← Back to the platform
Platform layer Line 03 of 07

Forge

The raw power, integrated.

Racks are easy to buy and hard to make sing. Forge is the compute engine itself, the right accelerators, matched to your science, racked, tuned and benchmarked before a single job runs.

8 compute platforms
PFLOP per rack
Any silicon, one stack
1/6 Hexa: Compute
What you actually buy

The compute your science needs, not the chip on the sales sheet.

Every Forge platform carries the same three commitments. Compute is chosen and proven, not simply shipped.

The silicon

Right chip for the science

We match accelerators to the workload, training, inference, memory-bound, not to a vendor quota.

The integration

Racked, tuned, benchmarked

Arrives as a proven machine with real numbers on your codes, not a parts list to assemble.

The neutrality

Silicon-agnostic

NVIDIA, AMD or custom, one portable software stack, so your code never marries a vendor.

Pre-configured · ready to ship

Compute bricks, built before you order.

A family of modular, pre-validated chassis, deskside to rack-scale, burned-in and benchmarked on our floor. When the science can’t wait eighteen weeks, a brick ships in days.

Família de bricks de computação modulares e pré-configurados da ARETE
Pre-integrated & burned-in
Deskside to rack-scale
Ships in days, not months
One portable software image
Eight platforms, one engine

From a single tuned node to rack-scale coherent compute.

One continuous engine. Pick the accelerator your science needs today; every tier above is more power, never a re-port.

Accelerators, the engine

Forge Hopper

talk to architect

A proven GPU platform for training and HPC, the workhorse for most scientific loads.

Forge Blackwell

talk to architect

Latest-generation accelerators for frontier-model training and the heaviest simulation.

Forge Inference

talk to architect

Throughput-optimized nodes tuned for serving models at scale, not training them.

Systems, racked & ready

Forge Node

talk to architect

A single, fully tuned GPU server, installed, monitored and ready to run today.

Forge Server

talk to architect

A dense multi-GPU system with a coherent NVLink domain, one big accelerator.

Forge SuperNode

talk to architect

Rack-scale coherent compute presented as a single machine for the largest jobs.

Silicon strategy, future-proof

Forge CPU

talk to architect

A high-core CPU tier for pre/post-processing and memory-bound science beside the GPUs.

Forge Custom

talk to architect

An accelerator-agnostic path: pick the best silicon per workload, one stack across all.

Why it holds together

You never re-port to change chips.

Most vendors lock your code to one accelerator, then charge you to move. Forge keeps one portable stack, so silicon generations swap under your science, not through it.

SPEC

Match silicon to the workload

We profile your codes and pick the accelerator that earns its watts on your science.

BUILD

Integrated, tuned, benchmarked

Racked and validated with real numbers before it ships, no assembly required.

EVOLVE

Swap generations, keep the code

Next-gen silicon drops in under the same stack, capability rises, your code stays.

“We don’t sell chips. We sell the compute your science actually needs.”

Powered by Hexa

Forge is the raw Compute face.

HC, Compute, is where Forge lives: the accelerated engine at the heart of the ring. Connect, Cool and the rest wrap around what Forge delivers.

HC Compute
See the framework →
Tell us the workload. We’ll forge the compute.
Architect my Forge → All seven lines