ERNIE 4.5 21B-A3B (文心) VRAM Requirements, KV Cache & Best GPUs
ERNIE 4.5 21B-A3B (文心) (21B total, 3B active, Sparse MoE + GQA) uses 11.96 GiB for 4-bit weights (Q4_K_M), requiring 12.96 GiB at 8K context and 14.27 GiB at 32K. · Hugging Face: baidu/ERNIE-4.5-21B-A3B-PT ↗ · GGUF: unsloth/ERNIE-4.5-21B-A3B-PT-GGUF ↗
Fits with Safe Headroom
Config-derived formula
HF config.json ↗
Required GPU Memory
12.96 GiB
13.92 GB (decimal)
Usable Hardware Memory
15.2 GiB
16 GB (physical)
Safe Headroom
+2.24 GiB
81% utilization
Est. Decode Throughput
~65.5 tok/s
3B active params
Est. Prefill Throughput
~3813 tok/s
4K Prompt TTFT: ~1074 ms
Est. Cost / 1M Tokens
$0.52
Sparse MoE + GQA · Apache-2.0
Recommended Hardware for Current Configuration
Best Single GPU
RTX 4060 Ti 16GB (16GB)
~$430
Best Value / Used GPU
RTX 3090 (24GB)
~$750
Apple Unified Memory
Ryzen AI Max+ 395 128GB
~$1,999
Cloud From
Vast.ai · $0.12 / hr
1x RTX 4060 Ti 16GB
Production Launch Commands
llama.cpp (GGUF)
llama-server -hf unsloth/ERNIE-4.5-21B-A3B-PT-GGUF:Q4_K_M \ --ctx-size 8192 -ngl 999 --flash-attn
vLLM
vllm serve baidu/ERNIE-4.5-21B-A3B-PT \ --max-model-len 8192 --gpu-memory-utilization 0.92
Live Token Speed Simulator (TTFT ~1074ms · ~65.5 tok/s)
[ERNIE 4.5 21B-A3B (文心) · 12.96 GiB] allocated on target GPU. Prefill completed in TTFT window; streaming autoregressive tokens at memory-bandwidth-bound rate...
Cloud GPU Comparison & Rent-vs-Buy TCO
Verified public rates across RunPod, Vast.ai, Vultr, and DigitalOcean.
Purchase break-even
~18.5 months
Est. local power: $3.51 / mo
| Provider | Recommended Instance | Total VRAM | Hourly Rate | Monthly at Selected Hours | Billing Mode | Deploy |
|---|---|---|---|---|---|---|
| RunPod | 1x RTX 3090 (Community Cloud) | 24 GB | $0.22 / hr | $26.75 / mo | Per-second | Deploy GPU ↗ |
| Vast.ai | 1x RTX 4060 Ti 16GB (Verified Host) | 16 GB | $0.12 / hr | $14.84 / mo | Spot / On-demand | Deploy GPU ↗ |
| Vultr ($300 Free Credit*) | 1x L40S 48GB (Dedicated Instance · $300 Credit*) | 48 GB | $1.67 / hr | $203.19 / mo | Hourly | Deploy GPU ↗ |
| Vultr Cloud GPU | 1x L40S 48GB (Dedicated Instance) | 48 GB | $1.67 / hr | $203.19 / mo | Hourly | Deploy GPU ↗ |
| DigitalOcean | 1x RTX 6000 Ada (Dedicated Instance) | 48 GB | $1.57 / hr | $190.91 / mo | Per-second | Deploy GPU ↗ |
Rates reflect public listings on the verification date and fluctuate with regional inventory.
Fine-Tuning VRAM Requirements (QLoRA / LoRA / Full) — ERNIE 4.5 21B-A3B (文心)
QLoRA (4b)
Fits with Safe Headroom
14.19 GiB
Model Weights: 11.96 GiB · Gradients + Optimizer: 1.2 GiB · Act: 1.03 GiB
Recommended GPU: RTX 4060 Ti 16GB (16GB)
LoRA (16b)
Exceeds Available GPU Memory
42.28 GiB
Model Weights: 39.12 GiB · Gradients + Optimizer: 1.8 GiB · Act: 1.36 GiB
Recommended GPU: A40 48GB (48GB)
Full BF16
Exceeds Available GPU Memory
162.79 GiB
Model Weights: 39.12 GiB · Gradients + Optimizer: 121.27 GiB · Act: 2.4 GiB
Recommended GPU: Instinct MI300X 192GB (192GB)
Engineering Sizing Notes & Optimization Playbook
Quick Hardware Verdict
RTX 4060 Ti 16GB (16GB physical, 15.2 GiB usable) runs 4-bit (Q4_K_M) ERNIE 4.5 21B-A3B (文心) at 8K context using 12.96 GiB with 2.24 GiB of safe headroom (~65.5 tok/s single-stream decode).
KV Cache & Runtime Tuning
ERNIE 4.5 21B-A3B (文心) uses Sparse MoE + GQA (21B total, 3B active per token). Moving from 8K to 32K context expands FP16 KV cache from 0.44 GiB to 1.75 GiB; enabling 8-bit KV cache (llama.cpp -ctk q8_0 -ctv q8_0 or vLLM --kv-cache-dtype fp8) saves 0.87 GiB at 32K (reducing total VRAM to 13.4 GiB).
VRAM Scaling Across Context Windows (4K to 128K, 4-bit Weights)
ERNIE 4.5 21B-A3B (文心) × 49 GPUs (4-bit Q4_K_M)
| Target Hardware | Total VRAM | Memory Bandwidth | 4-bit @ 8K | 4-bit @ 32K | Decode Speed | Compatibility | Full Compatibility Report |
|---|---|---|---|---|---|---|---|
| NVIDIA RTX 3060 12GB | 12 GB (11.4 GiB) | 360 GB/s | 12.96 GiB | 14.27 GiB | ~65.1 tok/s | Exceeds Available GPU Memory | Full Compatibility Report → |
| NVIDIA RTX 4060 Ti 16GB | 16 GB (15.2 GiB) | 288 GB/s | 12.96 GiB | 14.27 GiB | ~65.5 tok/s | Fits with Safe Headroom | Full Compatibility Report → |
| NVIDIA RTX 4070 Ti Super | 16 GB (15.2 GiB) | 672 GB/s | 12.96 GiB | 14.27 GiB | ~128.3 tok/s | Fits with Safe Headroom | Full Compatibility Report → |
| NVIDIA RTX 4080 Super | 16 GB (15.2 GiB) | 736 GB/s | 12.96 GiB | 14.27 GiB | ~136.8 tok/s | Fits with Safe Headroom | Full Compatibility Report → |
| NVIDIA RTX 5070 12GB | 12 GB (11.4 GiB) | 672 GB/s | 12.96 GiB | 14.27 GiB | ~92.1 tok/s | Exceeds Available GPU Memory | Full Compatibility Report → |
| NVIDIA RTX 5070 Ti | 16 GB (15.2 GiB) | 896 GB/s | 12.96 GiB | 14.27 GiB | ~156.3 tok/s | Fits with Safe Headroom | Full Compatibility Report → |
| NVIDIA RTX 5080 | 16 GB (15.2 GiB) | 960 GB/s | 12.96 GiB | 14.27 GiB | ~163.5 tok/s | Fits with Safe Headroom | Full Compatibility Report → |
| NVIDIA RTX 3090 | 24 GB (22.8 GiB) | 936 GB/s | 12.96 GiB | 14.27 GiB | ~160.8 tok/s | Fits with Safe Headroom | Full Compatibility Report → |
| NVIDIA RTX 4090 | 24 GB (22.8 GiB) | 1008 GB/s | 12.96 GiB | 14.27 GiB | ~168.6 tok/s | Fits with Safe Headroom | Full Compatibility Report → |
| NVIDIA RTX 4090 D | 24 GB (22.8 GiB) | 1008 GB/s | 12.96 GiB | 14.27 GiB | ~168.6 tok/s | Fits with Safe Headroom | Full Compatibility Report → |
| NVIDIA RTX 5090 | 32 GB (30.4 GiB) | 1792 GB/s | 12.96 GiB | 14.27 GiB | ~232.6 tok/s | Fits with Safe Headroom | Full Compatibility Report → |
| NVIDIA RTX 5090 D | 32 GB (30.4 GiB) | 1792 GB/s | 12.96 GiB | 14.27 GiB | ~232.6 tok/s | Fits with Safe Headroom | Full Compatibility Report → |
| NVIDIA RTX A6000 | 48 GB (45.6 GiB) | 768 GB/s | 12.96 GiB | 14.27 GiB | ~140.9 tok/s | Fits with Safe Headroom | Full Compatibility Report → |
| NVIDIA RTX 6000 Ada | 48 GB (45.6 GiB) | 960 GB/s | 12.96 GiB | 14.27 GiB | ~163.5 tok/s | Fits with Safe Headroom | Full Compatibility Report → |
| NVIDIA RTX PRO 6000 Blackwell | 96 GB (91.2 GiB) | 1792 GB/s | 12.96 GiB | 14.27 GiB | ~232.6 tok/s | Fits with Safe Headroom | Full Compatibility Report → |
| NVIDIA DGX Spark (GB10 128GB) | 128 GB (99.8 GiB) | 273 GB/s | 12.96 GiB | 14.27 GiB | ~67.8 tok/s | Fits with Safe Headroom | Full Compatibility Report → |
| NVIDIA L4 24GB | 24 GB (22.8 GiB) | 300 GB/s | 12.96 GiB | 14.27 GiB | ~67.9 tok/s | Fits with Safe Headroom | Full Compatibility Report → |
| NVIDIA A40 48GB | 48 GB (45.6 GiB) | 696 GB/s | 12.96 GiB | 14.27 GiB | ~131.5 tok/s | Fits with Safe Headroom | Full Compatibility Report → |
| NVIDIA L20 48GB | 48 GB (45.6 GiB) | 864 GB/s | 12.96 GiB | 14.27 GiB | ~152.6 tok/s | Fits with Safe Headroom | Full Compatibility Report → |
| NVIDIA L40S 48GB | 48 GB (45.6 GiB) | 864 GB/s | 12.96 GiB | 14.27 GiB | ~152.6 tok/s | Fits with Safe Headroom | Full Compatibility Report → |
| NVIDIA A800 80GB SXM | 80 GB (76 GiB) | 2039 GB/s | 12.96 GiB | 14.27 GiB | ~247.3 tok/s | Fits with Safe Headroom | Full Compatibility Report → |
| NVIDIA A100 80GB SXM | 80 GB (76 GiB) | 2039 GB/s | 12.96 GiB | 14.27 GiB | ~247.3 tok/s | Fits with Safe Headroom | Full Compatibility Report → |
| NVIDIA H800 80GB SXM | 80 GB (76 GiB) | 3350 GB/s | 12.96 GiB | 14.27 GiB | ~301 tok/s | Fits with Safe Headroom | Full Compatibility Report → |
| NVIDIA H100 80GB SXM | 80 GB (76 GiB) | 3350 GB/s | 12.96 GiB | 14.27 GiB | ~301 tok/s | Fits with Safe Headroom | Full Compatibility Report → |
| NVIDIA H20 96GB HBM3 | 96 GB (91.2 GiB) | 4000 GB/s | 12.96 GiB | 14.27 GiB | ~318.5 tok/s | Fits with Safe Headroom | Full Compatibility Report → |
| NVIDIA H20 141GB HBM3e | 141 GB (133.9 GiB) | 4800 GB/s | 12.96 GiB | 14.27 GiB | ~335.2 tok/s | Fits with Safe Headroom | Full Compatibility Report → |
| NVIDIA H200 141GB SXM | 141 GB (133.9 GiB) | 4800 GB/s | 12.96 GiB | 14.27 GiB | ~335.2 tok/s | Fits with Safe Headroom | Full Compatibility Report → |
| NVIDIA B200 180GB | 180 GB (171 GiB) | 8000 GB/s | 12.96 GiB | 14.27 GiB | ~374.5 tok/s | Fits with Safe Headroom | Full Compatibility Report → |
| NVIDIA B300 288GB Blackwell Ultra | 288 GB (273.6 GiB) | 8000 GB/s | 12.96 GiB | 14.27 GiB | ~374.5 tok/s | Fits with Safe Headroom | Full Compatibility Report → |
| Huawei Ascend 910B 64GB (昇腾) | 64 GB (60.8 GiB) | 1200 GB/s | 12.96 GiB | 14.27 GiB | ~187.5 tok/s | Fits with Safe Headroom | Full Compatibility Report → |
| Huawei Ascend 910C 128GB (昇腾) | 128 GB (121.6 GiB) | 2400 GB/s | 12.96 GiB | 14.27 GiB | ~265.5 tok/s | Fits with Safe Headroom | Full Compatibility Report → |
| Huawei Atlas 300I Duo 96GB (310P) | 96 GB (91.2 GiB) | 408 GB/s | 12.96 GiB | 14.27 GiB | ~87.6 tok/s | Fits with Safe Headroom | Full Compatibility Report → |
| Cambricon MLU590 80GB (寒武纪思元) | 80 GB (76 GiB) | 2048 GB/s | 12.96 GiB | 14.27 GiB | ~247.8 tok/s | Fits with Safe Headroom | Full Compatibility Report → |
| Hygon DCU K100-AI 64GB (海光深算) | 64 GB (60.8 GiB) | 1638 GB/s | 12.96 GiB | 14.27 GiB | ~222.4 tok/s | Fits with Safe Headroom | Full Compatibility Report → |
| MetaX MetaX C500 64GB (沐曦曦云) | 64 GB (60.8 GiB) | 1800 GB/s | 12.96 GiB | 14.27 GiB | ~233.1 tok/s | Fits with Safe Headroom | Full Compatibility Report → |
| Kunlunxin Kunlunxin P800 64GB (昆仑芯) | 64 GB (60.8 GiB) | 1600 GB/s | 12.96 GiB | 14.27 GiB | ~219.8 tok/s | Fits with Safe Headroom | Full Compatibility Report → |
| Moore Threads MTT S4000 48GB (摩尔线程) | 48 GB (45.6 GiB) | 768 GB/s | 12.96 GiB | 14.27 GiB | ~140.9 tok/s | Fits with Safe Headroom | Full Compatibility Report → |
| AMD Radeon RX 7900 XTX 24GB | 24 GB (22.8 GiB) | 960 GB/s | 12.96 GiB | 14.27 GiB | ~163.5 tok/s | Fits with Safe Headroom | Full Compatibility Report → |
| AMD Radeon AI PRO R9700 32GB | 32 GB (30.4 GiB) | 640 GB/s | 12.96 GiB | 14.27 GiB | ~123.8 tok/s | Fits with Safe Headroom | Full Compatibility Report → |
| AMD Ryzen AI Max+ 395 128GB | 128 GB (99.8 GiB) | 256 GB/s | 12.96 GiB | 14.27 GiB | ~64.1 tok/s | Fits with Safe Headroom | Full Compatibility Report → |
| AMD Instinct MI300X 192GB | 192 GB (182.4 GiB) | 5300 GB/s | 12.96 GiB | 14.27 GiB | ~343.7 tok/s | Fits with Safe Headroom | Full Compatibility Report → |
| AMD Instinct MI325X 256GB | 256 GB (243.2 GiB) | 6000 GB/s | 12.96 GiB | 14.27 GiB | ~353.8 tok/s | Fits with Safe Headroom | Full Compatibility Report → |
| AMD Instinct MI355X 288GB | 288 GB (273.6 GiB) | 8000 GB/s | 12.96 GiB | 14.27 GiB | ~374.5 tok/s | Fits with Safe Headroom | Full Compatibility Report → |
| Intel Arc Pro B60 24GB | 24 GB (22.8 GiB) | 456 GB/s | 12.96 GiB | 14.27 GiB | ~95.7 tok/s | Fits with Safe Headroom | Full Compatibility Report → |
| Intel Gaudi 3 128GB HBM2e | 128 GB (121.6 GiB) | 3700 GB/s | 12.96 GiB | 14.27 GiB | ~310.9 tok/s | Fits with Safe Headroom | Full Compatibility Report → |
| Apple Mac mini M4 Pro 64GB | 64 GB (49.9 GiB) | 273 GB/s | 12.96 GiB | 14.27 GiB | ~67.8 tok/s | Fits with Safe Headroom | Full Compatibility Report → |
| Apple MacBook Pro M4 Max 128GB | 128 GB (99.8 GiB) | 546 GB/s | 12.96 GiB | 14.27 GiB | ~117.9 tok/s | Fits with Safe Headroom | Full Compatibility Report → |
| Apple Mac Studio M4 Ultra 256GB | 256 GB (199.7 GiB) | 819 GB/s | 12.96 GiB | 14.27 GiB | ~156.6 tok/s | Fits with Safe Headroom | Full Compatibility Report → |
| Apple Mac Studio M3 Ultra 512GB | 512 GB (399.4 GiB) | 819 GB/s | 12.96 GiB | 14.27 GiB | ~156.6 tok/s | Fits with Safe Headroom | Full Compatibility Report → |