Representative catalog

Advertised memory by hardware

Dedicated VRAM/HBM and unified memory are labeled separately. Bars use a logarithmic scale.

Accelerator or unified-memory capacityVendor-advertised GB · logarithmic bar scale
GeForce RTX 409024 GB · dedicated
Radeon RX 7900 XTX24 GB · dedicated
GeForce RTX 509032 GB · dedicated
Radeon PRO W790048 GB · dedicated
H100 SXM80 GB · dedicated
H200141 GB · dedicated
Instinct MI300X192 GB · dedicated
Instinct MI325X256 GB · dedicated
Representative AI hardware memory capacities from official vendor specifications
HardwareClassMemoryTypeQualificationSource
NVIDIA GeForce RTX 4090Consumer GPU; software support and usable free VRAM depend on the system configuration.consumer gpu24 GBGDDR6XDedicatedVendor ↗
AMD Radeon RX 7900 XTXBackend and kernel support should be checked separately from capacity.consumer gpu24 GBGDDR6DedicatedVendor ↗
NVIDIA GeForce RTX 5090Consumer GPU; advertised capacity does not guarantee runtime compatibility.consumer gpu32 GBGDDR7DedicatedVendor ↗
AMD Radeon PRO W7900Workstation GPU with ECC support.workstation gpu48 GBGDDR6 ECCDedicatedVendor ↗
NVIDIA H100 SXMH100 NVL is a separate 94GB configuration; this row represents the 80GB SXM capacity.datacenter accelerator80 GBHBMDedicatedVendor ↗
NVIDIA H200Datacenter accelerator capacity.datacenter accelerator141 GBHBM3eDedicatedVendor ↗
AMD Instinct MI300XDatacenter accelerator capacity.datacenter accelerator192 GBHBM3DedicatedVendor ↗
AMD Instinct MI325XDatacenter accelerator capacity.datacenter accelerator256 GBHBM3EDedicatedVendor ↗
Apple Mac Studio M3 Ultra (maximum configuration)Maximum configurable unified memory shared by CPU, GPU, and the operating system; not equivalent to dedicated VRAM.unified memory system512 GBUnified memoryUnified / sharedVendor ↗

Connect models to memory

See which open weights fit.

The open-model catalog estimates Q4 and Q8 weight memory with conservative format and loading overhead, then compares Q4 estimates against capacity tiers from 8GB through 512GB.

Open the model hardware-fit matrix →