Skip to main content

Apple M1 Ultra (128GB Unified Memory)

Overview

Apple M1 Ultra is Apple Silicon's first-generation Ultra flagship, announced at WWDC 2022-03-08. It uses UltraFusion interconnect (1st gen) to merge two M1 Max dies into a single chip, delivering 128GB unified memory + 800 GB/s memory bandwidth (the industry's first consumer-grade 128GB UMA solution), marking a milestone in Apple Silicon history.

Historical significance: The M1 Ultra is the founding product of the Apple Silicon Ultra series, laying the architectural foundation for subsequent M2/M3/M4/M5 Ultra generations. It launched with Mac Studio 2022 and was replaced by M2 Ultra in 2023-06.

Core Specifications

ItemParameter
ArchitectureApple Silicon (M1 Max × 2, UltraFusion 1.0)
ProcessTSMC 5nm
CPU Cores20 (16P + 4E)
GPU Cores64
Neural Engine32-core, 22 TOPS
Unified Memory64GB / 128GB LPDDR5
Memory Bandwidth800 GB/s
FP32 GPU21 TFLOPS
FP16 GPU42 TFLOPS
Media EngineHardware H.264/HEVC/ProRes
TDP60-120 W
Launch PriceMac Studio M1 Ultra 128GB: $5,999 (2022-03 launch price)
Launch Date2022-03-18 (Mac Studio launch)
Discontinued2023-06 (replaced by M2 Ultra)

UltraFusion 1.0 Interconnect

DimensionUltraFusion 1.0 (M1)UltraFusion 2.0 (M3+)
Bandwidth2.5 TB/s2.5 TB/s
Channels4 channels4 channels
Latency~100ns~80ns
First appeared2022 M1 Ultra2023 M3 Ultra

Key technologies:

  • Die-to-die silicon interposer: Similar to TSMC CoWoS
  • Transparent to software: System sees a single chip
  • Unified memory pool: CPU/GPU/Neural Engine all access 128GB

Apple Silicon Evolution Timeline

GenerationFlagshipProcessMemoryBandwidthCPU CoresGPU CoresNE
M1M1 Ultra (2022)5nm128GB800 GB/s20 (16P+4E)6422 TOPS
M2M2 Ultra (2023)5nm192GB800 GB/s24 (16P+8E)7631.6 TOPS
M3M3 Ultra (2023-12)3nm256GB800 GB/s32 (24P+8E)8072 TOPS
M4M4 Ultra (2025-Q4 estimated)3nm256GB819 GB/s32 (24P+8E)8076 TOPS
M5M5 Ultra (2026 H2 estimated)2nm384GB~1 TB/s40 (24P+16E)96-120~120 TOPS

Comparison with M1 Max

MetricM1 UltraM1 MaxImprovement
Process5nm5nmSame
CPU Cores20 (16P+4E)10 (8P+2E)
GPU Cores6432
Neural Engine22 TOPS11 TOPS
Memory128GB LPDDR564GB LPDDR5
Bandwidth800 GB/s400 GB/s
UltraFusion2.5 TB/s (dual Max interconnect)-New
Price (Mac Studio 128GB)$5,999$3,999+50%

Local LLM Inference Performance (128GB variant)

ModelQuantizationPerformance (tok/s)Notes
Llama 2 7BFP16~35 tok/sMainstream
Llama 2 13BQ4_K_M~18 tok/sPerformance limited
Llama 2 70BQ4_K_M~3-4 tok/s128GB can load quantized
Llama 1 65BQ4_K_M~4 tok/sEarly model
Mistral 7BQ4_K_M~38 tok/sVery fast
Falcon 40BQ4_K_M~6 tok/s
MPT 30BFP16~5 tok/sFull load possible
Code Llama 34BQ4_K_M~7 tok/sCoding

Significance of 128GB M1 Ultra: First time in 2022 that consumer hardware could run a 70B LLM (albeit with Q4 quantization), at a price of $5,999 (far below NVIDIA DGX Station at $50K+).

Vendor Information

ItemDetails
VendorApple Inc.
ProductMac Studio (launched 2022-03, discontinued 2023-06)
Mac Studio M1 Ultra 128GB$5,999 (2022-03 launch price)
Discontinued2023-06 (replaced by M2 Ultra)
Target MarketCreators, local AI (early), ML researchers
Historical PositionFounding product of Apple Silicon Ultra series

Use Cases

  • Local LLM inference (128GB UMA, revolutionary in 2022)
  • Stable Diffusion 1.x image generation
  • Final Cut Pro / Logic Pro hardware acceleration
  • 4K video editing
  • Early Apple Silicon on-device AI development
  • Not suitable for: datacenter training
  • Not suitable for: large-scale inference
  • Not suitable for: FP8 training (FP16/BF16 only)

M1 Ultra vs M2 Ultra vs M3 Ultra

MetricM1 Ultra (2022)M2 Ultra (2023)M3 Ultra (2023-12)
Process5nm5nm (N5P)3nm
CPU Cores20 (16P+4E)24 (16P+8E)32 (24P+8E)
GPU Cores647680
NE22 TOPS31.6 TOPS72 TOPS
Memory128GB LPDDR5192GB LPDDR5256GB LPDDR5
Bandwidth800 GB/s800 GB/s800 GB/s
Price$5,999$5,499$5,899

Key Features

  • UltraFusion 1.0: Industry-first consumer inter-die silicon interconnect at 2.5 TB/s
  • 128GB UMA: Largest consumer unified memory in 2022
  • 800 GB/s: Among the highest consumer GPU/SoC bandwidths
  • 5nm TSMC: First-gen Apple Silicon flagship
  • Drawbacks: No FP8 support, $6K price, discontinued

Historical Significance

  • 2022 Revolution: First time consumer hardware could run a 70B LLM
  • Architecture foundation: M2/M3/M4/M5 Ultra all evolved from this architecture
  • Ecosystem beginning: Early llama.cpp / MLX adaptation
  • NVIDIA alternative: Starting point for local AI workstations