FPS: --

Compute

The hardware underneath

Every eval and every training run has to land on real silicon. The approach is simple: own the machines that pay off sitting on a desk, and rent datacenter GPUs by the hour when a run needs more than a desk can give. Both halves are Blackwell.

On the desk · unified memory

NVIDIA DGX Spark

In the lab

GB10 Grace-Blackwell

128 GB unified memoryNVFP4-native tensor cores~1 PFLOP FP4

A full Grace-Blackwell stack on a desk. The workhorse for local inference and for hosting the eval harness.

ASUS Ascent GX10

In the lab

GB10 Grace-Blackwell

128 GB unified memoryBlackwell FP4aarch64 / Arm

The same GB10 platform in a different chassis — a second Spark-class node to cross-check results and scale runs out.

NVIDIA DGX Station

Roadmap

GB300 Grace-Blackwell Ultra

Up to 784 GB coherentHBM3e + LPDDR5XBlackwell Ultra FP4

NVIDIA's first-party GB300 desktop — the big-memory station tier, where the heavier training runs and full eval sweeps go once they outgrow a Spark.

Dell Pro Max with GB300

Incoming

GB300 Grace-Blackwell Ultra

GB300 desktop workstationMassive coherent memoryBlackwell Ultra FP4

Dell's take on the Grace-Blackwell Ultra desktop — the same GB300 tier from another builder. On the list to add as it ships.

In the cloud · rented on demand

When a training run or a full sweep outgrows the owned hardware, we burst to datacenter Blackwell — spun up for the job, torn down after.

NVIDIA B200

On demand

Blackwell · datacenter

192 GB HBM3e8 TB/s bandwidthSingle card → nodes

Datacenter Blackwell we rent when a job needs more than a desk can give — ganged from one card up to large multi-GPU nodes for training and big sweeps.

NVIDIA B300

On demand

Blackwell Ultra · datacenter

288 GB HBM3e8 TB/s bandwidthMulti-GPU

Blackwell Ultra in the cloud — the most memory per GPU in the lineup. Where the largest training runs land when the owned hardware isn't enough.

Why unified memory changes the math

On a normal GPU, model weights have to fit in a few tens of gigabytes of dedicated VRAM, and anything bigger gets copied back and forth across a PCIe straw. Grace-Blackwell stitches the CPU and GPU together over a coherent link into one large memory pool the GPU addresses directly. A quantized 30-billion-parameter model lands in the memory footprint of a mid-size game — and it runs on tensor cores built to do 4-bit math natively rather than emulating it. That’s the lever the whole program pulls on.

This page grows as the fleet does. As we bring platforms onto the bench and collect real measurements, the specs and results land here — and feed straight into the Evals matrix.