Apple M3 Ultra vs NVIDIA GeForce RTX 5090 — Specs & Performance Compared

Verdict: On llama3.3:70b, the NVIDIA GeForce RTX 5090 generates 27.0 tokens/s (per Local AI Master) against 16.0 for the Apple M3 Ultra (per LLMCheck), best community run on file for each.

Side-by-side comparison of the Apple M3 Ultra and the NVIDIA GeForce RTX 5090. Both released in 2025. Specs, synthetic benchmarks, AI inference tok/s, and Amazon availability below; each column header links to that part's benchmark page.

Apple · M3

Apple M3 Ultra

2025

Full Apple M3 Ultra benchmarks →

NVIDIA · Blackwell

NVIDIA GeForce RTX 5090

32 GB VRAM575W TDP$1,999 MSRP2025

Full NVIDIA GeForce RTX 5090 benchmarks →

*Price sourced from Amazon.com. Price and availability subject to change.

Where to buy

A listing for each part; the current price and stock are on the retailer's page.

Specs side-by-side

Spec	Apple M3 Ultra	NVIDIA GeForce RTX 5090
vram gb	—	32
vram type	—	GDDR7
tdp w	—	575
cuda cores	—	21760

Synthetic benchmarks

Each row says whether a higher or a lower number is better (ranks and times are lower-is-better). Scores come from TechPowerUp, 3DMark, Overclocking.com, Tom's Hardware, Notebookcheck, GadgetVersus, MacRumors, AppleInsider, Geekbench Browser, GSMArena and Hostbor; each row links its sources, and where they differ the cell shows their median. Empty cells mean we don't yet have that benchmark on file for the SKU.

Benchmark	Apple M3 Ultra	NVIDIA GeForce RTX 5090
3DMark Port Royal (higher is better) Source: TechPowerUp, 3DMark	—	41,238 points
3DMark Speed Way (higher is better) Source: Overclocking.com, 3DMark	—	15,488 points
3DMark Steel Nomad (higher is better) Source: 3DMark	—	11,906 points
3DMark Steel Nomad (4K) (higher is better) Source: Tom's Hardware	—	14,133 points
3DMark Time Spy (higher is better) Source: TechPowerUp, 3DMark	—	39,365 pts
Cinebench 2024 multi (higher is better) Source: Notebookcheck	3,000 points	—
Cinebench 2024 Multi-Core (higher is better) Source: Notebookcheck	3,000 points	—
Cinebench R23 multi (higher is better) Source: GadgetVersus	46,243 points	—
Cinebench R23 single (higher is better) Source: GadgetVersus	2,011 points	—
Geekbench 6 Metal (GPU) (higher is better) Source: MacRumors	259,668 points	—
Geekbench 6 multi (higher is better) Source: AppleInsider, MacRumors, Geekbench Browser	27,749 points	—
Geekbench 6 Multi-Core (higher is better) Source: Geekbench Browser, GSMArena, Hostbor	27,759 points	—

AI inference (tokens/sec)

Top community-reported tokens-per-second results for popular LLMs whose weights fit the card's memory, each linked to its source. See the per-SKU benchmark page for the full set including quantization and prompt length.

Apple M3 Ultra

llama3.1:8b: 128.2 tok/s (Creative Strategies)
gpt-oss-20b: 115.5 tok/s (llama.cpp GitHub (ggml-org guide: running gpt-oss with llama.cpp))
deepseekr1:32b: 100.0 tok/s (LocalLLaMA)

All Apple M3 Ultra AI benchmarks →

NVIDIA GeForce RTX 5090

Qwen2.5-Coder-7B-Instruct: 5841.0 tok/s (Runpod)
Llama 3.1 8B FP8: 420.0 tok/s (Runpod)
gpt-oss:20b: 298.2 tok/s (Hardware Corner)

All NVIDIA GeForce RTX 5090 AI benchmarks →

Quick answers

Which is better for running local LLMs?

On llama3.3:70b, the NVIDIA GeForce RTX 5090 generates 27.0 tokens/s (per Local AI Master) against 16.0 for the Apple M3 Ultra (per LLMCheck), best community run on file for each.