Mango DPU
Free BottleneckJBOF SystemBoostX™ NRT
The Ultimate Full-Duplex DPU for AI-Scale Storage.
Driving The Next Era of Scalable Storage.
BoostX™ NRT drives your data faster, Line-rate 200GbE storage with zero overhead. Plug into your JBOF storage system.
Stable Throughput. Proven at Scale.
Stable Throughput. Proven at Scale.
Drives 200GbE throughput in under 75W.
Save Rack Cost beyond TCO.
Save Rack Cost beyond TCO.
Low-power LP form factor, ready for headless JBOF system.
Full-Duplex. Full Speed.
Full-Duplex. Full Speed.
Full-duplex line rate maximizes throughput in mixed I/O.
Stable Throughput. Proven at Scale.
Drives 200GbE throughput in under 75W.
Save Rack Cost beyond TCO.
Low-power LP form factor, ready for headless JBOF system.
Full-Duplex. Full Speed.
Full-duplex line rate maximizes throughput in mixed I/O.
CPU Free. Bottleneck Free.
Bypass the CPU. Skip the DRAM. That's BoostX™ NRT.
We took out the CPU and DRAM from the data path.
BoostX™ NRT drives pure line-rate 200Gbps directly to your SSDs.
Before
CPU handles all data I/O, causing bottlenecks.
After ｜ With BoostX™ NRT
Achieves Headless JBOF by eliminating CPU/DRAM.
The Leanest Path from Ethernet to Flash.
BoostX™ runs the NVMe over RoCEv2 target in hardware: Standard Ethernet in, PCIe Gen5 NVMe out, and no server CPU or DRAM in between.
Standard Ethernet in, NVMe out, and a headless target in between. Cost falls at every stage of the path.
Standard Ethernet
RoCEv2 · in
any switchany NICany optic
KESAR-E · Headless JBOF · 2U
BoostX™ NRT
target · canister A
BoostX™ NRT
target · canister B
1.6 Tbps · HA
Server CPUDRAM
no server CPU or DRAM
NVMe SSD
26 × PCIe Gen5 · out
Standard Ethernet
RoCEv2 runs on standard Ethernet, so you are free to choose any switch, NIC, and optic vendor. That avoids the InfiniBand premium and the single-vendor lock it carries.
InfiniBand premium on the network fabric · widens on open whitebox · lower is better · illustrative
Lower
Higher
NRT (whitebox Eth)InfiniBand
Headless Offload
BoostX offloads the NVMe over RoCEv2 target into hardware, so the enclosure carries no server CPU, DRAM, or software stack. It is a true headless JBOF, immune to memory price swings.
Compute overhead to add the same capacity · CPU tax verified at zero · lower is better · illustrative
Lower
Higher
NRT (headless)Server target
NVMe Efficiency
PCIe Gen5 NVMe attaches directly, with no SAS HBA, expander, or RAID layer. At equal performance it draws far less power in far less rack space.
Power to deliver the same performance · typ active, peak excluded · lower is better · illustrative
Lower
Higher
NRT (NVMe)SAS target
Basis · Network: KESAR's reference fabric is open whitebox Ethernet you can source from competing vendors of your choice. The premium runs from an audit-safe floor up to a wider gap on open whitebox; InfiniBand has no whitebox path. Media: PCIe Gen5 NVMe versus 24G SAS on an iso-performance basis. Compute: a reseller street estimate of the server BOM a software target carries; CPU tax verified at zero. All figures model only the delta a conventional alternative carries — no absolute total is fabricated.
Performance.
From Microseconds to Petabytes, One Drive for Every Workload.
Full Duplex Architecture. Read and Write, Simultaneously Unleashed. A 5:5 mixed workload boosts IOPS by 73% and throughput by 83% beyond single-direction limits.
4KiB Random I/O (IOPS)
Random Read5.6M
Random Write5.7M
RW (7:3)7.2M
RW (5:5)peak9.7M
128KiB Sequential I/O (GB/s)
Seq. Read24.3
Seq. Write24.4
Seq. RW (7:3)33.8
Seq. RW (5:5)peak44.4
Available on the System.
Need to Scale? Explore Our JBOF.
The DPU is the foundation. See the MangoBoost products built to extend and complete the system around it.
Kesar Enclosure I Storage
- Daos-based high-performance storage engine
- BoostX™ DPU-based acceleration & offload technology
- User-friendly management solution
What workloads is it built for?
Explore the full range of workloads MangoBoost DPU is built to accelerate, from AI infrastructure to large-scale cloud and beyond.
Hyperscale Data Centers
Powers the storage backbone of large-scale cloud and web service providers with high density and low TCO.
CloudWeb-scale
AI/ML Training Infrastructure
Feeds GPU clusters with the throughput and capacity needed for large-model training and inference pipelines.
AI/MLGPU
Content Delivery & Streaming
Delivers the sustained read bandwidth required for CDN edge caches, video streaming, and live media services.
CDNMedia
HPC & Research Computing
Provides high-bandwidth shared storage for scientific simulation, genomics, and large-dataset analytics workloads.
HPCResearch
Latest from MangoBoost.
Benchmark milestones, engineering deep-dives and company announcements.
Related news and blog posts will appear here as they are published.
Need a Boost?
Our team is at the ready to create a customized plan for you to optimize and scale your business.