ai.canshu.com
中文
VRAM Calculator Advisor & TCO API Pricing Compare Models Leaderboard GPUs

Search models, GPUs, and tools

↑↓ Navigate · Enter Open · ★ Watchlist prioritized ai.canshu.com
Q4_K_M: Smooth Fit 25.03 GB / 94.1 GB Usable

Qwen3-32B on NVIDIA H20 96GB (HBM3): VRAM Requirement & Speed Estimate

Yes. At Q4_K_M 4-bit quantization with a 16K context window, Qwen3-32B requires 25.03 GB of VRAM (19.88 GB weights + 4.29 GB KV cache + 0.85 GB overhead), fitting within H20 96G's 94.1 GB usable memory at ~130.7 tokens/s.

Concurrency (Batch):
Dense 32.8B·64L (d=5120)·64Q/8KV·Ctx 128K
Runnable Models →
GPUs
96 GB VRAM·4000 GB/s·Safe Usable 94.1 GB
Smooth Fit25.03 / 94.1 GB~130.7 tok/s
3. Weight Quantization
4. Per-Request Context16K (16,384 tok)
Verified: 2026-11-20·Roofline bandwidth/compute estimate (Batch=1)
Smooth Fit
25.03/ 94.1 GB usable

Physical 96 GB · Load 26.6%

Theoretical Decode Speed
~130.7tokens/s

Actual speed varies ~±25% by runtime

Max Context at Batch=1
128K/ 128K

Enable Q8/Q4 KV quantization to scale batch concurrency

Weights: 19.88 GBKV Cache (1×16K): 4.29 GBOverhead: 0.85 GB
🚀 Blazing Fast (Instant CoT)~3.8s for a 500-token response
Reference Launch CLI:
OLLAMA_CONTEXT_LENGTH=16384 ollama run qwen3:32b

H20 96G Fits Qwen3-32B · Cost & TCO

Approx. $14,000 (~$146/GB VRAM) · 400W TDP

📊 All Quantization Levels for Qwen3-32B on H20 96G

Quantization Total VRAM (16K) Verdict Est. Speed Max Context
FP16 / BF16 71.66 GB Smooth Fit ~42.5 tok/s 99K
FP8 / Q8_0 40.29 GB Smooth Fit ~77.8 tok/s 128K
GGUF Q6_K 32.18 GB Smooth Fit ~99.2 tok/s 128K
GGUF Q5_K_M 28.54 GB Smooth Fit ~113 tok/s 128K
Q4_K_M / AWQ 25.03 GB Smooth Fit ~130.7 tok/s 128K
GGUF IQ4_XS 22.52 GB Smooth Fit ~147.1 tok/s 128K
GGUF Q3_K_M 21.1 GB Smooth Fit ~158.4 tok/s 128K
GGUF Q2_K 17.92 GB Smooth Fit ~191.2 tok/s 128K

💡 Recommended Alternative GPUs & Macs for Qwen3-32B

Frequently Asked Questions

Can NVIDIA H20 96GB (HBM3) run Qwen3-32B locally?

Yes. At Q4_K_M 4-bit quantization with a 16K context window, Qwen3-32B requires 25.03 GB of VRAM (19.88 GB weights + 4.29 GB KV cache + 0.85 GB overhead), fitting within H20 96G's 94.1 GB usable memory at ~130.7 tokens/s.

What is the best quantization for Qwen3-32B on H20 96G?

The highest-quality quantization that fits cleanly on H20 96G is FP16 / BF16 (Lossless Full Precision), using 71.66 GB VRAM at ~42.5 tokens/s with up to 99K context.

How much VRAM does Qwen3-32B need in full FP16 precision?

In unquantized FP16/BF16 precision (16K context), Qwen3-32B requires 71.66 GB VRAM (65.6 GB for weights alone).