MANGO Alphonso™
High EndAMD GPUsFull Stack Server Optimized for AI
High performance, high efficiency, turnkey AI infrastructure solution.
Turnkey Solution. Managed as one.
Full stack AI system. LLMBoost™, BoostX™ RoCE AI and MAC tune throughput, cost and operations as one.
Full stack AI system.
Built for large scale training and high volume inference, delivered as one validated turnkey stack.
Network and TCO, optimized in house.
LLMBoost™ and onboard BoostX™ RoCE AI offload the data path to raise throughput and lower total cost of ownership.
Unified system. One screen.
MAC (Mango AI Center) manages distributed nodes as a single system, with a real time dashboard for easy operations.
Compute and networking, balanced by design.
Compute and networking balanced to keep every GPU fed, never idle.
MANGO ALPHONSO™ · single node
System Memory
Gen5 x16
CPU 1
Gen5 x16
2× PCIe Gen5 switch
PCIe Switch
PCIe Switch
GPU
MI355X
GPU
MI355X
GPU
MI355X
GPU
MI355X
BoostX™
RoCE AI
BoostX™
RoCE AI
BoostX™
RoCE AI
BoostX™
RoCE AI
Up to 400 Gbps per GPU
Ethernet, RoCE v2, UEC-ready
System Memory
Gen5 x16
CPU 2
Gen5 x16
2× PCIe Gen5 switch
PCIe Switch
PCIe Switch
GPU
MI355X
GPU
MI355X
GPU
MI355X
GPU
MI355X
BoostX™
RoCE AI
BoostX™
RoCE AI
BoostX™
RoCE AI
BoostX™
RoCE AI
Up to 400 Gbps per GPU
Ethernet, RoCE v2, UEC-ready
More performance. Lower cost.
Alphonso™ is tuned as a full stack system, not a collection of parts. That shows up as more throughput from every GPU, and a lower cost to run the cluster over its life.
01 · Performance
Superior performance per dollar.
Same class of GPU, same cluster, same workload. What changes is the server, tuned as a full stack system to deliver more throughput.
Performance/$ = throughput (tok/s/GPU) ÷ modelled 3 year TCO · indexed, baseline = 100 · higher is better
vs MI355X(Base) Server
UP TO
+41.7%
more performance per dollar
100
141.7
MI355X (Base) ServerMango Alphonso
vs NVIDIA B200 Server
UP TO
+47.0%
more performance per dollar
100
147
B200 ServerMango Alphonso
GLM 5/5.1 FP8 · coding agent, 8K in / 1K out · best observed result
02 · TCO
Lower TCO per cluster.
The saving comes from the node and the fabric together, with UEC-ready features on the DPU instead of premium switch silicon.
3 year TCO · 256 GPU cluster · indexed, lower is better
NVIDIA B200 cluster
100
INDEX
Mango Alphonso™
79
INDEX
~21% lower
3 year TCO on a 256 GPU build.
$4.6M saved over 3 years
Modelled 3 year TCO on a 256 GPU build, at matched scale and topology.
Same scale, same topology
256 GPU, 32 nodes, 8 leaf + 4 spine at 1 : 1, 400G end to end.
UEC-ready on the DPU
Packet spraying and congestion control move off the switch, so a standard 400GbE switch is enough.
Ethernet, not a proprietary fabric
RoCE v2 on the spine you already run, with standard optics and standard skills.
Baseline = 8× B200 server · TCO includes power and maintenance · pricing varies by vendor and contract.
Every node, managed as one.
An agent on each node streams live telemetry to a single master, so you monitor, validate and operate the whole cluster from one screen.
Agent AALPHONSO™ node
Agent BALPHONSO™ node
Agent CALPHONSO™ node
Telemetry
Unified control
Mango AI Center · Master
MASTER: RUNNINGSYSTEM HEALTHY
Dashboard
Software
Inventory
Monitoring
Fabric
Logs
Cluster Overview
Monitoring 3 active nodes
eval2Online
eval4Online
smc21Offline
CPU
42%
Memory
58%
GPU
8/8
Health
OK
Single pane of glass
MAC treats distributed nodes as one system, with a GUI dashboard for every layer of the stack.
Real time health
Cluster overview with live CPU, memory, GPU, temperature and power across every active node.
Package consistency
Version drift checks keep drivers, toolkits and runtimes aligned across every node.
Network and fabric visibility
Per NIC RDMA bandwidth and switch to port topology, visualized in real time.
Latest from MangoBoost.
Benchmark milestones, engineering deep-dives and company announcements.
First-ever Multi-Region Heterogeneous Inference: MangoBoost Sets New Standards for Global AI Scaling in MLPerf Inference v6.0
MangoBoost
Learn more
MangoBoost Pushes AI Scalability in MLPerf Training v5.1: First-Ever Multi-Node MI325X Results Highlight AMD's Scaling Capabilities
MangoBoost
Learn more
Dell spotlighted MangoBoost’s LLMBoost™ for MLPerf Inference v5.1
Dell Technologies
Learn more
Pioneering Multinode Heterogeneous Inference: MangoBoost and AMD, in Collaboration with OEM Partners, Sets Record Llama2-70B Performance in MLPerf Inference v5.1
MangoBoost
Learn more
How Supermicro/AMD servers boost AI performance with MangoBoost
Performance Intensive Computing
Learn more
Supermicro GPU Servers With MangoBoost LLMBoost™ AI Enterprise Software : Unlocking the full potential of AMD Instinct™ GPUS
Supermicro
Learn more