← AI Hashrate 中文

Tesla P4 8GB vs T4 16GB for Local LLMs

The T4 16GB has about 67% more memory bandwidth (320 vs 192 GB/s). Batch-1 decode tracks bandwidth closely, so a model that fits both cards should decode roughly 67% faster on the T4 16GB. The T4 16GB fits more of the catalog at Q4/8K — 24 vs 16 of 69 models (16 vs 8 GB VRAM). Largest model that fits the T4 16GB but not the Tesla P4 8GB: gemma-4-26B-A4B-it (25.2B). MSRP is $80 for the Tesla P4 8GB vs $2,000 for the T4 16GB. MSRP is a launch list price, not live retail — Amazon prices move; click through for the current price. In the Q4/8K head-to-head below (39 shared models), the T4 16GB posts the higher tok/s in 38 rows vs 1. Methodology.

Spec comparison

Tesla P4 8GBT4 16GB
VendorNVIDIANVIDIA
VRAM8 GB16 GB
Memory typeGDDR5GDDR6
Memory bandwidth192 GB/s320 GB/s
TDP75 W70 W
MSRP (list)$80$2,000
BuySearch Amazon (affiliate)Search Amazon (affiliate)

MSRP is the launch list price, not live retail. Buy links are Amazon affiliate links — the price after you click is the live price.

Q4 / 8K head-to-head — top 15 of 39 shared models

ModelParams (B)Tesla P4 8GB tok/sTesla P4 8GB fitsT4 16GB tok/sT4 16GB fitsWinner
MiniCPM5-1B1.08113.1 est.Yes188.6 est.YesT4 16GB
Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF1630.024.1 est.No161.1 est.NoT4 16GB
Qwen3-30B-A3B-Instruct-250730.521.9 est.No146.5 est.NoT4 16GB
GLM-4.7-Flash31.221.0 est.No140.6 est.NoT4 16GB
Qwen3.5-35B-A3B34.718.2 est.No121.8 est.NoT4 16GB
qwen35b-a3b-fable-sft-abliterated34.717.3 est.No115.6 est.NoT4 16GB
Ornith-1.0-35B35.017.0 est.No113.9 est.NoT4 16GB
Qwen3.6-35B-A3B36.017.0 est.No113.5 est.NoT4 16GB
Qwen3.5-2B2.061.1 est.Yes101.8 est.YesT4 16GB
DeepSeek-R1-0528-Qwen3-8B-layer-mix-bpw-3.8-mlx2.353.1 est.Yes88.5 est.YesT4 16GB
Llama-3.2-3B-Instruct3.238.2 est.Yes63.6 est.YesT4 16GB
gpt-oss-20b21.533.6 est.No56.6 est.YesT4 16GB
gemma-4-26B-A4B-it25.223.9 est.No53.6 est.YesT4 16GB
Qwen3-4B-Instruct-25074.030.5 est.Yes50.9 est.YesT4 16GB
NVIDIA-Nemotron-3-Nano-4B-BF164.030.5 est.Yes50.9 est.YesT4 16GB

est. = bandwidth-based estimate; measured = curated real benchmark that overrides the estimate. Fits = weights + KV(8K) + 1 GB overhead ≤ 95% of VRAM.

Fit coverage at Q4 / 8K

Tesla P4 8GBT4 16GB
Catalog models that fit16 of 6924 of 69
Largest exclusive fitgemma-4-26B-A4B-it (25.2B)

Tesla P4 8GB details → · T4 16GB details →