Skip to content
Hardware

Supported GPUs

Tier eligibility is computed from total VRAM across a system, not from a single card, so two RTX PRO 6000s reach Frontier the same way a DGX Spark does.

Entry class

12–17 GB. Runs the Entry tier.
GPUVRAMBandwidthTDPArchitectureLanesTiersResults
GeForce RTX 3060 12GB12 GB360 GB/s170 WampereBF16 · INT4Entry
GeForce RTX 4060 Ti 16GB16 GB288 GB/s165 WadaBF16 · INT4Entry
GeForce RTX 407012 GB504 GB/s200 WadaBF16 · INT4Entry
GeForce RTX 4080 SUPER16 GB736 GB/s320 WadaBF16 · INT4Entry
GeForce RTX 507012 GB672 GB/s250 WblackwellBF16 · INT4 · NVFP4Entry
GeForce RTX 5070 Ti16 GB896 GB/s300 WblackwellBF16 · INT4 · NVFP4Entry
GeForce RTX 508016 GB960 GB/s360 WblackwellBF16 · INT4 · NVFP4Entry

Enthusiast class

24–48 GB. Runs Entry and Enthusiast.
GPUVRAMBandwidthTDPArchitectureLanesTiersResults
GeForce RTX 309024 GB936 GB/s350 WampereBF16 · INT4Entry, Enthusiast
GeForce RTX 3090 Ti24 GB1,008 GB/s450 WampereBF16 · INT4Entry, Enthusiast
GeForce RTX 409024 GB1,008 GB/s450 WadaBF16 · INT4Entry, Enthusiast
GeForce RTX 509032 GB1,792 GB/s575 WblackwellBF16 · INT4 · NVFP4Entry, Enthusiast
RTX A600048 GB768 GB/s300 WampereBF16 · INT4Entry, Enthusiast
RTX 6000 Ada48 GB960 GB/s300 WadaBF16 · INT4Entry, Enthusiast
L40S48 GB864 GB/s350 WadaBF16 · INT4Entry, Enthusiast

Frontier class

64 GB and up, including unified memory. Runs all three tiers.
GPUVRAMBandwidthTDPArchitectureLanesTiersResults
RTX PRO 6000 Blackwell96 GB1,792 GB/s600 WblackwellBF16 · INT4 · NVFP4Entry, Enthusiast, Frontier
DGX Spark (GB10)128 GB273 GB/s140 Wblackwell-gb10 · unifiedBF16 · INT4 · NVFP4Entry, Enthusiast, Frontier

A note on DGX Spark

GB10 is an aarch64 system with 128 GB of LPDDR5X shared between CPU and GPU. It has enormous capacity and modest bandwidth, which is exactly the profile the tier system exists to represent: expect it to hold models nothing else here can, and to trail discrete cards on decode throughput.

Its power telemetry is measured at the SoC module, not a GPU board. Those are different physical quantities, so a Spark result is never ranked for energy against a discrete card — the wattage is shown, clearly labelled, and left out of the efficiency view.