Skip to content
llm-speed

RTX 5090 vs M3 Ultra

A single shareable card for the RTX 5090 vs M3 Ultra matchup. Numbers are individual submitted decode results; each available side shows its source configuration. These are not repeated controlled comparisons.

RTX 5090 vs M3 Ultra

Individual submitted decode results; model and workload shown below.

NVIDIA Ada/Blackwell
356.1tok/s
decode (individual submitted result)
Model
stable-code-instruct-3b
Reported hardware
RTX 5090 (32GB) + AMD Ryzen 7 9850X3D 8-Core Processor (8c) + 3…
Workload / suite
concurrent-decode · suite-v1
Runtime
llama.cpp · version not reported
Quantization field
Not reported
Submitted
2026-07-01
source: r_q9f15lz6831
Apple M3
429.2tok/s
decode (individual submitted result)
Model
mlx-community/Qwen3-0.6B-4bit
Reported hardware
M3 Ultra (60-core GPU) + 96GB unified
Workload / suite
chat-short · suite-v1
Runtime
mlx · 0.31.3
Quantization field
Not reported
Submitted
2026-07-12
source: r_6rgao6ao735
These independently selected results can use different models, workloads and settings. They do not establish a hardware or model winner.

The memory shown in a source configuration belongs to that submission. For M3 Ultra purchases, Apple’s UK specification checked September 6, 2026 lists 96 GB configurable to 256 GB. Its 2025 launch announcement described up to 512 GB. Confirm the exact configuration; a family maximum is not the machine measured here.

Need the long-form table? Open the RTX 5090 vs M3 Ultra comparison for every overlapping (model × hardware) row, source runs, and methodology.