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 7.94 GB / 94.1 GB Usable

Qwen3-8B on NVIDIA H20 96GB (HBM3): VRAM Requirement & Speed Estimate

Yes. At Q4_K_M 4-bit quantization with a 16K context window, Qwen3-8B requires 7.94 GB of VRAM (4.97 GB weights + 2.42 GB KV cache + 0.55 GB overhead), fitting within H20 96G's 94.1 GB usable memory at ~420 tokens/s.

Concurrency (Batch):
Dense 8.2B·36L (d=4096)·32Q/8KV·Ctx 128K
Runnable Models →
GPUs
96 GB VRAM·4000 GB/s·Safe Usable 94.1 GB
Smooth Fit7.94 / 94.1 GB~420 tok/s
3. Weight Quantization
4. Per-Request Context16K (16,384 tok)
Verified: 2026-11-20·Roofline bandwidth/compute estimate (Batch=1)
Smooth Fit
7.94/ 94.1 GB usable

Physical 96 GB · Load 8.4%

Theoretical Decode Speed
~420tokens/s

Actual speed varies ~±25% by runtime

Max Context at Batch=1
128K/ 128K

Enable Q8/Q4 KV quantization to scale batch concurrency

Weights: 4.97 GBKV Cache (1×16K): 2.42 GBOverhead: 0.55 GB
🚀 Blazing Fast (Instant CoT)~1.2s for a 500-token response
Reference Launch CLI:
OLLAMA_CONTEXT_LENGTH=16384 ollama run qwen3:8b

H20 96G Fits Qwen3-8B · Cost & TCO

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

📊 All Quantization Levels for Qwen3-8B on H20 96G

Quantization Total VRAM (16K) Verdict Est. Speed Max Context
FP16 / BF16 19.59 GB Smooth Fit ~163.6 tok/s 128K
FP8 / Q8_0 11.75 GB Smooth Fit ~290.3 tok/s 128K
GGUF Q6_K 9.72 GB Smooth Fit ~363.1 tok/s 128K
GGUF Q5_K_M 8.81 GB Smooth Fit ~409.1 tok/s 128K
Q4_K_M / AWQ 7.94 GB Smooth Fit ~420 tok/s 128K
GGUF IQ4_XS 7.31 GB Smooth Fit ~420 tok/s 128K
GGUF Q3_K_M 6.95 GB Smooth Fit ~420 tok/s 128K
GGUF Q2_K 6.16 GB Smooth Fit ~420 tok/s 128K

💡 Recommended Alternative GPUs & Macs for Qwen3-8B

Frequently Asked Questions

Can NVIDIA H20 96GB (HBM3) run Qwen3-8B locally?

Yes. At Q4_K_M 4-bit quantization with a 16K context window, Qwen3-8B requires 7.94 GB of VRAM (4.97 GB weights + 2.42 GB KV cache + 0.55 GB overhead), fitting within H20 96G's 94.1 GB usable memory at ~420 tokens/s.

What is the best quantization for Qwen3-8B on H20 96G?

The highest-quality quantization that fits cleanly on H20 96G is FP16 / BF16 (Lossless Full Precision), using 19.59 GB VRAM at ~163.6 tokens/s with up to 128K context.

How much VRAM does Qwen3-8B need in full FP16 precision?

In unquantized FP16/BF16 precision (16K context), Qwen3-8B requires 19.59 GB VRAM (16.4 GB for weights alone).