Skip to content
Voltarum Local LLM Hardware Guide
ES

Rockchip · Edge AI · 2022

Rockchip RK3588 boards (NPU)

For advanced users

Cheap Arm boards whose small NPU runs converted models through RKLLM. Slow, but very frugal.

Memory
8 GB–32 GB
Bandwidth
Not published
Power
20 W
Largest popular model at 4-bit
Qwen3.5 9B

—

Researched from manufacturer data and public sources; not tested by us.· Last reviewed Oct 11, 2026

View on Amazon (opens in a new tab)
All retailersCompare with another device

What it can run

Estimate

Popular models at Q4_K_M with a 32K-token context. Memory needed from real model files; speed is the theoretical ceiling.

This device runs only models compiled by its vendor (2 of our models).

Change model size or context

Model weightsContext cache (KV)RuntimeUsable limit (colour = verdict)

ModelVerdictMemorySpeed ceiling
Qwen3 1.7BRuns
6.4 GB / 14 GB · with 16 GB		—
Qwen3.5 9BRuns
8.3 GB / 14 GB · with 16 GB		—
Gemma 4 12BNot supported——
gpt-oss-20bNot supported——
Qwen3.6 27BNot supported——
Qwen3.6 35B A3BNot supported——
Llama 3.3 70BNot supported——
gpt-oss-120bNot supported——
Qwen3.5 122B A10BNot supported——
Qwen3.5 397B A17BNot supported——

Theoretical maximum for generating text: memory bandwidth divided by the bytes read per token. Real runs usually reach about half to three quarters of it.

Where to buy

Prices are editorial references (launch price and a typical street price on the date shown), not live prices. In 2026 the memory shortage has pushed many prices well above launch.

Prices are shown only if updated in the last 24 hours and may change; the retailer’s price at checkout applies. We may earn a commission from these links at no cost to you. Affiliate disclosure

Overview

Editorial opinion

Best for: Tinkerers who want an always-on, few-watt board running small converted models.

Editorial opinionWould we buy it today?

A hobby platform. For useful local chat, any 12–16 GB GPU is far faster.

Pros and cons

Editorial opinion

Pros

  • Up to 32 GB on cheap boards
  • Very low power

Cons

  • Models must be converted with RKLLM first
  • 6 TOPS NPU: a few tokens per second on 7–9B models
  • llama.cpp only uses the CPU here

Good to know

RKLLM supports Qwen 2–3.5, Llama, Gemma, Phi, MiniCPM and others after conversion to its own format.

Specifications

Factual data
Manufacturer
Rockchip
Chip
RK3588
Architecture
Arm Cortex-A76/A55 + 6 TOPS NPU
Type
Edge board / accelerator
Form factor
Developer kit / board
Released
2022
Availability
On sale
Memory options
8 GB / 16 GB / 32 GB
Memory type
LPDDR4X / LPDDR5 (board dependent)
Memory bandwidth
Not published
Usable for models (default)
29 GB / 32 GB
INT8 AI performance
6 TOPS
Board / max power
20 W
Compute stack
RKLLM (NPU)
Models
Vendor-compiled models only
Several units for one model
No

Source: Manufacturer specifications

Software support

Editorial opinion

How well the popular local-AI tools run on it today (editorial assessment).

  • llama.cppPoor
  • OllamaPoor
  • LM StudioNot supported
  • vLLMNot supported
  • MLXNot supported

Measured speeds

Factual data

Measured results with their source. We never invent speeds.

No measured results yet. The table above shows the theoretical ceiling instead.

Alternatives

NVIDIA GPUs

NVIDIA GeForce RTX 5090

Memory
32 GB
Bandwidth
1.79 TB/s

Largest popular model at 4-bit: Qwen3.6 35B A3B

The fastest consumer card for local AI: 32 GB and 1.8 TB/s make 30B-class models fly.

from $4,329 (typical price)Recommended

AMD GPUs

AMD Radeon AI PRO R9700

Memory
32 GB
Bandwidth
645 GB/s

Largest popular model at 4-bit: Qwen3.6 35B A3B

32 GB for about a third of an RTX 5090: the cheapest new 32 GB card.

from $1,550 (typical price)Recommended

NVIDIA GPUs

NVIDIA RTX PRO 4000 Blackwell

Memory
24 GB
Bandwidth
672 GB/s

Largest popular model at 4-bit: Qwen3.6 27B

24 GB in a single-slot, 140 W card: the homelab-friendly way to get 24 GB new.

from $1,500 (typical price)Recommended