hao-ai-lab-FastVideo/docs/cookbook/capacity/hardware-coverage.md

Hardware coverage checklist

This checklist tracks the hardware lookup catalog, not FastVideo runtime or kernel support. A searchable device can still be unsupported by a model, backend, or kernel. Capacity is never inferred from a marketing family name.

Batch 1 is the reviewable foundation in hardware.yaml: 52 variants across 40 families — 31 NVIDIA, 6 AMD, 3 Intel, and 12 Apple entries; 38 dedicated-memory devices and 14 unified-memory systems. Every entry has a vendor source URL and a checked date.

Batch 1 — shipped in this change

Priority order: devices named by existing recipes/baselines and the previous picker, then common consumer, workstation, and data-center variants.

Priority	Family	Variants	Memory	Why now
1	NVIDIA H100	SXM 80 GB, NVL 94 GB	Dedicated	Recipe/baseline anchor; SXM MIG profiles
1	NVIDIA H200	SXM 141 GB	Dedicated	Common data-center target
1	NVIDIA B200	180 GB	Dedicated	Previous picker
1	NVIDIA GB200	Blackwell GPU 186 GB	Dedicated	Previous picker; per-GPU, not pooled
1	NVIDIA A100	SXM 80 GB	Dedicated	Common data-center target; MIG profiles
1	NVIDIA L40S / L4	48 GB / 24 GB	Dedicated	Common serving cards
1	NVIDIA RTX 5090	32 GB	Dedicated	FastH3 consumer baseline
1	NVIDIA RTX PRO 6000 Blackwell	Workstation 96 GB	Dedicated	FastH3 consumer baseline
1	NVIDIA RTX 4090	24 GB	Dedicated	FastH3 FP8 baseline
1	NVIDIA DGX Spark GB10	128 GB, 64 GB OEM	Unified	FastH3 Spark baseline
1	Apple M4 Max	36, 48, 64, 128 GB	Unified	FastMetal tested 36 GB recipe
1	Apple M5 Max / M5 Ultra	48/64/128 GB; 96/256/512 GB	Unified	Current Mac Studio/MacBook Pro options
2	Apple M1 Pro / M2	16 GB / 8 GB	Unified	M1 Pro FastMetal evidence; low-memory Apple systems
2	NVIDIA RTX 50 series	5080 16, 5070 Ti 16, 5070 12, 5060 Ti 16/8	Dedicated	Common consumer cards
2	NVIDIA RTX 40 series	4080 Super 16, 4070 Ti Super 16, 4060 Ti 16/8	Dedicated	Common consumer cards
2	NVIDIA RTX 30 series	3090 24, 3080 12/10	Dedicated	Widely installed
2	NVIDIA RTX 6000 Ada / A6000 / A5000 / A4000	48 / 48 / 24 / 16 GB	Dedicated	Workstations
2	NVIDIA RTX 5000 Ada / 4000 Ada	32 / 20 GB	Dedicated	Workstations
2	AMD Instinct MI300X / MI325X / MI355X	192 / 256 / 288 GB	Dedicated	Data-center alternatives
2	AMD Radeon RX 7900 XTX / RX 9070 XT	24 / 16 GB	Dedicated	Consumer alternatives
2	AMD Radeon PRO W7900	48 GB	Dedicated	Workstation alternative
2	Intel Arc A770 / B580	16 / 12 GB	Dedicated	Consumer alternatives
2	Intel Data Center GPU Max 1550	128 GB	Dedicated	Data-center alternative

Batch 2 — next reviewable additions

These are deliberately not asserted in the catalog yet. Add them only with a vendor source and checked date.

• NVIDIA H100 PCIe 80 GB, H200 NVL, B300/GB300 variants.
• NVIDIA A100 40 GB, A40, A10, A30, T4, V100 16/32 GB.
• NVIDIA laptop GPUs, separated by laptop power/board variant and VRAM.
• NVIDIA RTX 5070 laptop, 5060/5050 desktop, 4070/4060 desktop variants.
• NVIDIA RTX PRO 6000 Max-Q and Server Edition (separate from Workstation).
• NVIDIA DGX Station / OEM Grace-Blackwell systems.
• AMD Instinct MI250X, MI210, MI300A; Radeon RX 9060 XT, W6800.
• Intel Arc B570, Flex 170/140.
• Apple M3 Max/Ultra, M2 Max/Ultra, M4 Pro, and remaining older configurable sizes.
• Additional MIG/vGPU profiles where a vendor documents the partition size.
• Systems with a documented host-RAM configuration for offload estimates.

Validation rules

The build fails on duplicate IDs, names, aliases, or family variants; duplicate or malformed sources; missing checked dates; unknown units, vendors, kinds, backends, or dtypes; usable budgets above nominal capacity; partitions above the full device; and unified entries that hide capacity in a generic memory_gb field.

• memory.nominal and memory.usable keep the authored value/unit (GB or GiB); the build also exports normalized decimal memory_gb and usable_memory_gb.
• Dedicated records expose VRAM. Unified records expose shared system memory and never expose it as VRAM; the estimator uses a labeled 75% application default and accepts an explicit memoryBudgetGiB override.
• partitions are separate selectable allocations. Selecting one never silently inherits the full-card capacity; a vendor-documented profile may explicitly span the full memory. If no partition is selected, the full device is used.
• aliases preserve old UI links and search names; canonical IDs remain stable.
• sources[].checked records when the vendor page was verified, not when the device was released.

Deliberately out of scope

• Claiming kernel, dtype, backend, or model support from catalog presence.
• Browser-side detection of the remote generation host. Lookup is user input.
• Summing per-GPU capacity into a pool. GPU count and topology are separate.
• Inventing usable-memory reserves for macOS, Windows, or OEM systems. A budget must come from a documented policy or the user.
• Treating a successful generation as a measured minimum-VRAM result.