gigaRAM
Open AI Infrastructure

AI infrastructure is becoming a stack, not a box

Open rack, networking and software standards reduce vendor lock-in. gigaRAM is the memory layer inside that stack.

Explore GigaRAM

Open AI infrastructure

AI workloads
Training and inference
Open software stack
vLLM · PyTorch · ROCm
Linux
Operating system
Open interconnect / networking
Open standards
Accelerator
GPU or AI accelerator
Open Rack
OCP Open Rack Wide (ORW)

Two routes

One stack, two architectures

Vertically integrated stack

GPUNVLink / NVSwitchCUDAVendor librariesAI frameworks

Open infrastructure

Open RackAcceleratorOpen interconnect / networkingLinuxOpen software stackAI workloads

NVIDIA remains the benchmark for highly integrated AI infrastructure. The open route is developing alongside it — a second path, not a replacement.

OCP Open Rack Wide

ORW in plain terms

01

A standard, not a product

ORW is an open AI rack specification within the Open Compute Project. It is not an accelerator, not an AMD product and not a new class of GPU.

02

One architecture, many vendors

Servers, accelerators, networking, storage, power and liquid cooling are designed to a shared form factor and stay compatible with each other.

03

AMD Helios is one example

The AMD Helios rack system with MI400 / MI455X-generation accelerators is one implementation of this approach. An important one — not the only one.

One rack, several suppliers

OCP ORW
Compute
Vendor A · B · C
Accelerators
Vendor A · B · C
Networking
Vendor A · B · C
Storage
Vendor A · B · C
Power
Vendor A · B · C
Liquid cooling
Vendor A · B · C

The open standard defines mechanics, power and cooling. Who supplies each component is the customer's decision.

The Open AI Software Stack

The open software stack

ORW itself does not define the software stack. Its significance lies elsewhere: hardware infrastructure becomes more standardized and open, while software is increasingly built around Linux and the open-source ecosystem.

AI Models
Models and applications
vLLM · SGLang
Inference engines
PyTorch · JAX
Frameworks
ROCm · open accelerator software
Accelerator layer
Linux
Operating system
Kubernetes · Slurm
Orchestration and scheduling
Open compute infrastructure
Open infrastructure (ORW and others)

Where gigaRAM sits

A memory layer, not another AI framework

gigaRAM sits between AI workloads and a server's physical memory, intelligently distributing data across DRAM and NVMe.

AI Applications & Models
Workloads
PyTorch · vLLM · databases · containers
Runtime
gigaRAM Intelligent Memory Layer
Intelligent memory layer
DRAM ↔ NVMe
Physical memory and flash
CPU / GPU infrastructure
Hardware

Open compute needs open memory infrastructure.

  • Transparent to applications
  • Linux-native
  • Hardware-independent approach
  • Designed for open infrastructure
  • Reduces dependence on expensive DRAM
  • Enables larger workloads on existing servers
  • Suitable for AI, databases and memory-intensive applications

gigaRAM does not replace accelerator HBM and is not part of ORW or Helios. It is a separate software infrastructure layer that can complement open AI infrastructure.

Measured on real workloads

Redis−50%
DRAM replacedTotal RPS +5.6%p99 SET −21.7%
PostgreSQL−40%
DRAM replacedTPS ≈ −2%

Redis: 48 vCPU, 50M keys at 16 KiB. PostgreSQL is a deliberately conservative projection. Both are verified by a pilot on your own workload.

Comparison

A trade-off, not a winner

Vertically Integrated

  • One vendor
  • GPU
  • Interconnect
  • Networking
  • Software ecosystem

Maximum integration and a mature ecosystem.

Open Infrastructure

  • Multiple vendors
  • Open rack standards
  • Open networking / interconnect standards
  • Linux
  • Open-source AI software
  • Independent infrastructure software

Flexibility, supplier choice and room for independent infrastructure software.

Neither approach is universally right. The choice depends on timelines, scale, team skills and procurement strategy.

Trend 2027+

AI infrastructure is becoming a stack, not a box.

The next phase of the market is not only a GPU race. Competition moves up to the level of the whole AI factory.

Compute
Memory
Networking
Storage
Cooling
Orchestration
Software

As open standards spread, space opens up for independent companies that build the best component of each layer.

The future of AI infrastructure will not be defined by compute alone.

gigaRAM is building the intelligent memory layer for that future.

Explore GigaRAM

In short

Frequently asked questions

What is OCP Open Rack Wide (ORW)?

An open AI rack standard developed within the Open Compute Project. It defines mechanics, power, cooling and hardware integration, so different manufacturers can build compatible servers, accelerators, networking and storage.

Is ORW an AMD product or a new type of GPU?

No. ORW is an open rack specification — neither a specific vendor's product nor a class of accelerator. Any hardware manufacturer can build to it.

How do AMD Helios and MI400 / MI455X relate to it?

AMD Helios is a rack-scale AI system built on open rack architecture principles, using MI400 / MI455X-generation accelerators. It is an important example of the approach — not the standard itself.

Does ORW define the software stack?

No. ORW standardizes hardware infrastructure. Software is increasingly built around Linux and the open-source ecosystem: ROCm and other open accelerator layers, PyTorch and JAX, vLLM and SGLang, Kubernetes and Slurm.

Where does gigaRAM sit in this stack?

gigaRAM is a memory infrastructure layer between AI workloads and a server's physical memory. It transparently distributes data across DRAM and NVMe, letting larger workloads run on existing servers. It is not part of ORW or Helios and does not replace accelerator HBM.