Unverified Performance Data Emerges for Apple M5 Ultra

Preliminary performance metrics for Apple’s anticipated M5 Ultra chip have surfaced via unverified reports, sparking discussion within technical communities. According to data shared on Reddit, benchmarks hosted on the OMLX website suggest specific performance capabilities for the chip when running large language models (LLMs), though these figures lack official confirmation from Apple Inc.

LLM Benchmark Specifications

The reported data focuses on the performance of the Qwen 3.8 27B model (q4 quantization). According to the unverified OMLX metrics, the M5 Ultra achieved a throughput of 50 tokens per second (tok/s) and a prefill (pp) rate of 1,800 tok/s at an 8k context window, notably without the use of Multi-Token Prediction (MTP).

While these figures suggest a level of capability for local AI processing, technical commentators on Reddit have raised questions regarding the efficiency of these results. Some users noted that the reported speeds appear low when compared to high-end discrete GPU configurations. For instance, one commentator suggested that a setup utilizing four NVIDIA RTX 3090 GPUs could achieve between 40 and 80 tok/s on Q8 quantization when utilizing MTP.

Comparative Hardware Performance

The emergence of these benchmarks has prompted a comparison between Apple’s silicon and competing hardware architectures. Some users argued that Apple’s current trajectory may face challenges from high-performance alternatives, such as the Strix Halo, which some claim offers superior performance-to-power efficiency compared to the M3 Ultra.

Further comparisons were drawn between the unverified M5 Ultra data and existing high-end hardware. Users noted that a dual NVIDIA RTX 4090 configuration could theoretically reach over 4,700 pp at 70 tok/s, while a single RTX 5090 could achieve 8,600 pp at 95 tok/s when utilizing NVFP4. Additionally, a user cited a 2019 Mac Pro equipped with a 16-core Xeon and an RTX 6000 Max-Q as achieving 10.9 tok/s prefill at 64k context and 137.2 tok/s generation.

Market Context and Data Availability

At the time of reporting, some users attempting to verify the OMLX data reported receiving “No matching results” messages, suggesting that the data may be transient or subject to change.

The discussion regarding Apple’s hardware performance occurs alongside broader market speculation regarding the company’s supply chain and pricing strategies. While the M5 Ultra benchmarks remain unconfirmed, they represent the latest in a series of speculative technical reports circulating within developer and enthusiast forums.