GPU catalog: NVIDIA and AMD GPUs, filters
What the catalog lists per GPU, how to filter it and where the numbers come from.
On this page · 5 sections
The GPU catalog is the spec sheet behind every estimate: 48 GPUs · 10 Jetson modules — NVIDIA and AMD GPUs from GeForce cards to rack-scale NVL72 systems, and the Jetson modules the Edge tool sizes. A few cards are listed for reference only — the China-market H20, for example, appears in the catalog but is not offered as a sizing target in the calculator, What fits, Optimize or TCO.
What each row holds
- Memory — VRAM, memory type (HBM2e / HBM3 / HBM3e, GDDR, LPDDR) and bandwidth in GB/s, the number that bounds decode speed.
- Compute — dense tensor TFLOPS for BF16, FP8, FP4 and INT8 (no 2:4 sparsity; "—" where the hardware has no native path), SM count and SM version.
- Scale-out — NVLink support, bandwidth and the largest NVLink domain, intra-node bandwidth, the NCCL workspace per tensor-parallel size, MIG and vGPU support and form factor (SXM, PCIe, OAM, rack-scale).
- Power and lifecycle — TDP, BF16 TFLOPS per watt, release year, tier (datacenter, prosumer, consumer, edge) and an end-of-life flag.
Specs come from vendor datasheets and are verified per row; a test suite checks every catalog change against cross-field invariants — NVLink against form factor, TFLOPS ratios between precisions, MIG and PCIe rules, unique names. Jetson rows add power modes and the memory reserved by the OS; they carry no prices.
Finding a card
Search matches the name and highlights it. Quick presets narrow the list in one click: Production datacenter (live, NVLink datacenter cards), Latest Blackwell, LLM inference ready (40 GB or more with FP8 or FP4), Home lab (consumer cards up to 24 GB and 450 W), Rack-scale and Edge / Jetson. Filters cover vendor, architecture, memory size and type, form factor and tier; an EOL switch hides end-of-life rows, which are otherwise dimmed. Sort by any numeric column, high to low or low to high.
The Columns panel chooses what the table shows — VRAM, bandwidth, BF16 / FP8 / FP4 TFLOPS, TDP, TFLOPS per watt and year by default; memory type, INT8, SMs, SM version, NVLink, MIG, vGPU and form factor on request. On a phone the table becomes cards, with filters and sort in sheets. Recently viewed GPUs appear above the list; they are kept in your browser only.
The GPU page
Each row opens /gpus/<slug>: key specs, the full spec groups, the current rental price and market position, links to its price page and comparisons, Models that fit — the catalog models that run on this card at a context and precision you set — and Similar GPUs, ranked by architecture, tier, VRAM and BF16 compute, with end-of-life cards last. Jetson pages show their power modes.
AMD GPUs
AMD Instinct cards (MI250X, MI300X, MI325X, MI350X, MI355X) are modeled with the same byte budget and roofline as NVIDIA. Three differences: the throughput calibration set is NVIDIA-heavy, so AMD figures carry wider uncertainty; multi-GPU units over xGMI are modeled at NVLink-class efficiency without a measured point yet; and precision support differs — FP8 is native on MI300-class cards and emulated on the MI250X, NVFP4 is NVIDIA-only, and TensorRT-LLM and Dynamo do not run on AMD. Fine-tuning recommendations cover NVIDIA cards only.
From the catalog
Tick up to four rows to compare them side by side, open What fits my GPU for a card you already have, or check live rental prices.