27 GB | Smooth Fit | ~357.3 tok/s | 128K |
| GGUF Q2_K | 14.31 GB | Smooth Fit | ~410.5 tok/s | 128K |
💡 Recommended Alternative GPUs & Macs for Qwen3-30B-A3B (MoE)
Frequently Asked Questions
Can 华为昇腾 Ascend 910B 64GB (HBM2e) run Qwen3-30B-A3B (MoE) locally?
Yes. At Q4_K_M 4-bit quantization with a 16K context window, Qwen3-30B-A3B (MoE) requires 20.92 GB of VRAM (18.49 GB weights + 1.61 GB KV cache + 0.82 GB overhead), fitting within 昇腾 910B 64G's 61.4 GB usable memory at ~307.9 tokens/s.
What is the best quantization for Qwen3-30B-A3B (MoE) on 昇腾 910B 64G?
The highest-quality quantization that fits cleanly on 昇腾 910B 64G is FP8 / Q8_0 (99.9% Near-Lossless), using 35.11 GB VRAM at ~200.4 tokens/s with up to 128K context.
How much VRAM does Qwen3-30B-A3B (MoE) need in full FP16 precision?
In unquantized FP16/BF16 precision (16K context), Qwen3-30B-A3B (MoE) requires 64.28 GB VRAM (61 GB for weights alone). Because it uses a Mixture-of-Experts (MoE) architecture, all 30.5B parameters must sit in memory, while only 3.3B active parameters are read per token.