Generated March 4, 2026
Sean's three machines compared head-to-head. This report covers specs, benchmarks, real-world performance, and whether upgrading to the M5 Max MacBook Pro makes sense.
| Spec | 16" MacBook Pro M5 Max | Mac Studio M3 Ultra | 15" MacBook Air M2 |
|---|---|---|---|
| Chip | Apple M5 Max | Apple M3 Ultra | Apple M2 |
| CPU Cores | 18 (6 super + 12 performance) | 28 (20 performance + 8 efficiency) | 8 (4 performance + 4 efficiency) |
| GPU Cores | 40 | 60 | 10 |
| Neural Engine | 16-core | 32-core | 16-core |
| Unified Memory | 128 GB | 96 GB | 24 GB |
| Memory Bandwidth | 614 GB/s | 819 GB/s | 100 GB/s |
| SSD | 2 TB | 1 TB | 2 TB |
| Display | 16.2" Liquid Retina XDR, 3456x2234, 120Hz ProMotion | None (desktop) | 15.3" Liquid Retina, 2880x1864, 60Hz |
| External Displays | Up to 4 | Up to 8 | Up to 1 |
| Ports | 3x Thunderbolt 5, HDMI, SDXC, MagSafe 3, 3.5mm | 4x Thunderbolt 5, 2x USB-C, 2x USB-A, HDMI, 10Gb Ethernet, SDXC, 3.5mm | 2x Thunderbolt/USB 4, MagSafe, 3.5mm |
| Wireless | Wi-Fi 7, Bluetooth 6, Thread | Wi-Fi 6E, Bluetooth 5.3 | Wi-Fi 6, Bluetooth 5.3 |
| Weight | 4.7 lbs (2.15 kg) | ~8.0 lbs (3.64 kg) | 3.3 lbs (1.51 kg) |
| Form Factor | Laptop | Desktop | Laptop |
| Year | 2026 | 2025 | 2023 |
| Benchmark | M5 Max MBP 16" | Mac Studio M3 Ultra | MacBook Air M2 |
|---|---|---|---|
| Single-Core | ~4,500 (ESTIMATE) | 3,213 (confirmed) | ~2,600 (confirmed) |
| Multi-Core | ~30,000 (ESTIMATE) | 26,530 (confirmed) | ~9,709 (confirmed) |
Sources and methodology:
- Mac Studio scores from Geekbench Browser (890 user submissions average) [browser.geekbench.com]
- MacBook Air M2 scores from Geekbench Browser Mac benchmarks (15" 2023 model) [browser.geekbench.com]
- M5 Max scores are estimates - no Geekbench results exist yet as of March 4, 2026
M5 Max estimation rationale:
- The M5 base chip (10 cores, 4.6 GHz) scores 17,464 multi-core on Geekbench 6
- The M4 Max (16 cores, 4.5 GHz) scores 25,702 multi-core and 4,028 single-core
- The M5 Max adds 2 more cores (18 total), all high-performance (no efficiency cores), with the new "super core" tier and improved IPC
- Single-core estimate: ~4,500 (roughly 10-12% IPC gain over M4 Max)
- Multi-core estimate: ~30,000 (18 cores vs 16 + IPC gains, ~15-17% over M4 Max)
The Mac Studio M3 Ultra still holds a strong multi-core advantage thanks to its 28 CPU cores. Even with the M5 Max's architectural improvements, the Ultra's raw core count likely keeps it ahead in heavily threaded workloads by roughly 10-15%. However, the M5 Max should have a significant single-core advantage (~40% over M3 Ultra), which matters for lightly threaded tasks, app launch times, and single-threaded compilation steps.
The MacBook Air M2 is roughly 3x slower in multi-core and 35-40% slower in single-core compared to the Mac Studio. It's a different class of machine entirely.
| Benchmark | M5 Max (40-core GPU) | Mac Studio M3 Ultra (60-core GPU) | MacBook Air M2 (10-core GPU) |
|---|---|---|---|
| Metal | ~200,000-220,000 (ESTIMATE) | 220,119 (confirmed) | ~30,000 (ESTIMATE) |
| OpenCL | ~110,000-120,000 (ESTIMATE) | 124,904 (confirmed) | ~18,000 (ESTIMATE) |
Sources:
- Mac Studio M3 Ultra Metal score: 220,119 (Geekbench Browser average) [browser.geekbench.com]
- M5 Max and M2 GPU estimates based on core count scaling and generational improvements
Analysis:
The M3 Ultra's 60-core GPU is significantly more powerful than the M5 Max's 40-core GPU in raw compute. Even with the M5 Max's 3rd-generation ray tracing engine and Dynamic Caching improvements, the 50% core count advantage of the M3 Ultra is hard to overcome. The Mac Studio likely retains a 10-20% GPU compute advantage.
The M4 Max (40-core GPU) in the Mac Studio 2025 scores 115,830 OpenCL. The M5 Max should improve on this by 10-15%, but won't match the M3 Ultra's 60 GPU cores.
For external display support: the Mac Studio supports up to 8 displays (M3 Ultra), the M5 Max MBP supports 4 (plus built-in), and the Air supports just 1.
This is where the comparison gets really interesting for LLM inference workloads.
| Spec | M5 Max | M3 Ultra | M2 |
|---|---|---|---|
| Neural Engine Cores | 16 | 32 | 16 |
| Estimated TOPS | ~38 TOPS (ESTIMATE) | ~31 TOPS (confirmed) | ~15.8 TOPS |
The M3 Ultra's 32-core Neural Engine provides double the dedicated ML acceleration hardware compared to the M5 Max. However, for LLM inference, the Neural Engine is far less important than memory bandwidth and capacity.
| Metric | M5 Max | M3 Ultra | M2 |
|---|---|---|---|
| Bandwidth | 614 GB/s | 819 GB/s | 100 GB/s |
| Unified Memory | 128 GB | 96 GB | 24 GB |
| Bandwidth per GB | 4.8 GB/s/GB | 8.5 GB/s/GB | 4.2 GB/s/GB |
Why this matters: LLM inference speed is almost entirely bottlenecked by memory bandwidth. Each token generated requires reading the entire model weights from memory. The formula is roughly:
Tokens/second ≈ Memory Bandwidth / Model Size (in bytes)
| Model | M5 Max (128GB) | M3 Ultra (96GB) | M2 (24GB) |
|---|---|---|---|
| Llama 3.1 8B (Q4) ~5GB | ~123 tok/s | ~164 tok/s | ~20 tok/s |
| Llama 3.1 70B (Q4) ~40GB | ~15 tok/s | ~20 tok/s | Won't fit |
| Llama 3.1 70B (Q8) ~70GB | ~9 tok/s | ~12 tok/s | Won't fit |
| Llama 3.1 405B (Q4) ~200GB | Won't fit | Won't fit | Won't fit |
| DeepSeek V3 (Q2) ~100GB | ~6 tok/s | Won't fit | Won't fit |
| Mixtral 8x22B (Q4) ~80GB | ~7.5 tok/s | ~10 tok/s (tight) | Won't fit |
Estimates assume llama.cpp Metal inference with reasonable batch sizes. Real-world performance varies by quantization, context length, and implementation.
The M5 Max wins on capacity (128GB vs 96GB) - it can load larger models. The M3 Ultra wins on speed (819 GB/s vs 614 GB/s) - it runs models faster. This is a genuine tradeoff:
- If you want to run very large models (70B+ at higher quantization, or 100B+ models), the M5 Max's 128GB is critical
- If you want faster inference on models that fit in either machine, the M3 Ultra's 33% higher bandwidth wins
- The M5 Max can fit DeepSeek V3 Q2 (~100GB); the M3 Ultra cannot
| Spec | M5 Max MBP | Mac Studio M3 Ultra | MacBook Air M2 |
|---|---|---|---|
| Capacity | 2 TB | 1 TB | 2 TB |
| Protocol | Apple Fabric (NVMe) | Apple Fabric (NVMe) | Apple Fabric (NVMe) |
| Est. Sequential Read | ~7,500 MB/s | ~7,500 MB/s | ~3,000 MB/s |
| Est. Sequential Write | ~6,000 MB/s | ~6,000 MB/s | ~2,500 MB/s |
The M5 Max and M3 Ultra both use the latest Apple Fabric SSD controllers with similar performance. The M2 in the Air uses an older controller with roughly 40% of the throughput.
The Mac Studio's 1TB is notably small for a daily driver workstation, especially for ML datasets, Docker images, and model files. The MBP's 2TB provides more breathing room.
Memory bandwidth is the single most important spec for local LLM inference. Here's how these compare:
| Metric | M5 Max | M3 Ultra | M2 |
|---|---|---|---|
| Total Bandwidth | 614 GB/s | 819 GB/s | 100 GB/s |
| Memory Bus Width | 512-bit (ESTIMATE) | 800-bit | 128-bit |
| Memory Type | LPDDR5X | LPDDR5X | LPDDR5 |
| Ratio vs M2 | 6.1x | 8.2x | 1.0x |
The M3 Ultra's bandwidth advantage comes from being a die-doubled chip (two M3 Max dies fused via UltraFusion interconnect), effectively doubling the memory bus width. The M5 Max, as a single die, cannot match this despite using faster memory.
For context: NVIDIA's RTX 4090 has ~1,008 GB/s bandwidth but only 24GB VRAM. Apple Silicon's advantage is the massive unified memory pool, even if bandwidth per unit is lower than dedicated GPUs.
| Task | M5 Max MBP | Mac Studio M3 Ultra | MacBook Air M2 |
|---|---|---|---|
| 4K ProRes timeline scrubbing | Excellent | Excellent | Good |
| 8K RAW editing | Very Good | Excellent | Poor |
| ProRes encode/decode engines | 2 | 4 | 1 |
| Color grading (GPU-heavy) | Very Good | Excellent | Fair |
The Mac Studio wins here due to more GPU cores and 4 ProRes engines (the Ultra doubles the Max's 2 engines). The M5 Max MBP is still extremely capable for 4K and competent at 8K.
| Task | M5 Max MBP | Mac Studio M3 Ultra | MacBook Air M2 |
|---|---|---|---|
| Full Xcode project build | ~1.0x (baseline) | ~0.85x (faster) | ~3.0x (slower) |
| Rust cargo build (large crate) | Very Fast | Fastest | Slow |
| Parallel compilation scaling | 18 threads | 28 threads | 8 threads |
The M3 Ultra's 28 cores provide roughly 55% more compile parallelism. For large projects with many compilation units, this translates to significantly faster builds. The M5 Max's higher single-core speed helps with linking and single-threaded compilation bottlenecks.
See Section 4 for detailed estimates. Summary:
- M5 Max: Fits the biggest models (128GB). Best choice if you need 70B+ Q8 or 100B+ Q4 models.
- M3 Ultra: Fastest inference for models up to ~80GB. 33% faster token generation.
- M2 Air: Limited to small models (7-13B). Functional for quick queries but not serious inference work.
| Feature | M5 Max MBP | Mac Studio M3 Ultra | MacBook Air M2 |
|---|---|---|---|
| Available RAM for containers | ~100+ GB | ~70+ GB | ~16 GB |
| CPU cores for multi-container | 18 | 28 | 8 |
| Storage for images | 2 TB | 1 TB (tight) | 2 TB |
| Network (wired) | TB5 adapter needed | 10Gb Ethernet built-in | USB adapter needed |
The Mac Studio is the clear winner for containerized workloads: more cores, built-in 10Gb Ethernet, and desktop reliability. The 1TB SSD is the only weak point - Docker images can eat storage fast.
| Metric | M5 Max MBP | Mac Studio M3 Ultra | MacBook Air M2 |
|---|---|---|---|
| Cooling | Active (fans) | Active (large fan + heatsink) | Passive (fanless) |
| Sustained Performance | 90-95% of peak | ~100% of peak | 70-80% of peak (throttles) |
| Fan Noise Under Load | Moderate | Low-moderate | Silent (but throttles) |
| Ambient Noise at Idle | Silent | Near-silent | Silent |
| Max TDP | ~120W (estimated) | ~480W system max | ~22W |
The Mac Studio's desktop-class cooling allows it to sustain peak performance indefinitely. It has a much larger thermal envelope and never needs to throttle.
The M5 Max MBP has good cooling for a laptop, but sustained workloads (continuous compilation, prolonged LLM inference, video encoding) will cause some thermal throttling. The laptop will maintain 90-95% of peak performance under continuous load, which is excellent for a laptop but not desktop-class sustained performance.
The MacBook Air has no fan. Under sustained load, it throttles significantly (20-30% performance loss after 10-15 minutes of heavy work). It's designed for bursty workloads, not sustained compute.
| Machine | Configuration | Estimated Price |
|---|---|---|
| 16" MacBook Pro M5 Max | 18c CPU, 40c GPU, 128GB, 2TB | ~$5,499 |
| Mac Studio M3 Ultra | 28c CPU, 60c GPU, 96GB, 1TB | ~$3,999 (original purchase price) |
| 15" MacBook Air M2 | 8c CPU, 10c GPU, 24GB, 2TB | ~$1,999 (original purchase price) |
M5 Max MBP pricing is estimated based on M4 Max pricing patterns. The base 16" M5 Max (32-core GPU, 36GB, 1TB) likely starts at $3,499. Upgrading to 40-core GPU, 128GB RAM, and 2TB SSD adds roughly $2,000 in Apple's typical config pricing.
| Metric | M5 Max MBP | Mac Studio M3 Ultra | MacBook Air M2 |
|---|---|---|---|
| GB6 Multi-Core per $1000 | ~5,455 | ~6,634 | ~4,855 |
| Memory GB per $1000 | ~23 GB | ~24 GB | ~12 GB |
| Memory Bandwidth per $1000 | ~112 GB/s | ~205 GB/s | ~50 GB/s |
The Mac Studio delivers the best value for raw compute and memory bandwidth. The M5 Max MBP charges a premium for portability + 128GB capacity. The MacBook Air is the worst value for compute but the best value for portability.
Wins:
- 32GB more unified memory (128 vs 96 GB) - this is the headline advantage. Fits models the Mac Studio can't.
- Portability - it's a laptop. You can take it to Kyoto.
- Built-in display - 16.2" XDR, 120Hz ProMotion, gorgeous
- Higher single-core speed - ~40% faster per-core than M3 Ultra
- Wi-Fi 7 + Bluetooth 6 - newer wireless
- Thunderbolt 5 - same as Mac Studio, but the MBP also has HDMI and SD card
- 2TB SSD - double the Mac Studio's current 1TB
- Battery - up to 22 hours, works unplugged
Losses:
- 33% less memory bandwidth (614 vs 819 GB/s) - models that fit on both machines run faster on the Studio
- 10 fewer CPU cores (18 vs 28) - compilation, video encoding, and multi-threaded work is slower
- 20 fewer GPU cores (40 vs 60) - GPU compute and rendering is weaker
- Half the Neural Engine (16 vs 32 cores)
- Thermal throttling - sustained workloads will slow down on a laptop
- No 10Gb Ethernet - need an adapter
- No desktop-class I/O - fewer ports, no USB-A
If you're moving to Kyoto (which you are), the M5 Max MBP is the right buy. You can't ship a Mac Studio easily, and you need a portable daily driver that can handle serious LLM inference, development, and AI work. The 128GB unified memory is the killer feature - it lets you run models that simply won't fit on the 96GB Mac Studio.
The Mac Studio remains the superior workstation when at home. Its 819 GB/s bandwidth and 28 CPU cores make it meaningfully faster for sustained workloads, compilations, and inference on models that fit in 96GB. It should continue running OpenClaw as the always-on home server.
The ideal setup: Use the M5 Max MBP as the primary machine in Kyoto and on the go. Keep the Mac Studio running at home as a remote development server, OpenClaw host, and inference backend you can SSH into. The Air becomes the ultra-light travel/couch machine (and Yan Li's backup).
| If your priority is... | Best choice |
|---|---|
| Running the largest possible LLMs locally | M5 Max MBP (128GB) |
| Fastest inference on 70B models | Mac Studio M3 Ultra (819 GB/s) |
| Compilation speed on large projects | Mac Studio M3 Ultra (28 cores) |
| Portability + serious power | M5 Max MBP |
| Always-on server / OpenClaw host | Mac Studio M3 Ultra |
| Light travel / basic tasks | MacBook Air M2 |
| GPU-heavy rendering | Mac Studio M3 Ultra (60 GPU cores) |
| Maximum number of external displays | Mac Studio M3 Ultra (up to 8) |
- Apple MacBook Pro specs: apple.com/macbook-pro/specs
- Apple Mac Studio specs: apple.com/mac-studio/specs
- Geekbench 6 Mac Studio (2025) M3 Ultra 28c/60g: browser.geekbench.com - SC: 3,213 / MC: 26,530 / OpenCL: 124,904 / Metal: 220,119
- Geekbench 6 Mac benchmarks list: browser.geekbench.com
- Mac Studio system specs verified via
node os.cpus()andos.totalmem()(running on the machine) - MacBook Air specs verified via SSH (
system_profiler SPHardwareDataType, Model: Mac14,15) - M5 Max benchmarks: NOT YET AVAILABLE on Geekbench as of March 4, 2026. All M5 Max performance numbers are estimates extrapolated from M5 base chip scores (17,464 MC), M4 Max scores (25,702 MC), and architectural scaling.
Report generated by Xinshu for Sean Grove, March 4, 2026. M5 Max performance estimates will need updating once real benchmarks become available.