Comparison of consumer and prosumer compute options for running/training ML models locally: Apple’s MacBook M4 Pro, Nvidia’s RTX 3060 (12GB), RTX 3090 (24GB), and RTX 5090 (32GB), and Nvidia’s DGX Spark desktop AI computer. Figures below are approximate and drawn from vendor specs and community benchmarks as of 2026; used-market prices in particular fluctuate.
Comparison table
| MacBook M4 Pro | RTX 3060 (12GB) | RTX 3090 (24GB) | RTX 5090 (32GB) | DGX Spark | |
|---|---|---|---|---|---|
| Price (€) | ~2,200–3,000 (whole laptop) | ~300–350 (new) | ~800–1,300 (used only) | ~2,300–2,800 | ~3,800–4,200 |
| VRAM / memory | 24/48GB unified | 12GB GDDR6 | 24GB GDDR6X | 32GB GDDR7 | 128GB unified LPDDR5x |
| Memory bandwidth | 273 GB/s | 360 GB/s | 936 GB/s | ~1,792 GB/s | 273 GB/s |
| FP32 TFLOPS (dense) | ~9.5 (GPU) | ~12.7 | ~35.6 | ~105 | ~31 (dense FP32 equivalent, approx.) |
| Tensor / AI throughput | ~38 TOPS (Neural Engine, INT8) | ~101 TFLOPS FP16 (tensor) | ~285 TFLOPS FP16 (tensor) | ~1,676 TFLOPS FP16 (with sparsity) | ~1,000 TFLOPS FP4 (sparse, marketing figure) |
| TDP / power | ~40–60W (whole system) | 170W | 350W | 575W | ~140W (whole unit) |
| Memory type | Shared with CPU (unified) | Dedicated VRAM | Dedicated VRAM | Dedicated VRAM | Shared with CPU (unified) |
Notes
- M4 Pro and DGX Spark share an architecture class: both are unified-memory SoCs rather than discrete GPUs with dedicated VRAM. This is why they post the same 273 GB/s bandwidth figure — it’s the SoC’s shared memory bus, not a dedicated GPU memory controller. The DGX Spark is often criticized for this: it has huge memory capacity (128GB) but bandwidth on par with a laptop chip, so it is compute/bandwidth-starved for large-batch work despite being marketed for LLM inference.
- Bandwidth, not FLOPs, is usually the bottleneck for LLM inference. Autoregressive token-by-token generation is memory-bound, which is why the RTX 3090 (936 GB/s) often outperforms cards with higher FLOPs but lower bandwidth for local LLM serving, and why the DGX Spark’s 273 GB/s is a real limitation despite its capacity advantage.
- VRAM capacity determines what model sizes fit at all, independent of speed. This is where the DGX Spark (128GB) and RTX 5090 (32GB) win over the 3060/3090 — they can hold larger models (or longer context / bigger batches) even if not always fastest per-token.
- The RTX 3090 is EOL and used-market only now; prices vary a lot by region and condition.
- Price is not apples-to-apples: the Mac and DGX Spark prices are for a complete compute unit (CPU + RAM + storage + GPU/NPU), while the RTX card prices are for a component only — a full PC build (CPU, RAM, PSU, case) adds several hundred euros on top before it’s a usable system.