2026 GPU Spec List for Local AI, VR and Image Generation

2026年7月5日

How to choose a GPU and why VRAM matters so much are covered in separate articles.

This article lists the full specs and prices of every GPU you can buy as of April 2026. If you’re weighing your options, bookmark it and use it for comparison.

Full GPU spec list (as of April 2026)

NVIDIA GeForce RTX 50 series (Blackwell, released 2025–2026)

MSI GeForce RTX 5060 Ti 16GB VENTUS 2X OC PLUS

¥102,353 (as of 2026-06-22)

Amazon楽天市場Yahoo!

NVIDIA GeForce RTX 40 series (Ada Lovelace, released 2022–2024)

NVIDIA GeForce RTX 30 series (Ampere, released 2020–2022)

【中古】ELSA GeForce RTX 3060 12GB

¥52,800 (as of 2026-06-22)

Amazon楽天市場Yahoo!

【中古】MSI GeForce RTX 3090 GAMING X TRIO 24GB

¥148,000 (as of 2026/5/1)

Amazon楽天市場Yahoo!

AMD Radeon RX 9000 series (RDNA 4, released 2025)

AMD Radeon RX 7000 series (RDNA 3, released 2022–2024)

Quick recommendation table by VRAM

Notes on the spec tables

Why memory bandwidth matters

LLM generation speed is roughly proportional to memory bandwidth. Even at the same VRAM capacity, a GPU with wider bandwidth generates text faster. For example, between the RTX 3060 (12GB / 360 GB/s) and the RTX 3080 Ti (12GB / 912 GB/s), the same model shows a 2x+ difference in generation speed.

About the LLM score

The scores in this article are relative values with the RTX 4090 set to 100, based on text-generation speed with Q4_K_M quantized models. Actual speed varies with the model and context length. AMD GPU scores are reference values in a ROCm (Linux) environment.

Cautions when buying a used GPU

Used prices are based on shop stock (with a warranty). Flea-market apps can be cheaper, but there’s a risk of mining history and fan wear, so I recommend buying from a shop.

How generation speed is estimated

The “estimated tok/s" figures in this and related articles are calculated with the following formula.

What this formula means

Every time a GPU generates one token, it reads all of the model’s parameters from memory. The ratio between “the minimum time to read all of it" and “the time it actually takes to generate one token" is about 0.75. The remaining 0.25 goes to reading and writing the past conversation history (the KV cache) and to decompressing the compressed data (Q4 quantization).

Basis for 0.75: measured data

Below are results measured on an RTX 3090 and an RTX 3060 under conditions where the model fully fits in VRAM (April 15, 2026, Ollama 0.20.2, 256-token generation, Linux).

* Theoretical tok/s = memory bandwidth ÷ model size. The theoretical ceiling, assuming “one token can be generated in the time it takes to read all parameters once."

“Measured ÷ theoretical" falls in the 0.69–0.89 range, with a median of about 0.75. The narrower a GPU’s bandwidth (RTX 3060), the more memory-read waiting dominates, so results land close to the theoretical value. On a wide-bandwidth GPU (RTX 3090), the non-read work (reading/writing conversation history, decompression) stands out relatively more, so the ratio comes out a bit lower.

Limits of the estimate

This formula is only valid when the model fully fits in VRAM. If VRAM runs short and the model spills into CPU memory, speed drops dramatically (measured, it can fall to a tenth or less).

Also, when a model is split across two GPUs, you can combine the bandwidth of both, so it’s faster than a single card. In my setup, qwen3.5:27b (17GB) split across an RTX 3090 + RTX 3060 ran at 25.5 tok/s.

Quick estimated-speed table by GPU (8B and 14B models)

★ = author-measured (April 2026, Ollama 0.20.2, Linux). The rest are estimates from the formula (bandwidth ÷ model size × 0.75). Valid only when the model fully fits in VRAM. AMD GPUs are estimated in a ROCm environment.

Full GPU value ranking

How to read this chart: the longer the bar, the higher the “LLM performance per unit price" — i.e., the better the value. The value metric is “LLM score ÷ GPU price (¥10k units)." Marked [new]/[used].

* The LLM score is a relative value of text-generation speed with Ollama (Q4_K_M quantized), based on RTX 4090 = 100. AMD is a reference value under ROCm (Linux). Prices are as of April 2026. Used prices are the midpoint of the shop range.

The used RTX 3060 12GB and RTX 3080 12GB stand out on value. Among new cards, the RTX 5070 12GB and RTX 5060 Ti 16GB rank high. The pricey RTX 5090 and RTX 4090 are high-performance but rank low on the value metric. That said, choosing on value alone biases you toward “cheap GPUs with little VRAM," so the key is to secure the VRAM your use case needs first, then look at value.

Conclusion: if you’re torn, these three

MSI GeForce RTX 5060 Ti 16GB VENTUS 2X OC PLUS

¥102,353 (as of 2026-06-22)

Amazon楽天市場Yahoo!

【中古】ELSA GeForce RTX 3060 12GB

¥52,800 (as of 2026-06-22)

Amazon楽天市場Yahoo!

【中古】MSI GeForce RTX 3090 GAMING X TRIO 24GB

¥148,000 (as of 2026/5/1)

Amazon楽天市場Yahoo!

Summary

This article is a reference of the full GPU spec list and prices as of April 2026. Prices move, so check the latest street prices before buying.

For how to choose a GPU (recommendations by use case, VRAM guidelines, a budget-by-budget guide), see the related articles below.

The prices and specs in this article are as of April 2026. Prices change daily, so check the latest before buying.

Related

Running a Local AI Chatbot at Home: A Budget-by-Budget Guide

Starting Local AI with a Used GPU: The Value of the RTX 30/40 Series

2026 GPU prices — by VRAM

What matters for local AI is the size of the graphics card’s working space (VRAM — you can run as much model as fits here). I looked up the main cards on sale as of June 2026 on price-comparison sites (kakaku.com, etc.) and grouped them by VRAM size. Prices move by shop and timing, so treat them as ballpark figures (across 2026, prices have trended high overall due to a memory shortage).

For local AI, the two realistic choices in 2026 are: to start affordably new, 16GB (RTX 5060 Ti or the RX 9060/9070 line); to aim for 24GB, a used RTX 3090 (recently spiking). New 24GB has all but vanished from the current generation (even the RTX 5080 is 16GB), and beyond that you jump straight to the 32GB RTX 5090 (~¥700k). Decide first how large a model you want to run, then choose from the VRAM class that fits it and you won’t go wrong. Note that AMD (RX line) support varies by software, so it’s wise to confirm the tools you want to use actually run before buying.