M5 Ultra 256GB Benchmarks: 27B LLM 51 tok/s, GPU +82% in Cyberpunk 4K RT

The 256GB unified memory M5 Ultra Mac Studio finally has real-device data. This article compiles all raw numbers from a hands-on test video, comparing each item against the M3 Ultra — conclusion first: CPU and GPU across the board, local LLM single-stream decoding speed up nearly 50%.

Test Methodology: Clarifying the Scope

This machine tests Qwen3.8-27B-4bit , with weights of 15.7GB , running on Mac Studio. Two key methodological points must be stated upfront to avoid misinterpretation:

MTP disabled : Inference uses LM-only baseline (speculative decoding MTP turned off), so the reported decoding speeds represent a conservative lower bound , not the maximum achievable in daily use.

ANE alignment enabled : Test items enable ANE alignment, causing pp1024/pp4096 to appear as pp1025/pp4097 in the tables (an extra token to align prefill to 2048-block boundaries).

In short: these are conservative "capped" numbers; real-world daily usage will be faster.

Table 1: Qwen3.8-27B-4bit local inference, M5 Ultra vs M3 Ultra (MTP disabled conservative baseline)

1. AI Local Inference: Single-Stream Decoding +45.7%, Multi-Concurrency Prefill Up 3.6x

First, the most watched metric — how fast can it run large models locally .

Single-stream decoding speed is the most direct daily experience indicator: M5 Ultra reaches 51 tok/s , M3 Ultra 35 tok/s , a +45.7% improvement. Converting to effective bandwidth, the 48.4 tok/s tier corresponds to ~760 GB/s, while 51 tok/s reaches ~800 GB/s — essentially hitting the practical ceiling of this generation's unified memory bandwidth. Single-stream speed has plateaued; further gains rely on MTP and continuous batching.

Multi-concurrency is where this machine truly shines:

Aggregate decoding: 2 concurrent +46.3%, 4 concurrent +30.3%, but at 8 concurrent M5 (248.9) slightly trails M3 (257.4) by −3.3% — the only test where M5 falls behind.

Prefill : 2 concurrent 1243.3 vs 345.5 ( 3.6× ), 8 concurrent 1016.1 vs 282.6 ( 3.6× ) — massive advantage when multiple users query simultaneously.

8-concurrent end-to-end latency (pp1024/tg512): M5 12.176s vs M3 32.971s, 2.8× faster .

Conclusion: running a single large model, M5 is ~50% faster than M3; under multi-user/multi-task concurrency, M5's lead widens further — only in extreme 8-stream high-concurrency decoding do the two essentially tie.

2. CPU: Single-Core +25%, Multi-Core +43.6%

Table 2: CPU, GPU synthetic benchmarks and Cyberpunk 2077 real-world (vs M3 Ultra)

CPU tested with Geekbench 7 and Cinebench 2026:

Geekbench 7 single-core 3,711 (M3 2,963, +25.2% ), multi-core 51,760 (M3 40,271, +28.5% ).

Cinebench 2026 single-core 744 (M3 676, +10.1% ), multi-core 17,348 (M3 12,082, +43.6% ).

Single-core gains are modest, multi-core gains pronounced — consistent with the large-chip strategy of adding more cores.

3. GPU: Overall +34% to +68%, Graphics Performance Nears Desktop Class

GPU shows the largest generational leap:

Geekbench 7 GPU: 352,444 (M3 263,242, +33.9% ).

Cinebench 2026 GPU: 140,734 (M3 83,865, +67.8% ), now surpassing the RTX 5080 (100k-class) in this benchmark.

Blender 5.2.1 LTS (samples/minute, higher better): Junkshop scene 3,795 (M3 2,048, +85.3% ), Lion scene 4,806 (M3 2,861, +68% ) — nearly double the performance.

3DMark Steel Nomad: 8,052 (M3 5,548, +45.1% ); 3DMark Solar Bay Extreme: 26,105 (M3 16,426, +58.9% ).

In absolute terms, M5 Ultra's GPU scores exceed most desktop cards beyond the RTX 5080 (Steel Nomad 8,800) and RTX 5070 Ti (Solar Bay 27,915), representing a tier-breaking lead within the Mac lineup.

4. Real-World Gaming: Cyberpunk 2077 with Ray Tracing +82.8%

Ray tracing has been Apple Silicon's weak spot, but the gap is visibly narrowing:

At 4K resolution · Ultra Ray Tracing preset · MetalFX Quality mode , Cyberpunk 2077 average frame rates:

Ray tracing on: M5 37.28 FPS (M3 20.39, +82.8% ).

Ray tracing off: M5 60.32 FPS (M3 47.78, +26.2% ).

The ray-tracing improvement (+82.8%) is the second-highest single-item gain, indicating major generational GPU improvements in the ray-tracing pipeline. However, 37 FPS remains "playable but not smooth"; 4K ray-traced gaming is still not Apple Silicon's forte.

5. Gain Breakdown: Which Areas Improved Most?

Table 3: M5 Ultra vs M3 Ultra gain ranking (red indicates sole negative item)

Sorting the 14 test items by gain reveals a clear pattern:

First tier (+60% and above) : Blender Junkshop +85.3%, Cyberpunk RT on +82.8%, Blender Lion +68%, Cinebench GPU +67.8%, Solar Bay +58.9% — all graphics/rendering/ray-tracing related.

Second tier (+30% to +46%) : Multi-concurrent decoding, 3DMark, CPU multi-core, GPU overall.

Third tier (~+25%) : Geekbench CPU single/multi-core, Cyberpunk RT off.

Sole negative item : 8-concurrent aggregate decoding −3.3% .

In one sentence: M5 Ultra concentrates its gains on GPU and concurrent throughput , CPU improves steadily, extreme high-concurrency decoding roughly matches previous generation — fully consistent with the "larger die + more cores + stronger GPU" generational positioning.

6. Methodology Notes and Buying Advice

Several methodological caveats:

Inference data is MTP-disabled conservative baseline ; enabling MTP/continuous batching yields higher speeds.

Comparisons must account for identical engine settings — both MTP off, both LM-only baseline — otherwise architectural differences conflate with configuration differences.

For multi-concurrency, examine both "aggregate tok/s" and "per-stream tok/s" to avoid masking per-stream slowdown at 8-way behind aggregate throughput.

Only Qwen3.8-27B-4bit was tested; the community awaits 70B/200B head-to-heads, which are not yet available.

Buying advice :

If you need "single-machine local LLM + occasional video editing/rendering" , the M5 Ultra 256GB's graphics and concurrency gains are compelling , especially for multi-user shared inference and local agent orchestration.

If you only run single chat sessions, code, or edit 4K video — M4 Max (64–128GB) is already ample . The 256GB version costs nearly 7× the entry model; the value of paying for memory and concurrency only exists when you actually saturate them .

Final Thoughts

M5 Ultra 256GB proves with real benchmarks that it is not merely "more memory" but a generational upgrade across CPU, GPU, and local inference — especially in graphics rendering and concurrent throughput where it nearly doubles performance. 256GB unified memory lets you fit 200B+ parameter models, and these benchmarks show it runs them far faster than the previous generation.