From morning until now, based on the patterns I use, future tasks, and the expectation that memory prices won't drop for the next three years, I've put together a performance table.
Conclusion: It's cheap.
Conclusion 2: Cancel the pro subscription.
Conclusion 3: I'm going to buy a Mac mini for home…
M5 Ultra 256GB vs RTX 5090 + RTX PRO 6000 — Local AI Benchmark
Item | Mac Studio M5 Ultra 256GB | RTX 5090 + RTX PRO 6000 PC |
|---|
CPU | Up to 36 cores | Ryzen 9950X3D 16C/32T |
GPU | Up to 80 cores GPU + Neural Accelerator per core | 21,760 + 24,064 CUDA Cores |
AI Processing | Apple: Up to 4.3x faster than M3 Ultra | 3,352 + 4,000 = Theoretically 7,352 AI TOPS* |
Memory | 256GB Unified Memory | 32GB + 96GB = 128GB VRAM |
Memory Architecture | Single unified memory pool | VRAM of two GPUs is separate |
Memory Bandwidth | 1.2TB/s | Each GPU 1.792TB/s |
GPU Communication | UltraFusion within the chip | GPU communication via PCIe |
CUDA | X | O |
Power/Noise | Very favorable | Approximately 1.2kW for RTX 5090 + PRO 6000 alone |
Domestic Price | M5 Ultra starting price 9,490,000 won + 256GB option | Approximately 36,000,000~44,000,000 won |
Maximum Model Size | 200~400B possible in Q4 | Practically limited to 96GB single GPU / 128GB distributed |
* This is a theoretical value obtained by simply summing NVIDIA's AI TOPS, and it does not mean that the two GPUs will achieve 7,352 TOPS together as a single GPU.
Expected LLM Inference Performance
Model / Condition | M5 Ultra 256GB Estimated | RTX PRO 6000 Center PC |
|---|
70B Dense Q4 | About 20~25 tok/s | About 30~45 tok/s |
GPT-OSS 120B MXFP4 | About 100~120 tok/s | About 160~190 tok/s |
Qwen3.5 397B-A17B Q4 | About 55~60 tok/s | Memory Insufficient |
Qwen 397B Q4 / 64K context | About 38~42 tok/s | Memory Insufficient |
Qwen 397B Q4 / 128K context | About 28~32 tok/s | Memory Insufficient |
M5 Ultra figures are still estimated before official release, based on M3 Ultra 256GB's actual MLX/oMLX benchmarks and
M5 Ultra's 1.2TB/s memory bandwidth, 80-core GPU, Neural Accelerator performance improvements.
Note that Qwen3.5 397B-A17B Q4 actually runs on M3 Ultra 256GB:
measured.
M5 Ultra has about 1.47 times (819GB/s → 1.2TB/s) increased memory bandwidth compared to M3 Ultra,
and according to Apple's announcement, LLM prompt processing is improved by up to about 4 times and AI compute by up to about 4.3 times.
Conclusion
RTX 5090 + RTX PRO 6000
Advantages
Absolute speed of models under 120B is definitely fast
CUDA / vLLM / SGLang / TensorRT ecosystem
Strong in learning, LoRA, image generation, parallel serving
Very powerful if different models are loaded on RTX 5090 and PRO 6000
Disadvantages
Even with two GPUs combined, VRAM is not a single 128GB pool
Memory becomes the biggest bottleneck from 200~400B models
Price about 40 million won
Power consumption and heat generation are very large
M5 Ultra 256GB
Advantages
256GB is practically used as a single memory pool by the GPU
Qwen 397B models can be run on a single machine
1.2TB/s memory bandwidth
GPU Neural Accelerator further enhances AI operations compared to previous Ultra
Very good large LLM computing capability per price/power/noise
Disadvantages
If the same model fits inside NVIDIA VRAM, then NVIDIA is faster
CUDA/SGLang cannot be used as is
Apple Silicon runtimes such as MLX/llama.cpp are required
One-line Conclusion
If the goal is to run 70~120B models at maximum speed, then NVIDIA.
If the goal is large models of 200~400B scale, long context, or local AI agents, then the M5 Ultra 256GB is much more reasonable.
Ultimately
NVIDIA = You are buying computational performance and speed
M5 Ultra = You are buying a large pool of high-speed memory
As the model size increases in local LLMs, the value of the latter grows rapidly.
