← Back to Blog
RISC-V AI

DeepComputing DC-ROMA RISC-V Mainboard III Enters Production: SpacemiT K3, 60 TOPS, First RVA23 Laptop Board

DeepComputing DC-ROMA SpacemiT K3 RVA23 RVV 1.0 Framework Laptop 13 60 TOPS on-device LLM

Published: 2026-09-02 | Category: RISC-V AI | Reading time: ~5 min

DeepComputing has announced the first production batch of the DC-ROMA RISC-V Mainboard III, described as the first laptop mainboard supporting the RVA23 RISC-V profile and fully compatible with the Framework Laptop 13. The board is powered by the SpacemiT K3 and delivers up to 60 TOPS of general-purpose AI compute for on-device inference and AI-accelerated workloads.

Production is the operative word. Previous RISC-V laptop mainboards were developer runs; a first production batch is a different statement about supply readiness.

What makes RVA23 the headline

RVA23 is not a marketing label, and treating it as one is the mistake most coverage will make. The profile makes RVV 1.0 vector processing, hardware virtualisation (Hypervisor extension), and bit manipulation mandatory rather than optional.

That matters at the distribution level. Before RVA23, a Linux distribution building a single riscv64 package archive had to compile to the lowest common denominator across all shipping cores — no vector, no virtualisation in the default binary. RVA23 removes that constraint: distros can target a baseline that finally includes the features the silicon has had for years.

For a laptop mainboard this translates into three things: a working KVM-based virtualisation story, vector-accelerated multimedia and BLAS-class workloads without hand-tuned per-vendor code, and a mainstream distro image that boots rather than a board-specific fork.

SpacemiT K3: what has been published

DeepComputing's announcement states the SoC and the 60 TOPS figure and stops there. Detailed K3 specifications have been published by module partners rather than by SpacemiT in this news cycle. A SpacemiT K3 industrial module (Mini COM-E form factor) announced on 27 August 2026 lists the following, which should be attributed to the module vendor, not to DeepComputing or to the DC-ROMA board:

Separately, financial press reporting states that SpacemiT released the K3 in January 2026 as the first RVA23-compliant mass-produced RISC-V AI CPU, with 30× the AI compute of the K1 and 80× the supported model parameter scale, and that volume delivery began in April 2026. The K3 is also reported to include dual real-time CPU cores able to run a robot ROS stack independently of the application cores.

None of the above is confirmed against a SpacemiT datasheet here. Treat it as partner- and press-sourced until SpacemiT publishes authoritative documentation.

Why the Framework Laptop 13 compatibility is the real story

The Framework Laptop 13 compatibility is the most consequential line in the announcement, and it has nothing to do with the SoC.

It means the board drops into an existing, documented, repairable, third-party chassis with known thermals, a known battery, a known display panel, known expansion cards, and a large existing user base. Historically, RISC-V laptops have been self-contained units with bespoke enclosures, limited availability, and unclear spare-parts supply. Slotting into an established modular platform removes a whole class of adoption friction — for an engineer who wants to test a toolchain, measure RVV performance, or validate a vectorised library, this turns "wait for the next board run" into "order a mainboard."

Not disclosed by DeepComputing

The production announcement itself is short. Not disclosed in the source consulted: RAM configurations and maximum supported capacity; storage interface and included NVMe capacity; exact port and expansion layout; supported OS images and kernel version; thermal design power and sustained-performance behaviour under continuous inference load; price, order channels, and shipping date; batch size; and whether the 60 TOPS figure refers to INT8 or to the mixed-precision set quoted by the module vendor.

Engineering implications

The 60 TOPS figure needs precision context before you plan around it. If it tracks the module vendor's mixed-precision claim (BF16 / FP16 / FP8 / INT8 / INT4), then INT8 and FP8 throughput are different numbers, and achievable model throughput depends on which one the 60 TOPS refers to. Ask before sizing a design.

30B local models are a memory-bandwidth problem, not a TOPS problem. A 30B parameter model at FP8 needs roughly 30 GB of weights before KV cache. The module vendor lists 32 GB of LPDDR5, which is close to the practical floor. Whether the laptop board ships with enough memory to make that claim reachable is the question that matters.

RVA23 plus RVV 1.0 at 1024-bit is the differentiator against earlier RISC-V laptops. Prior riscv64 laptop hardware largely predates the profile, which meant per-board distro images and no portable vector code. If the K3's 1024-bit RVV 1.0 implementation is real and compiler-accessible, standard BLAS, OpenCV, and llama.cpp builds should reach their vector paths without board-specific patches.

What to verify before publishing

  1. Get the SpacemiT K3 datasheet or product brief to confirm core counts, clocks, TOPS precision, and memory interface.
  2. Confirm from DeepComputing directly whether the 60 TOPS figure is INT8 or mixed-precision.
  3. Confirm the board's RAM and storage configuration, price, and first-batch shipping date.
  4. Confirm which OS images ship and at which kernel version.

Sources

  1. AGI Hunt — "DC-ROMA RISC-V Mainboard III enters production with 60 TOPS AI compute" (2 September 2026; aggregator coverage of the DeepComputing announcement): https://agihunt.info/en/p/1a05757c8d007a1e9e09550ce6a
  2. CSDN — 众达科技 SpacemiT K3 industrial edge AI compute module, Mini COM-E (27 August 2026; module-vendor specifications, secondary source): https://www.csdn.net/article/2026-08-27/164110013
  3. DeepComputing DC-ROMA product page (for specification confirmation): https://deepcomputing.io/product/dc-roma-risc-v-mainboard
  4. 中国证券报 via gzylhyzx.com — "从芯片到场景 余杭全链打造智能经济新高地" (K3 January 2026 release, April 2026 volume delivery, 30× and 80× claims; secondary source): http://gzylhyzx.com/detail.html?cpid=T17&newsId=35228853
  5. AGI Hunt — SpacemiT K3 mass production and robotics adaptation context (secondary source): https://www.toutiao.com/a7680213954810888746

Verification notes