# MangoBoost

> MangoBoost is a Bellevue- and Seoul-rooted AI infrastructure company that attacks the expensive idle time around GPUs. Its stack spans programmable data processing units, Ethernet networking, disaggregated all-flash storage, AI servers, inference and training software, and fleet management. The pitch is practical: offload networking and storage from host CPUs, keep accelerators supplied with data, support mixed generations of AMD and other hardware, and let operators buy an integrated system or adopt the pieces they need.

- **Founded:** 2022
- **Headquarters:** Bellevue, United States
- **Founders:** Jangwoo Kim (Co-Founder and CEO), Eriko Nurvitadhi (Co-Founder and CAIO), Dongup Kwon (Co-Founder and CTO), Junki Park (Co-Founder and COO/CFO)
- **Team size:** Approximately 120-140 employees; LinkedIn lists 51-200 and the company says 120+ experts.
- **Products:** BoostX RoCE AI, BoostX NRT, BoostX Base, LLMBoost, Mango Alphonso
- **Notable:** Raised $55 million in a 2023 Series A and reported more than $100 million in total capital by 2026., Published the first DPU-accelerated submission in MLPerf Storage v1.0 and reported the top-performing Ethernet-based single-host system in that round., Reported line-rate 400G Ethernet and near-local SSD behavior in its MLPerf Storage v2.0 configuration.

## Products & services

- **BoostX RoCE AI** — A UEC-ready DPU/RNIC that provides up to 400GbE aggregate bandwidth, RoCEv2 RDMA, packet spraying and programmable congestion control for multi-node AI clusters.
- **BoostX NRT** — A DPU for hardware-offloaded NVMe-oF target service, used in headless storage enclosures and disaggregated storage systems.
- **BoostX Base** — A programmable DPU platform and SDK for networking, storage, security, virtualization and custom offload workloads.
- **LLMBoost** — OpenAI-compatible enterprise software for LLM inference and training, with auto-tuning, parallelism, batching, memory management and orchestration across AMD GPU systems.
- **Mango Alphonso** — A turnkey AMD GPU server that combines LLMBoost, BoostX RoCE AI networking and Mango AI Center management.
- **Mango Kesar Controller** — A 2U, DAOS-based all-flash storage controller for AI and HPC, advertised at up to 300 GB/s of RoCEv2 bandwidth.
- **Mango Kesar Enclosure** — A CPU-free, DRAM-free all-flash NVMe-oF enclosure using BoostX NRT, with redundant paths and up to 26 dual-port NVMe drives.
- **Mango AI Center** — A fleet operations platform for monitoring GPUs, DPUs, NICs, storage, power, thermals, packages and inference services from one dashboard.
- **MASS** — Mango Accelerated Scalable Storage, a GUI control plane for provisioning and managing Kesar storage systems.

## Achievements

- Raised $55 million in a 2023 Series A and reported more than $100 million in total capital by 2026.
- Published the first DPU-accelerated submission in MLPerf Storage v1.0 and reported the top-performing Ethernet-based single-host system in that round.
- Reported line-rate 400G Ethernet and near-local SSD behavior in its MLPerf Storage v2.0 configuration.
- Submitted the first public multi-node AMD Instinct MI325X MLPerf Training result in v5.1.
- Demonstrated MLPerf Inference v6.0 across three AMD GPU generations and two regions, reporting 95.5% scaling efficiency.
- Joined the Linux Foundation OPI Project, Ultra Ethernet Consortium and UALink Consortium.
- Built a portfolio the company describes as 50+ patents on top of more than 15 years of research.

## Latest updates

- **2026-07** — Prepared to demonstrate its full agentic AI stack, InferenceX work with AMD and a preview of LLM as a Service at AMD Advancing AI 2026.
- **2026-06** — Presented the BoostX-NTI NVMe/TCP acceleration architecture in the ISCA 2026 industry session.
- **2026-06** — Selected for Presight's AI Accelerator Cohort II, connecting the company with Presight, G42 and prospective UAE deployments.
- **2026-04** — Reported a first multi-region, three-generation AMD GPU submission in MLPerf Inference v6.0 with 95.5% scaling.
- **2025-12** — Closed a Series B round reported in supplied company data at $81.33 million.
- **2025-11** — Reported first multi-node AMD MI325X results in MLPerf Training v5.1 and joint submissions with AMD, Dell and Supermicro.
- **2025-08** — Published MLPerf Storage v2.0 results for DPU-accelerated NVMe/TCP over a 400G Ethernet link.

## Links

- Website: https://www.mangoboost.io/
- LinkedIn: https://www.linkedin.com/company/mangoboost
- Twitter/X: https://x.com/MangoBoost_Inc
- GitHub: https://github.com/MangoBoost

---

Profile page: https://yespress.io/mangoboost
Published by YesPress — https://yespress.io
Last updated: 2026-08-21