SpacemiT K3: Flagship RISC-V AI CPU Deep Dive
At the 5th Dishui Lake China RISC-V Industry Forum, SpacemiT detailed the K3 — the first mass-production RISC-V chip to claim RVA23 compliance, 1024-bit RVV, native FP8 inference, and full chip-level virtualization. We break down the architecture, benchmarks, and what it means for the RISC-V ecosystem.
Overview
On August 13, 2026, at the 5th Dishui Lake China RISC-V Industry Forum in Shanghai, SpacemiT market director Peng Huacheng presented the company's flagship Key Stone K3 SoC. Positioned as a "CPU 2.0" part that fuses general-purpose compute and AI acceleration on one die, K3 is SpacemiT's answer to the question of whether RISC-V can compete in high-performance computing and edge AI.
The presentation emphasized four claimed world-firsts for a mass-production RISC-V chip: RVA23 Profile compliance, 1024-bit RVV vector width, native FP8 inference, and complete chip-level virtualization. SpacemiT also disclosed real-world LLM benchmarks and a humanoid-robot half-marathon demonstration.
K3 Key Specifications
| Parameter | Value |
|---|---|
| Product Positioning | Flagship RISC-V AI CPU / "CPU 2.0" |
| CPU Cores | 8x SpacemiT X100 @ up to 2.4 GHz |
| Real-Time Cores | 2x RISC-V real-time cores |
| AI Cores | 8x SpacemiT A100 |
| General Compute | ~130K DMIPS |
| AI Compute | 60 TOPS |
| Memory | Up to 32 GB LPDDR5 |
| Real-Time Cache | 3 MB real-time compute cache |
| CAN-FD | 10x CAN-FD interfaces |
| CPU ISA | RVA23 Profile (64-bit application processor) |
| Vector Extension | RVV 1.0, 1024-bit VLEN |
| Precision Support | Native FP8 inference |
| Virtualization | Full chip-level virtualization (RVH) |
| Single-Core Perf | ~9 SPECint2006/GHz (X100, comparable to ARM Cortex-A76) |
| Software | Ubuntu, Bianbu Linux, OpenKylin, openEuler, ROS2 |
Four Claimed World-Firsts
SpacemiT highlighted four features that it says are firsts for a mass-production RISC-V chip:
- First RVA23-compliant mass-production chip: RVA23 defines a stable 64-bit application-processor baseline, so operating systems and distributions compiled for RVA23 can run on K3 without per-chip patches.
- First 1024-bit RVV implementation: The X100 cores support a 1024-bit vector length, widening the datapath for AI, signal processing, and HPC workloads.
- First native FP8 inference: FP8 is increasingly used for LLM inference because it preserves accuracy while halving memory bandwidth versus FP16. Native hardware support avoids software emulation.
- First full chip-level virtualization: K3 implements the RISC-V hypervisor extension (RVH) with IOMMU and AIA, enabling Type-1 hypervisors and secure multi-tenant workloads.
X100 and A100: The Compute Blocks
K3 is built from SpacemiT's self-designed X100 application cores and A100 AI cores. SpacemiT's roadmap shows a clear performance escalation:
| Core | Microarchitecture | SPECint2k6/GHz |
|---|---|---|
| X60 | Dual-issue, 8-stage pipeline | ~4 |
| X100 | Quad-issue, 12-stage out-of-order | ~9 |
| X200 | Six-issue, 10-stage | ~16 |
| X300 (2027 target) | Next-gen high-performance | ~20 |
The A100 AI cores scale from the edge-class A60 (2 TOPS) through the desktop-class A100 (32 TOPS per core class) to the server-class A200 (256 TOPS). On K3, eight A100 cores deliver a combined 60 TOPS of general AI compute.
On-Device LLM Benchmarks
SpacemiT showed that K3 can run most models on HuggingFace that do not rely on FP4/FP6 formats. One concrete benchmark disclosed at the forum:
| Model | First-Token Latency | Throughput |
|---|---|---|
| Qwen 30B-A3B | 0.9 s | ~15 tokens/s |
The company also stated that K3 can run 30B-parameter class models locally, putting it in the same conversation as entry-level AI PC and edge-NPU platforms, but on a fully open RISC-V ISA.
Real-Time Subsystem for Robotics
K3 includes a dedicated real-time computing domain for embodied AI and robotics:
- 2x RISC-V real-time cores
- 3 MB real-time compute cache
- 10x CAN-FD interfaces for motor drives and sensors
This "motion-control + AI" architecture is intended to replace the traditional CPU + MCU + NPU three-chip partition, reducing inter-chip latency and improving system stability. SpacemiT demonstrated the concept on April 19, 2026, when multiple "Linglong 2.0" humanoid robots powered by K3 completed the Beijing Yizhuang Humanoid Robot Half Marathon (21.0975 km). The chips had only been back from tape-out for about two months at that point.
Software and Open-Source Ecosystem
SpacemiT said that K3 source code was pushed to the mainline Linux kernel at launch, making it the first mass-production chip to fill the RVA23 kernel gap. The company lists active upstream work across GCC, LLVM, Ubuntu, OpenKylin, and openEuler.
Available software support includes:
- Bianbu Linux — SpacemiT's Ubuntu/Debian-based distribution
- Ubuntu 26.04 LTS — upstream distribution support
- ROS2 — for robotics development
- OpenHarmony — in development
- SpacemiT Studio — online AI-native IDE and flashing workbench
SpacemiT Roadmap
| Product Line | Current / Planned | Notes |
|---|---|---|
| K series | K1, K3 in production; K5, K7 planned | Edge to AI-compute SoCs |
| V series | V100 taped out; V200 in development | Server-oriented chips |
| AI cores | A60, A100 in production; A200 coming | 2 / 32 / 256 TOPS classes |
| CPU cores | X60, X100 in production; X200 ready, X300 2027 | Performance scaling to 20 SPECint2k6/GHz |
K3 vs K1: What Changes
| Feature | K1 | K3 |
|---|---|---|
| CPU | 8x X60 @ 1.6 GHz | 8x X100 @ 2.4 GHz |
| DMIPS | ~50K | ~130K |
| AI | 2 TOPS | 60 TOPS |
| Memory | LPDDR4X up to 16 GB | LPDDR5 up to 32 GB |
| ISA Profile | RV64GCV | RVA23 + RVV 1024-bit |
| LLM Scale | 0.5-1B params | Up to 30B params local |
| Real-Time Domain | Limited | 2x RT cores + 3 MB cache + 10x CAN-FD |
Resources
- SpacemiT Official Website
- SpacemiT Document Center
- Sohu: SpacemiT K3 Details (Chinese)
- QQ News: K3 Forum Report (Chinese)
- EDN China: K3 "CPU 2.0" Analysis (Chinese)
- Bianbu Linux SDK (Gitee)
Multi-Language Summary
EN: SpacemiT K3 is a flagship RISC-V AI CPU with 8x X100 cores @ 2.4 GHz, 8x A100 AI cores delivering 60 TOPS, and up to 32 GB LPDDR5. It is the first mass-production RISC-V chip to claim RVA23 compliance, 1024-bit RVV, native FP8, and full virtualization.
RU: SpacemiT K3 — флагманский AI CPU на базе RISC-V с 8 ядрами X100 @ 2,4 ГГц, 8 ядрами A100 (60 ТОПС) и до 32 ГБ LPDDR5. Первый серийный чип RISC-V с заявленной поддержкой RVA23, 1024-битного RVV, нативного FP8 и полной виртуализации.
ES: SpacemiT K3 es una CPU AI RISC-V insignia con 8 núcleos X100 @ 2,4 GHz, 8 núcleos A100 de 60 TOPS y hasta 32 GB LPDDR5. Es el primer chip RISC-V en producción masiva con RVA23, RVV de 1024 bits, FP8 nativo y virtualización completa.
FR: Le SpacemiT K3 est un CPU AI RISC-V phare avec 8 cœurs X100 @ 2,4 GHz, 8 cœurs A100 de 60 TOPS et jusqu'à 32 Go de LPDDR5. Premier chip RISC-V en production de masse à revendiquer la conformité RVA23, le RVV 1024 bits, le FP8 natif et la virtualisation complète.
DE: Der SpacemiT K3 ist ein Flaggschiff-RISC-V-AI-CPU mit 8 X100-Kernen @ 2,4 GHz, 8 A100-Kernen mit 60 TOPS und bis zu 32 GB LPDDR5. Erster RISC-V-Chip in Massenproduktion mit behaupteter RVA23-Konformität, 1024-Bit-RVV, nativem FP8 und vollständiger Virtualisierung.
FA: SpacemiT K3 یک CPU AI پرچمدار RISC-V با ۸ هسته X100 با فرکانس ۲.۴ گیگاهرتز، ۸ هسته A100 با ۶۰ TOPS و تا ۳۲ گیگابایت LPDDR5 است. اولین چیپ RISC-V تولید انبوه با مطابقت RVA23، RVV ۱۰۲۴ بیتی، FP8 بومی و مجازیسازی کامل.
Conclusion
The K3 is SpacemiT's most aggressive RISC-V part yet. By combining a modern out-of-order CPU, wide-vector units, native low-precision AI formats, and virtualization in a single mass-production chip, it targets the same AI PC, robotics, and edge-server use cases that ARM and x86 currently dominate. Whether the four world-first claims hold up over time, the device is already in public demonstrations and shipping hardware, which makes it a concrete data point in the "RISC-V enters high-performance computing" narrative.