Rockchip unveiled two RK182X LLM/VLM accelerators at its developer conference last July, namely the RK1820 with 2.5GB RAM for 3B parameter models, and the RK1828 with 5GB RAM for 7B parameter models.
A few months have passed, and we have a few more details since a development kit based on Rockchip RK1820/RK1828 SO-DIMM module is now available, some basic documentation has been released, and the RK1828 has been benchmarked against the Rockchip RK3588’s NPU. I was also told that M.2 modules are coming soon.
Firefly RK182X 3D RAM Stacking Development Kit
The “RK182X 3D RAM Stacking Development Kit” is based on the AIO-GS1N2 carrier board, previously offered with a Rockchip RK3576 or RK3588(s) and with a NVIDIA Jetson Orin Nano or NX AI module.
Specifications:
- Main module (now only RK3588 variant for sale as part of the RK182x kit)
- Core-3588JD4 – Rockchip RK3588 with 6 TOPS NPU, 4GB to 32GB LPDDR4/LPDDR4x, 32GB to 256GB eMMC flash, 8Kp60 video decoder, 8Kp30 video encoder
- Core-3588SJD4 AI – Rockchip RK3588S with 6 TOPS NPU, 4GB to 32GB LPDDR5, 32GB to 128GB eMMC flash, 8Kp60 video decoder, 8Kp30 video encoder
- Core-3576JD4 – Rockchip RK3576 with 6 TOPS NPU, 2GB to 16GB LPDDR4/LPDDR4x, 16GB to 256GB eMMC flash, 8Kp30 video decoder, 4Kp60 video encoder
- Sub module
- Rockchip RK1820 SO-DIMM module (Jetson Nano/NX compatible) with 2.5GB on-chip DRAM, 20TOPS (INT8) computing power, for large language models up to 3B parameters
- Rockchip RK1828 SO-DIMM module (Jetson Nano/NX compatible) with 5GB on-chip DRAM, 20TOPS (INT8) computing power, for large language models up to 7B parameters
- Storage
- MicroSD card slot
- 2x SATA 3.0 ports (two ports available only when the main module is Core-3588JD4; note: 3x SATA connector on carrier, but not all usable with the Rockchip modules)
- NVMe SSD via M.2 Key-M 2280 sockets
- Video Output
- HDMI 2.0 with 1080p resolution by default
- VGA port up to 1080p
- Audio – Audio input and output RCA connectors
- Networking
- 9x Gigabit Ethernet RJ45 ports with PSE functionality with 48V ATX power supply
- 1x RJ45 management port (not available with Core-3588SJD4 AI)
- Optional 2.4GHz/5GHz dual-band WiFi via M.2 E-Key socket (shared with SATA0, not populated by default)
- USB
- 2x USB 3.0 ports
- 1x USB 3.0 via 20-pin wafer
- Expansion
- Up to 4x M.2 Key-M sockets (only two sockets soldered, and only one usable with Rockchip modules)
- Footprint for M.2 Key-E socket
- 2x 10-pin connectors for RS485, UART, and GPIO
- Debugging – 3-pin Debug connector for serial console
- Misc
- Power, MaskROM, and Reset buttons
- DIP switch for auto power-on configuration
- USB mode toggle switch (Host/Device)
- 4-pin fan connector
- 4-pin system fan connector
- Power Supply
- 24V DC /5A via 5.5 × 2.1mm power barrel jack
- 6-pin 12V ATX connector power supply
- 4-pin 48V ATX connector; enables PoE power (PSE) to terminal devices.
- Dimensions – 231.27mm × 164.13mm × 33.57mm
- Weight – 350 grams
- Temperature Range – Operating: -20°C to 60°C; storage: -20°C to 70°C
- Humidity – Storage: 10%～90%RH (non-condensing)
The company provides a Debian 12 image with Linux 6.1 for the development kit with the Core-3588JD4 system-on-module. The Rockchip RK1820 and RK1828 chips are supported by the RKNN3 Toolkit, which differs from the earlier RKNN and RKNN2 toolkits. It exposes a C API and supports model transformation, inference, and performance evaluation. The wiki provides basic documentation, but does not explain how get the RKNN3 Toolkit, which should eventually be released on the AIRockchip Github account. I suppose the tools are preinstalled on the Debian 12 images. Besides the RK182X 3D RAM Stacking Development Kit, Firefly also mentions that the RK182X modules will be supported on the GS1-N2 AI Network Video Recorder Server.
Rockchip RK1820/RK1828 LLM, VLM, and AI vision benchmarks
The company also shared some LLM benchmarks with both the RK1820 and RK1828 models.
It looks to work pretty well on 3B and 7B LLM and VLM models with 59 to 180 tokens/s using Qwen2.5, Qwen3, fastbvlm, and internalVL3 models.
Radxa also shared additional benchmarks comparing results from a Rockchip RK3588 system alone, against one with a Rockchip RK1828 LLM/VLM accelerator module.
They notably used CNN models such as Yolov5s and ResNet50, and there’s no benefit of using the RK1828 over the 6 TOPS NPU found on the Rockchip RK3588 SoC. So the RK182X chips are not suitable for AI vision workloads. They also tested the same Qwen2.5 models with 0.5B, 3B, and 7B parameters as Firefly did, with similar results. It was not compared directly against the Rockchip RK3588, but we wrote about the release of the Rockchip RKLLM toolkitfor NPU-accelerated large language models on RK3588, RK3588S, and RK3576 SoCs last year, and a Rockchip RK3588 could only achieve about 14 tokens/s with the Qwen2 1.8B model (W8A8). Quantizations are different (W4A16 vs W8A8), so I’m not sure to what extent results can be compared, but the RK1828 should still offer a significant performance boost for LLMs and VLMs.
Tom Cubie also provided some extra tidbits of information in comments. Rockchip does not sell the chips directly, but instead offers RK182X SO-DIMM to business customers. They are also working on an M.2 RK182x module with an extra 12V power connector, and a heatsink is required for the SO-DIMM and M.2 modules. Finally, the next-generation LLM/VLM chip – the Rockchip RK1860 – should be released in Q2/Q3 2026.
The RK182X 3D RAM Stacking Development Kit is sold for $889 with the RK1820 module, and $1,029 with the RK1828 module, both fitted with a RK3588 SoM with 8GB RAM, 64GB flash. I could not find the RK1820/RK1828 SO-DIMM modules sold separately. For reference, the price difference between the RK1820 and RK1828 kits is $140, so the modules themselves should sell for a few hundred dollars, unless Firefly’s markup is pretty high, since they are the first company to offer these.
Thanks to MIcael for the tip
Jean-Luc started CNX Software in 2010 as a part-time endeavor, before quitting his job as a software engineering manager, and starting to write daily news, and reviews full time later in 2011.
Support CNX Software! Donate via cryptocurrencies, become a Patron on Patreon, or purchase goods on Amazon or Aliexpress. We also use affiliate links in articles to earn commissions if you make a purchase after clicking on those links.
