Apple M5 Macs beat DGX Spark in AI benchmarks, rival RTX 6000 value

A detailed three-day benchmark battle between M5 Macs, DGX Spark, Strix Halo, and an RTX 6000 has settled some heated online debates. The tests, run by Redditor Signal_Ad657 in controlled conditions with good power and cooling, reveal that memory bandwidth is the primary driver of tokens-per-second performance. The RTX 6000 leads with ~1,800 GB/s, followed by the M5 at ~600 GB/s, while the DGX Spark and Strix Halo both sit around 256 GB/s. Token throughput closely follows that curve, making the M5 a genuine contender for local AI workloads—especially for its price point.

The biggest surprise was the M5 MacBook Pro's thermal performance. Despite running continuously for days, it stayed in the 80°C range—better than expected. However, the trade-off is noise: under sustained AI loads, the fans ramp up to loud, blow-dryer levels, contradicting the 'quiet' marketing claims. The EVO X2 thermals caused issues during extended runs, while the MacBook held up admirably. The dataset includes raw numbers for memory bandwidth, tokens/sec, and temperature logs, now publicly available in the MMBT repo. Future updates will add MLX on Mac and various backends for Strix Halo to explore software-level impacts.