[Tasks] Add computeCapability to NVIDIA GPUs in hardware.ts

[julien-c]

Summary

• Adds an optional computeCapability field to the HardwareSpec interface representing the CUDA Compute Capability version number
• Populates computeCapability for all 105 NVIDIA GPU entries, covering architectures from Maxwell (5.3) through Blackwell (12.0/12.1)
• Enables downstream consumers to filter/group GPUs by CUDA compatibility level

Compute Capability Mappings

CC	Architecture	Example GPUs
12.0	Blackwell (consumer)	RTX 5090, 5080, 5070, RTX PRO 6000
12.1	Blackwell (edge)	GB10
10.0	Blackwell (DC)	B200
9.0	Hopper	H200, H100
8.9	Ada Lovelace	L40s, L4, RTX 40-series
8.7	Orin	Jetson Orin variants
8.6	Ampere (consumer)	RTX A-series, RTX 30-series
8.0	Ampere (DC)	A100, A30
7.5	Turing	RTX 20-series, T4
7.2	Xavier	Jetson Xavier variants
7.0	Volta	V100
6.2	Parker	Jetson TX2
6.1	Pascal (consumer)	GTX 10-series, P40
6.0	Pascal (DC)	P100
5.3	Maxwell (Tegra)	Jetson Nano

Test plan

• npx tsc --noEmit passes with no errors
• Verified all 105 NVIDIA GPU entries have a computeCapability value

---

Note

Low Risk
Low risk, additive change: extends HardwareSpec with an optional field and populates static NVIDIA SKU data; main risk is minor downstream assumptions about SKU shape or increased bundle size.

Overview
Adds an optional computeCapability field to HardwareSpec in hardware.ts to capture NVIDIA CUDA compute capability.

Populates computeCapability across the NVIDIA GPU entries in SKUS, enabling consumers to filter/group hardware by CUDA compatibility without changing existing required fields.

^{Written by Cursor Bugbot for commit 4e1c621. This will update automatically on new commits. Configure here.}

[julien-c]

Stacked PR: #2011 adds gfxVersion (the AMD equivalent) to all AMD GPU entries.

[SBrandeis]

Why not a string btw?

[julien-c]

i would say simpler to do version comparisons that way (also it's used as an integer in the kernels package)

[julien-c]

but why: do you have a preference for strings?

[mfuntowicz]

It's quite usual to store the nvidia compute capabilities as major * 10 + minor so it becomes easy to do comparison. It also works seemlessly for AMDGPU gfx<compute_capability:integer> (ex: gfx1150).

[mfuntowicz]

LGTM - If there is a path where you'd like to display some more "informative" info on the Hub UI, maybe an enum for each provider would fit best:

enum NvidiaComputeCapabilities {
    BLACKWELL_RTX = 12.0,
    BLACKWELL = 10.0,
    HOPPER = 9.0,
    AMPERE_RTX = 8.6,
    AMPERE = 8.0
    ... 
}

Can make it easier for a reader to map these numbers to actual literals used by NVIDIA on their comm

[pcuenca]

LGTM - If there is a path where you'd like to display some more "informative" info on the Hub UI, maybe an enum for each provider would fit best:

enum NvidiaComputeCapabilities {
    BLACKWELL_RTX = 12.0,
    BLACKWELL = 10.0,
    HOPPER = 9.0,
    AMPERE_RTX = 8.6,
    AMPERE = 8.0
    ... 
}

Can make it easier for a reader to map these numbers to actual literals used by NVIDIA on their comm

Or even a comment could suffice for now, as a reference for contributors.

[pcuenca]

Amazing, thanks for taking the time to dive into this!