FitMyLLM
NVIDIA· Ampere

NVIDIA PG506-232

Running LLMs on the PG506-232 — the long read: which models fit at which quantisation, and the settings worth changing. · Or what a budget buys

VRAM
24 GB
High-end
Bandwidth
933
GB/s
Models Q4
312/498
63%
7B Q4 speed
~107
Blazing
Model coverage at Q463% of all
Estimated speed· By model size at Q4

Average speeds at Q4 quantization. Actual performance varies by model architecture and context length.

3B
~249
tok/s
7B
~107
tok/s
14B
~53
tok/s
32B
~23
tok/s
70B
—
39.4GB needed
Measured rig reports

We rent the machine and time every model on it: decode, VRAM peak, concurrency, watts and cost per million tokens. Yours may already be one of them — and two are free to read in full.

See the reports →
Specifications
VRAM
24 GB
Bandwidth
933 GB/s
FP16 Compute
10.3 TFLOPS
TDP
165W
Memory
HBM2
Architecture
Ampere
CUDA Cores
3,584
Tensor Cores
224
PCIe
Gen 4 x16
300
Fast models · >30 tok/s
Real-time chat speed
312
Usable · >10 tok/s
Comfortable for all tasks
312
Total compatible
Fit in VRAM at Q4
Don’t want to buy?

Test PG506-232 (or anything bigger) without committing. Pay by the second, cancel anytime.

Spin up in ~60s. Pay by the second. Cancel anytime.

Some links are affiliate links — we may earn a small commission at no extra cost to you. This helps keep FitMyLLM free and independent.

Compatible models· 312
S
Falcon-H1R Tiny 90M0.09B
Falcon·256K ctx· chat· reasoning
8293
tok/s · 2% VRAM
›
S
SmolLM2 135M0.135B
Smollm·2K ctx· chat
5529
tok/s · 2% VRAM
›
S
nomic-embed-text-v1.5 100M0.14B
Embedding·8K ctx· chat
5331
tok/s · 2% VRAM
›
S
GPT-2 124M0.14B
Gpt2·1K ctx· chat
5331
tok/s · 2% VRAM
›
S
SmolVLM 256M0.256B
Smollm·8K ctx· chat· vision
2916
tok/s · 3% VRAM
›
S
Gemma 3 270M0.27B
Gemma·32K ctx· chat
2764
tok/s · 3% VRAM
›
S
Snowflake Arctic Embed M v2.00.305B
Embedding·8K ctx· embedding· multilingual
2447
tok/s · 3% VRAM
›
S
bge-large-en-v1.5 335M0.335B
Embedding·1K ctx· chat
2228
tok/s · 3% VRAM
›
S
mxbai-embed-large-v10.335B
Embedding·1K ctx· embedding
2228
tok/s · 3% VRAM
›
S
Snowflake Arctic Embed L0.335B
Embedding·1K ctx· embedding
2228
tok/s · 3% VRAM
›
S
LFM2 350M0.35B
Lfm·125K ctx· chat· tool_use
2133
tok/s · 3% VRAM
›
S
SmolLM2 360M0.36B
Smollm·8K ctx· chat
2073
tok/s · 3% VRAM
›
S
GPT-2 Medium 345M0.38B
Gpt2·1K ctx· chat
1964
tok/s · 3% VRAM
›
S
Qwen 2.5 0.5B0.5B
Qwen·32K ctx· chat
1493
tok/s · 3% VRAM
›
S
SmolVLM 500M0.5B
Smollm·8K ctx· chat· vision
1493
tok/s · 3% VRAM
›
S
Falcon-H1 0.5B0.52B
Falcon·128K ctx· chat
1435
tok/s · 3% VRAM
›
S
BGE-M30.568B
Embedding·8K ctx· embedding
1314
tok/s · 3% VRAM
›
S
Snowflake Arctic Embed L v2.00.568B
Embedding·8K ctx· embedding· multilingual
1314
tok/s · 3% VRAM
›
S
Qwen3-Embedding 0.6B0.6B
Embedding·32K ctx· embedding· multilingual
1244
tok/s · 4% VRAM
›
S
Falcon-H1R Tiny 0.6B0.6B
Falcon·32K ctx· chat· reasoning
1244
tok/s · 4% VRAM
›
S
Falcon Perception 0.6B0.6B
Falcon·4K ctx· vision
1244
tok/s · 4% VRAM
›
S
Qwen 1.5 0.5B0.62B
Qwen·32K ctx· chat
1204
tok/s · 4% VRAM
›
S
LFM2 700M0.74B
Lfm·125K ctx· chat· tool_use
1009
tok/s · 4% VRAM
›
S
Qwen3 0.6B0.75B
Qwen·32K ctx· chat· reasoning
995
tok/s · 4% VRAM
›
S
GPT-2 Large 774M0.81B
Gpt2·1K ctx· chat
921
tok/s · 4% VRAM
›
S
Qwen 3.5 0.8B0.87B
Qwen·256K ctx· chat· coding· multilingual
858
tok/s · 4% VRAM
›
S
InternVL3 1B0.94B
Other·32K ctx· chat· vision
794
tok/s · 4% VRAM
›
S
MiniCPM5 1B1.08B
Minicpm·128K ctx· chat· reasoning· multilingual
691
tok/s · 5% VRAM
›
S
K2 Horizon 0.9B1.08B
Other·128K ctx· chat· coding· reasoning
691
tok/s · 5% VRAM
›
S
TinyLlama 1.1B1.1B
Llama·2K ctx· chat
679
tok/s · 5% VRAM
›
S
LFM2.5-1.2B-Thinking1.2B
Lfm·122K ctx· chat· reasoning· tool_use
622
tok/s · 5% VRAM
›
S
Llama-3.2-1B1.2B
Llama·4K ctx· chat
622
tok/s · 5% VRAM
›
S
LFM2 1.2B1.2B
Lfm·125K ctx· chat· tool_use· multilingual
622
tok/s · 5% VRAM
›
S
Zamba2 1.2B1.2B
Other·4K ctx· chat
622
tok/s · 5% VRAM
›
S
Ling 3.0 Tiny7.89BMoE
Other·128K ctx· chat· coding· reasoning
574
tok/s · 22% VRAM
›
S
EXAONE-4.0-1.2B1.3B
Exaone·64K ctx· chat
574
tok/s · 5% VRAM
›
S
OPT 1.3B1.3B
Opt·2K ctx· chat
574
tok/s · 5% VRAM
›
S
MiniCPM-V 4.61.3B
Other·256K ctx· chat· vision
574
tok/s · 5% VRAM
›
S
DeepSeek Coder 1.3B1.35B
Deepseek·16K ctx· coding
553
tok/s · 5% VRAM
›
S
Phi-1 1.3B1.42B
Phi·2K ctx· coding
526
tok/s · 6% VRAM
›
Next step

Get personalized recommendations.

See ranked models with benchmark scores, run commands, and precise speed estimates for your PG506-232.

What this card is worth

PG506-232 holds 312 of the models in our catalogue and is, in practice, a IQ4_XS card — the largest it takes is Nanbeige4.3 33B at IQ4_XS.

Tokens/sec per $100
—
8B at Q4_K_M, so cards compare like for like
VRAM per $100
—
what memory costs on this card
The biggest it takes
Nanbeige4.3 33B33.12B · IQ4_XS21.5 GB22 tok/sest
DeepSeek Coder 33B33B · IQ4_XS21.6 GB22 tok/sest
DeepSeek-R1-Distill-Qwen-32B32.8B · IQ4_XS21.5 GB22 tok/sest
Qwen3 32B32.8B · IQ4_XS21.5 GB22 tok/sest
Qwen2.5-32B32.5B · IQ4_XS21.3 GB22 tok/sest
Qwen 2.5 Coder 32B32.5B · IQ4_XS21.3 GB22 tok/sest
QwQ-32B32.5B · IQ4_XS21.3 GB22 tok/sest
Granite Switch 4.1 30B Preview32.24B · IQ4_XS21.1 GB23 tok/sest
Against cards of similar memory
M3 Max (36GB)27 GB$2,4991.5 tok/s per $100
M3 Pro (36GB)27 GB$1,9991.2 tok/s per $100
RTX 509032 GB$1,9997.9 tok/s per $100
RTX 5090 D32 GB$1,9997.9 tok/s per $100
PG506-23224 GB——

Neighbours in memory rather than in price: memory decides whether a card can do the job at all, so two cards of the same size at different prices is the comparison you are making. Ordered by memory, then bandwidth — not by the value column, which is worked out from bandwidth and price alone and therefore rewards a cheap card whatever its software stack does to that bandwidth in practice. Read it as one input, not as a ranking.

Device under test

NVIDIA PG506-232 — 24 GB VRAM.

PG506-232 spec
Brand
NVIDIA
VRAM
24 GB HBM2
Bandwidth
933 GB/s
FP16 Compute
10.3 TFLOPS
FP32 Compute
10.3 TFLOPS
CUDA Cores
3,584
Tensor Cores
224
TDP
165 W
Architecture
Ampere
AI capability
312/ 498 models @ Q4

With 24 GB VRAM and 933 GB/s bandwidth, this GPU handles models up to 30.5B parameters.

Speed ≈ bandwidth / model_size × efficiency. A 7B model at Q4 runs at ~107 tok/s.

Rent it — live prices
Checking what the clouds are charging…
Top models for PG506-232
312 fit · showing 20
ModelSizeVRAM Q4tok/sAvg
Qwen3 30B A3B30.5B19.1 GB24986.2
Qwen3-Coder 30B-A3B30.5B19.1 GB22636.9
Qwen3-30B-A3B Instruct 250730.5B19.1 GB22643.0
MPT-30B30B18.8 GB2526.8
OPT 30B30B18.8 GB256.3
Qwen3-Omni 30B-A3B30B18.8 GB24940.6
Granite 4.1 30B30B18.8 GB2522.6
Muse Glimmer 30B29.78B18.7 GB25233.5
Granite 4.2 30B29.28B18.4 GB2548.9
TranslateGemma 27B28.84B18.1 GB2638.6
Raon-1 mini (preview)28.78B18.1 GB17481.3
PaliGemma 2 28B28B17.6 GB2738.6
ERNIE 4.5 VL 28B A3B Thinking28B17.6 GB249—
Qwen3.5-27B27.8B17.5 GB2759.4
Qwen 3.8 27B27.78B17.5 GB27221.9
gemma-3-27b27.4B17.2 GB27121.7
gemma-2-27b27.2B17.1 GB2734.6
Qwen 3.6 27B27B17.0 GB2854.8
Gemma 4 26B A4B26B16.4 GB187174.3
Aria 25B A3.9B25.3B16.0 GB19164.8