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.
Notice how there are 0 words about actual software support (not the now-ubiquitous AI blurb) or any commitment to not end up like their past SoCs which run some old ass forks of heavily-patched U-boot and Linux, with mainline *still* not complete and largely at the mercy of Collabora and community.
Can’t wait for mainline support in 5-10 years. All driven by volunteers of course.
Wait, are they even going to publish the TRM for this one?
Expect this to always be the case when it comes to Arm chips, regardless of who makes them. There’s no will to create an open sauce solution or even something mainline. I have a feeling that at least some of it comes down to proprietary drivers from the IP owners, but it can’t just be that.
Companies like NXP tend to play well and have their support ready very early. Marvell used to be reasonably good at this as well, by relying on contractors to do that job, we’ve even discovered the imminent release of new chips in patches.
But neither are that good when it comes to open source drivers, not having publicly available drivers/software. Hardly an alternative for hobbyists or startups and small companies. That’s why so many businesses start out with the RPi, which in my book isn’t a great choice for a retail product.
Not sure what you mean by “not having publicly available drivers/software” because I was precisely speaking about the fact that for these ones, the drivers end up in mainline early, even sometimes before the chip appears, so by definition they are public. Maybe you’re speaking about datasheets instead ? In this case that’s true. And for Marvell particularly, the docs are not very thick.
So you are saying that there’s full, 100% support for things like the GPU, video encoders/decoders, 3D acceleration etc. in mainline from these companies? Sorry, but I don’t believe that.
There are no such components in their SoCs, they generally only contain PCIe, SATA, Ethernet, SD controllers as well as possibly a few accelerators (e.g. crypto, DMA, XOR engines). I think that GPU support will always remain horribly late in the Arm world, sadly, regardless of the chip vendor.
GPU for rk3588 has been mainline since 6.10 user space mesa took longer I run my boards off mainline.
Obviously as its even using cores and gpu that haven’t even got official Arm release.
If you want support then it needs to be released so firmware software support can be provided such as the rk3588.
https://gitlab.collabora.com/hardware-enablement/rockchip-3588/notes-for-rockchip-3588/-/blob/main/mainline-status.md
Hopefully others will do similar to the fantastic job collaboro have done, if others are just going to hack android BSP offerings then that is what it will be…
Really interesting SoC as at that point there is a gap in the market to heavyweights like Apple silicon.
Wow, that’s a massive jump to modern Arm cores all of a sudden, skipping several generations in one go. Wasn’t their highest end core to date the A76 from 2018 and all of a sudden they’re using 2025 cores. What happened?
3588 was first announced before the COVID 19.. and should first be available in 2020, the covid shifted this SoC of few years in time, and this is just the following generation after all. The 3688 specs are not the same than in previous show, but far better after the picture: 12 cores (8x Cortex-A730 + 4x Cortex-A530 Armv9.3 cores, sad to not see it in the text description, at least for disabled people. There is also 16K30FPS video output on the picture :). I wonder if there is still HDMI in, that was a very great feature of the RK3588??? At least 8K@60fps encoding is available. Nice to see that Rockchip continue to progress with RISC-V cores, beside the ARM/RISC-V mix of their video dedicated SoCs. Hope there will be a RISC-V main CPU based SoC in next generation.
So about 3 years after the release of the RK3588, we are still waiting for full function Linux based OS support, with no hope of a date for completion. It is essentially left for 3rd parties to hope to complete the job. I don’t think we have even seen a half decent Android build beyond 12 yet from anybody have we?
So perhaps 2029 for the RK3668? Never again for this company and it’s sub standard support of it’s own SOC.
They leave a very bad taste in my mouth, as does one of their partners, Orange Pi.
Big job was done in all domains. Vulkan 1.4+Zink is supported in Mesa git (and will be available on mesa 25.2, maybe available in few weeks or at most next month). Still not the overall perfs of vendor driver for desktop, but good ones for 3D applications, still due to ARM that don’t help so much, probably. you can boot on mainline, have display, sound, SATA (that didn’t work very well with vendor kernel, not up at each boot, but works perfectly with mainline on Radxa Rock 5 ITX), HDMI in/out, etc. Lot of new features was introduced in RK3588, inclusion process in Linux is slow due to strict rules, that help it to be kept clean was so much drivers.
Collabora gitlab has a table of supported features, and a hardware-enablement kernel patched git branch, for think in WIP/still not integrated in mainline Linux here:
https://gitlab.collabora.com/hardware-enablement/rockchip-3588/notes-for-rockchip-3588/-/blob/main/mainline-status.md
And FWIW my Rock5-ITX has been running mainline since 6.12 plus a couple backported patches from 6.13. So now it should be mostly fine I guess.
People forget RK3588 was announced back in 2020. That would be 5 years.
They announced RK3688 in 2025 with most likely release somewhere in 2026. I’d say it’s plenty of time for them to start upstreaming the SoC and show customers that their products are more than some Android TV Box toys.
Some mainline support on release day would be a huge selling point. Again, nobody expects it to be complete on day one – realistically that’s quite impossible as there are certain factors beyond their control (e.g. patch reviews) – but 1.5 years after release is the maximum I would consider reasonable.
> They leave a very bad taste in my mouth, as does one of their partners, Orange Pi.
It seems they are no longer partners after the OPi 5 ended up in Russian drones. OPi 6 will use the CIX P1 SoC.
> People forget RK3588 was announced back in 2020.
April 2019 to be precise.
just before COVID, so locking people at home didn’t help to finish hardware products in companies
> They announced RK3688 in 2025
They first announced it last year (October 2024), but the RK3668 is new to me.
That is armv9.4 so most def we will have no mainline support in sight have that in mind.. now what would be the expected price? i guess no less than 300USD for a board coming from radxa..
The ROCK 5B developer edition was launched for $129 and up (and even $79 for early backers). https://www.cnx-software.com/2022/01/09/rock5-model-b-rk3588-single-board-computer/
So I’m hoping for the Radxa ROCK 6 to start at $150 to $200.
Also, it cannot be too expensive because it only has 4 big cores for 6 small ones (the usual rockchip approach to core-stuffing) while Orion O6 has 8 big ones for 4 small, and already starts within that price range.
This is more like four medium cores and six small, since the X-cores are the big ones now and neither chip appears to have any of those.
RK3688 has 8+4 so aswell. Radxa already sell the O6 it for 250-350 range of price.. thats what i was saying, for them wouldnt make much sense not sure for others like orange pi
Perhaps with a bit of luck this will release early 2030.
3588 was announced 2018 and for the most part drivers have hit mainline in 2025. So 3688 should have for the most part mainline drivers in … 2024 + 7 = 2031. And the A730 will be comparable to the market then as the A76 is comparable to the market now. Yawn. Not terrible, but incapable of commanding a price for its compute… more for its balance of fixed-function hardware that makes it great for set-top boxes. Again.
You might want to take into account that (1) there was a pandemic in between, which was handled quite strongly in China (naliing doors shut and all), (2) it all depends how many new IPs will be in it – the rk3588 had a lot of them – but hopefully the rk36xx stuff will reuse a number of those, (3) Company policies might change over time, so while there was a definitive down from rk3399 to rk3588 in “engangement”, we’ll only see how this plays out once stuff is available.
But I really don’t subscribe to all those doom and gloom scenarios 😉
I think, after a hardware release in late 2024, support for the O6 just hit mainstream (kernel 6.17 and mesa 25.2 more or less).
How much is the Imortalis G730 different from the G720? Support might not need much more than adding a few new ID’s.
That mainstream support consists of a super basic device tree with only serial console working….If that took 7 months to merge, you can imagine what the actual driver support will take.
Also, the Mali GPU is just a 3D accelerator. It does not have any kind of display output on it – that is handled by separate blocks within the chip which need to be upstreamed individually by CIX.
We had mainline GPU support on RK3588 before any kind of display output.
> But I really don’t subscribe to all those doom and gloom scenarios
You think that’s unrealistic after soo many years of crappy Arm SoCs with poor vendor support?
> Company policies might change over time, so while there was a definitive down from rk3399 to rk3588 in “engangement”, we’ll only see how this plays out once stuff is available.
RK3399’s good support is mostly thanks to Google’s efforts, since they really wanted to use that particular SoC in Chromebooks.
For RK3588 we have Collabora, although they clearly don’t have as much money and interest to throw at the problem like Google did, which translates into the step down in engangement that you are seeing.
Must be fun putting out dozens of SoC variants, doing only the bare minimum of support via an outdated BSP kernel, then letting 3rd-parties handle the upstreaming and technical support work for basically free – all while boosting your sales.
RK182X is the best announcement of this conference. I hope it stays in the <30 range lol
why dont they make riscv socs ???
Maybe because RISC-V isn’t quite there yet and it costs a ton of money to design and tape out a chip?
Because to design an SOC around RISC-V you need roll your own extremely slow uArch on a budget, or license a very slow IP stack from the Foundation, or use something a bit slower from Alibaba or one of the American universities such as was done by TensTorrent with their TT-Ascalon. Even Keller’s product is quite slow compared to A76 let alone newer ARM IP. But it scales well horizontally.
There’s all kinds of media marketing promises of “fast” RISC-V IP, but there’s not enough vendor cooperation on uArch for there to be anything comparable to ARM, x86, or Power in the way of both dmips/clock or frequency scaling, and it’s those two things together that make an actual fast core. If you are happy with slow cores, stay with RISC-V for the next decade.
Maybe vendors will actually start cooperating on uArch instead of just ISA after another decade. Hopefully that comes with better free AI-driven EDA software. That could level the playing field, fingers crossed, but not holding my breath.
I agree with you on these points. RISC-V is successful in the ESP32 range of devices where it allows to lower license costs which are significant in such low-priced devices. But I strongly doubt we’ll see it anywhere performance or efficiency matters in the next decade given that it has made basically zero progress on these points over the last decade. It’s just found its place to replace proprietary cores in entry-level devices.
Any chance it will support SODIMM DDR5?
For a $250 motherboard with an additional $250 of memory soldered on it, if you fry the board, you loose everything.
Yes and in addition sometimes you buy the small RAM model first to evaluate the capabilities of the board, then you regret not having taken a larger one and you cannot upgrade it.
Forget about SO-DIMMs, we’re talking about LPDDR only with these kinds of SoCs.
But maybe Rockchip is considering LPCAMM2 support (most probably requires special timings so different boot BLOBs required) if LPCAMM2 is a thing in a few years when these new Rockchip SoCs are going to be found on SBC or ‘PC-style’ devices.
What do you meant a thing in a few years? You can already buy them from Crucial. Not cheap though.
https://www.crucial.com/memory/ddr5/CT32G75C2LP5X
Kind of a misunderstanding. I first wrote ‘still a thing’, then edited it (badly).
Since currently it is _not_ a thing given the low amount of devices supporting LPCAMM2 (is there even more than this single Thinkpad announced in April last year?) and even Framework not having adopted it.
W/o wide adoption prices won’t go down and with these high prices no adoption will likely happen (chicken/egg) especially in ‘as cheap as possible’ market segments.
Adoption takes time, but Dell is using it in some of their products, but for now it’s just high-end workstation type laptops.
I also thought shortly Dell would be using LPCAMM2 (since Google is telling this but by looking closer it’s wrongly labeled CAMM instead).
And the Framework story hints at an important detail: socketed RAM is (way) slower RAM
This is great. Whenever it is released, it will be the perfect board for emulation. We are talking about the possibility to emulate many PS3 games with this one 🙂
Will the RK3668 beat the RK3588 for overall processing speed? Will it take as long to unveil it as for the 3588?
Using the Dhrystone numbers provided by Rockchip:
RK3588 – 93K DMIPS (But I’ve seen some claim 130K DMIPS)
RK3668 – 200K DMIPS
RK3688 – 300K+ DMIPS
The usefulness of the Dhrystone benchmark is disputed, but that would mean RK3668 is twice as fast and RK3688 three times as fast as the RK3588 for multi-threaded integer tasks. Other benefits will be the higher memory bandwidth (important for LLM and other AI workloads), faster UFS 4.0 storage, etc…
> The usefulness of the Dhrystone benchmark
Nope, four decades after its invention the usefulness is not ‘disputed’ but doesn’t exist any more. Dhrystone was an improvement last century but the scores today are simply BS: https://github.com/ThomasKaiser/sbc-bench/blob/master/Benchmarking_some_benchmarks.md
The only reason marketing departments still spit out these garbage numbers is journalists/bloggers/youtoubers citing this nonsense.
Yes it should beat it. For example the A720 cores in the Orion O6 run at approximately 2/3 of the speed of skylake cores at the same frequency, and about twice the speed of an A76 at the same frequency. And we can expect a small improvement in A730.
GGUF (GPT-Generated Unified Format) 🙂
that’s funny
GGUF is the native format of llama.cpp
The GG are the initial of the creator of the project – https://github.com/ggerganov
I’m not gonna lie, and I got this from Grok, but GGUF = GPT-Generated United Format is all over the web, including on Hugging Face website itself, IBM, and others.
It’s the first time I’m seeing that one as well. Before GGUF it was called GGML (same as the lib that’s still in llama.cpp) with the Georgi’s initials. There were some discussions when everyone brought new stuff to the format to change it to various names to credit participants, and finally it was renamed to GGUF without seeing what “UF” standed for. Maybe it’s a retro-acronym now, though, and really stands for this nowadays.
The official format spec https://github.com/ggml-org/ggml/blob/master/docs/gguf.md does not define the acronym
https://huggingface.co/docs/hub/gguf does not spell it out either.
And searching in the whole docs returns nothing for GPT-Generated
if you search for GGUF(Georgi Gerganov Unified Format), you will also find plenty of matches, even on https://arxiv.org/abs/2505.01658
I guess, the joke is that GGUF backronym is indeed gpt-generated
That’s the link I was referring to: https://huggingface.co/collections/prithivMLmods/gpt-generated-unified-format-gguf-6730aa906d6fb0789d6bd13e
It looks like it might be user-generated content…
[ what Cadence, California, tells about LPDDR6
“Cadence Introduces Industry-First LPDDR6/5X 14.4Gbps Memory IP to Power Next-Generation AI Infrastructure”
“The complete PHY and controller memory system boasts a new high-performance, scalable and adaptable architecture based on Cadence’s proven and highly successful DDR5 12.8Gbps, LPDDR5X 10.7Gbps and GDDR7-36G product lines. This first offering in Cadence’s new LPDDR6 IP product line supports the LPDDR6 and LPDDR5X standards, including LPDDR5X CAMM2.”
LPDDR6 “8.800 bis 17.600 MT/s”, 1.05V, performance 1.1V, low power mode 0.9V
DDR6 “DDR6 is set to double the speed of DDR5″, at about 1.1V (DDR4 1.2V, 1.6-3.2GT/s, DDR5 ~4-8.8GT/s, DDR6 ~12.8-17.6-21GT/s “seems possible” (thx) ]
There is no more Cortex A730 anymore.. the naming changed to C1
Well spotted. Since it’s for IoT, would the RK3668 be an Arm Orbis C1-Pro (Cortex-A730)/C1-Nano (Cortex-A530) SoC?
Anyway Rockchip is used to seeing the CPU they use being renamed after announcements. Remember RK3288 with Cortex A12->A17 ?
yes most likely.
the GPU seems to be “G1” aswell (and its already supported in panthor!)