AMD Instinct™ MI355X GPUs
- Form
- OAM Accelerator
- Memory
- 288 GB HBM3E
- Best for
- Frontier training and latency-critical inference
harnessing amd
We build and run multi-megawatt AMD Instinct™ GPU clusters on infrastructure we own, in the jurisdiction you choose. Dedicated to your workload, fully customizable, and yours to command. Multi-megawatt capacity, order to live in 16+ weeks.
AMD Helios Rackscale Solution
- Form
- Rack-scale platform
- Memory
- 31 TB HBM4
- Best for
- Rack-scale training and inference at extreme density
GPU Specifications
- Architecture
- AMD CDNA 4 (4th Gen)
- Die design
- Multi-chiplet, 8× XCD
- Compute units
- 256
- Process
- TSMC 3 nm + 6 nm
- Typical board power (TBP)
- Up to 1400 W
- FP4 / FP6 peak
- 10.1 PFLOPS
- FP8 peak
- 5.0 PFLOPS
- FP16 / BF16 peak
- 2.5 PFLOPS
- Precision formats
- FP16, FP8, FP6, FP4 (incl. MXFP6 / MXFP4)
Memory
- Capacity
- 288 GB HBM3E
- Bandwidth
- 8 TB/s
Board
- Form factor
- OAM Accelerator
- Cooling
- Direct liquid or air cooled
Deployable node config
- Software stack
- Open ROCm
- Architecture
- 4th-Gen AMD CDNA
MI400 Series specs are subject to change. Coming Soon.
MI400 series GPUs
- Memory
- Up to 432 GB HBM4 per GPU
- Form factor
- Accelerator
- Best for
- Rack-scale training and inference at extreme density
Helios rackscale solution
- Platform
- Helios rack-scale (open / OCP)
- GPUs per rack
- 72
- Total HBM4
- 31 TB (72-GPU config)
- Aggregate bandwidth
- 1.4 PB/s
- Rack FP8 peak
- Up to 1.4 exaFLOPS
- Rack FP4 peak
- Up to 2.9 exaFLOPS
- Cooling
- Direct liquid
Flexible I/O & Open standards
- Interconnect
- Open UALink
- Design
- Open Rack Wide (ORW)
- Memory
- HBM4
our deep collaboration
AMD on the silicon. OneQode on the infrastructure.
A deep collaboration to deploy AMD Instinct™ GPUs in key global markets. MI355X anchors the rollout today, with AMD Helios planned on the road to 1GW.
01
Deep AMD collaboration
As an AMD Instinct™ Cloud Provider, OneQode deploys AMD Instinct™ GPUs globally, anchored on MI355X GPUs with AMD Helios planned. AMD on the silicon, OneQode on the infrastructure and market access.
02
Where hyperscalers don't reach
When you need an operator who can deploy hardware at scale in the markets you care about, we can help. Present in 30+ datacentres, we operate our own international network, and can source footprint where others can't.
03
Built on open standards
AMD's stack is open end to end: ROCm software, UALink and Ultra Ethernet interconnects, Open Rack Wide hardware. Workloads and tooling stay portable, not locked to one vendor's proprietary ecosystem.
04
Roadmap to 1GW and beyond
The 110 MW Namsskogan campus in Norway, under a binding LOI, is part of the next-gen AMD Helios conversation and is coming online. We're building and ramping AI infrastructure at scale.
From plan to live capacity
-
Sovereign by design
Jurisdictional control over where your compute lives and where data flows.
-
Low latency
Near-instant network: our global backbone routes traffic via the shortest path.
-
Deploy anywhere
Capacity stood up where hyperscalers don't reach.
-
Full-stack, one roof
Power, site, silicon, network, storage, and 24/7 operations; fully integrated.
Infrastructure
The rest of the stack
One operator, from power to GPU, across APAC and the Pacific.
