The Rockchip Developer Conference 2025 (RKDC!2025) is now taking place in Fuzhou, China, with some interesting announcements such as the Rockchip RK3668 10-core Arm Cortex-A730/A530 processor with a 16 TOPS NPU and the RK182X RISC-V co-processor with support for up to 7B parameters LLM (large Language Model)or VLM (Vision Language Model).
Rochchip RK3668 10-core Armv9 SoC
Let’s have a look at the Rockchip RK3668 SoC, which looks quite similar to the RK3688 SoC unveiled last year, but with some differences.
Preliminary Rockchip RK3668 specifications:
- CPU – 4x Cortex-A730 + 6x Cortex-A530 Armv9.3 cores delivering around 200K DMIPS; note: neither core has been announced by Arm yet
- GPU – Arm Magni GPU delivering up to 1-1.5 TFLOPS of performance
- AI accelerator – 16 TOPS RKNN-P3 NPU
- VPU – 8K 60 FPS video decoder
- ISP – AI-enhanced ISP supporting up to 8K @ 30 FPS
- Memory – LPDDR5/5x/6 up to 100 GB/s
- Storage – UFS 4.0
- Video Output – HDMI 2.1 up to 8K 60 FPS, MIPI DSI
- Peripherals interfaces – PCIe, UCIe
- Manufacturing Process- 5~6nm
The RK3688 will come with eight big cores and four SMALL cores, while the RK3668 is offered in a four big cores and six SMALL cores configuration. The RK3688 also offers a 32 TOPS AI accelerator, up to 200GB/s LPDDR6 memory bandwidth, a 16Kp30 video decoder, and an 8Kp60 video encoder.
There’s little public information about the RK3668, and I found it via Radxa on X, who plans to make a ROCK 6 SBC based on the new SoC.
Rockchip RK182X LLM/VLM accelerator
The second announcement I noticed, thanks to BG5SUN on X, is about the RK182X 3B/7B LLM/VLM co-processor.
It features a multi-core RISC-V CPU, 2.5GB or 5GB “ultra-high bandwidth” DRAM, and PCIe 2.0, USB 3.0, and Ethernet interfaces to connect to the host processor. The company indicates that INT4/FP4 7B parameter models can fit into 3.5GB of RAM. They are designed for the company’s Rockchip RK3576/RK3588 SoCs, already equipped with a 6 TOPS NPU, as well as other processors.
It will be supported by the RKNN3 Toolkit, and support PyTorch, ONNX, and TensorFlow frameworks, as well as HuggingFace GGUF (GPT-Generated Unified Format).
Rockchip also provided some performance numbers for popular distilled models like Qwen2.5 and DeepSeek R1.
We previously noted it was possible to run DeepSeek-R1-Distill-Qwen-1.5B on the RK3588 using its 6 TOPS NPU, and the performance to solve a simple math equation was 188.53 tokens/s for prefill and 14.93 tokens/s for the generate part. I’m not sure Qwen2.5-1.5B above is directly comparable, but it still gives an idea of the extra performance to be expected with the RK182X accelerators, with over 2000 tokens/s for prefill, and around 120 tokens/s for decode. So it’s about 8 to 10 times faster than when using the NPU (and memory) on the RK3588 SoC.
[Update July 21: The part numbers will be RK1820 for the 3B LLM chip with 2.5GB RAM, and RK1828 for the 7B LLM chip wth 5GB RAM.
]
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.
