Apple Beats NVIDIA and AMD on Local AI Speed — But That's Not the Whole Story

Apple's Mac Studio generates tokens two to three times faster than AMD's Ryzen AI Max+ 395 machines in independent testing, despite AMD marketing its chip as the brains of "the world's smallest AI development PC." That's the sharpest data point in a three-way local AI hardware race where the fastest box and the safest software bet turn out to be two different products.

LTT Labs, the independent testing outfit that ran the benchmarks, found the gap in plain terms: "the Apple Silicon Mac Studios outperform the AMD Ryzen AI Max+ 395 machines," producing tokens at two to three times the rate. For a developer iterating hundreds of times a day, that difference compounds fast — a workload that takes an hour on Apple silicon can take two or three on AMD.

Here's what each machine actually changes for the practitioner.

What does NVIDIA's DGX Spark get you?

The NVIDIA DGX Spark pairs 128GB of unified memory with a Grace Blackwell GB10 chip built around CUDA, which remains the deepest, most battle-tested software ecosystem in AI. Existing PyTorch code, TensorRT and cuDNN pipelines run on it without modification — something you can't say about the other two options yet.

The drawback is memory bandwidth. At 273 GB/s, the Spark is the slowest of the three when a model has to move a lot of data, and reviewers consistently find it underperforms Apple's Mac Studio on raw token generation despite the mature software. That's less than half the M5 Max's 614 GB/s and roughly a quarter of the M5 Ultra's 1.2 TB/s.

The fit: teams already deep in CUDA and TensorRT who want to prototype locally before deploying to NVIDIA cloud infrastructure without touching their code. Think quantized versions of Llama, Qwen and DeepSeek in the 70B to 200B range, where TensorRT-LLM optimization matters more than raw bandwidth.

Why does AMD sell one chip at two very different prices?

This is where things get confusing on purpose. AMD's Ryzen AI Max+ 395, marketed as capable of running 200-billion-parameter models locally, ships in two completely different products. Third-party mini PCs built around the same chip — the GMKtec EVO-X2 and the Framework Desktop — sell for a fraction of AMD's own official Ryzen AI Halo Developer Platform. Same processor, same 128GB memory pool, more than double the price on the official kit.

LTT Labs explains the gap plainly: the premium buys AMD's curated software configurations, "AI Playbooks," and first-party support, not extra hardware. Take the third-party route if you're comfortable managing software yourself and want the same silicon for less. Take AMD's kit for the support layer and the promise of continued driver and playbook maintenance.

There's an operating-system catch too. ROCm, AMD's answer to CUDA, is meaningfully more mature on Linux than Windows. Windows support runs through WSL and a newer compatibility layer that only recently expanded to more Ryzen hardware; Linux has had direct, better-supported access for longer. Budget-conscious developers doing training and fine-tuning experiments on Linux get the most memory for their money here — without needing NVIDIA-specific tooling.

AMD has leaned into the model side. The Ryzen AI Max+ 395 was the first consumer processor to run OpenAI's gpt-oss 120B, which needs around 61GB of memory at MXFP4 precision and runs at up to 30 tokens per second on the 128GB configuration.

When does Apple actually win?

The Mac Studio comes with the M5 Max and up to 128GB of unified memory, or the M5 Ultra with up to 512GB. The two chips aren't equal on bandwidth: the M5 Max reaches 614 GB/s, while the M5 Ultra doubles that to 1.2 TB/s. Even the lower tier beats NVIDIA and AMD at moving model data, which is why reviewers keep clocking the Mac Studio ahead of both rivals on local inference — even though Apple never markets these chips as AI accelerators first.

That 512GB ceiling is the real differentiator. It's enough to run a 70B model with zero quantization, a genuine advantage over the other two machines — though not enough for the largest open models.

The cost is the ecosystem. Your code needs to run through Apple's stack rather than CUDA or ROCm, which stings if your team's tooling assumes one of those. But you don't have to rewrite everything in MLX. Most people running models on a Mac load GGUF files directly through llama.cpp's Metal backend, Ollama or LM Studio, all of which work out of the box. MLX only matters if you're optimizing specifically for Apple Silicon.

So what does this change for buyers?

First, be clear about why you're buying a local box at all. It's not there to replace the cloud for large-scale training. Cloud GPUs still win for large training runs and burst capacity. A local box wins for three things:

• Keeping sensitive data off someone else's servers
• Avoiding metered API costs when you're iterating hundreds of times a day
• Getting a fixed, predictable machine you fully control instead of a shared, rate-limited one

Then match the ecosystem to your workflow before you match it to a benchmark. If your code already runs on CUDA, NVIDIA removes the most risk. If you want the most memory for the money and you're comfortable on Linux, AMD — official kit or third-party board — gets you there cheaper. If you want the fastest actual inference and don't mind stepping outside CUDA and ROCm entirely, Apple quietly wins the benchmarks.

None of these is a universal answer, and all three vendors are shipping new local AI hardware on a near-constant cycle. Pick the ecosystem your workflow already depends on first. The hardware comparison only matters after that.