Unified Memory Architecture
The memory architecture used in Apple's M-series chips, which integrates the memory used by the CPU, GPU, and Neural Engine into a single high-bandwidth shared memory pool packaged adjacent to the SoC, eliminating data-transfer latency while creating a strong dependency on high-quality LPDDR memory.
Core Facts
Starting with the M1 chip, Apple uses a Unified Memory Architecture (UMA) where the CPU, GPU, and Neural Engine share the same memory pool.
Apple M系列芯片采用统一内存架构（UMA），CPU和GPU共享同一块高带宽内存池
M1's unified memory architecture allows the CPU and GPU to physically share the same memory, eliminating the need for explicit data copying and requiring only the passing of pointers or virtual addresses.
Timeline (last 90 days)
UMA's memory bandwidth leads on mobile, but still lags behind the HBM bandwidth of data-center-grade GPUs, which becomes a bottleneck in high-concurrency inference or training scenarios.
UMA's memory capacity is tied to the chip package, and users cannot upgrade it themselves.
The unified memory architecture lets the CPU, GPU and Neural Engine share the same high-bandwidth memory, which is particularly beneficial for workloads such as video editing, 3D rendering and AI inference.
The unified memory architecture eliminates the data-copy overhead between CPU and GPU in traditional PCs.
The ratio between nominal memory and model-usable memory is not a universal formula; how much is taken varies by machine, system, drivers, and usage, and there is no all-purpose rule of "30% off."
The unified memory architecture (UMA) lets the CPU and GPU share the same high-bandwidth memory pool, and the GPU can directly address system memory.
The Windows system, background programs, GPU reservations, and so on first take a portion of memory from the unified memory pool, resulting in a gap between nominal memory and model-usable memory.
The cost of the unified memory architecture is lower memory bandwidth than the dedicated VRAM of high-end discrete GPUs, so throughput is usually below that of data-center GPUs.
For local inference scenarios, high-bandwidth memory (such as Apple's unified memory architecture, or the HBM of high-end GPUs) is more important than multi-core CPUs.
The benefit of a unified memory architecture is that data does not need to be copied back and forth between processors, giving lower latency and higher energy efficiency, which especially benefits graphics and machine learning tasks; but it also means memory capacity cannot be upgraded later, and the configuration at purchase is the configuration for life.
13 more timeline events
All Facts (20)
Starting with the M1 chip, Apple uses a Unified Memory Architecture (UMA) where the CPU, GPU, and Neural Engine share the same memory pool.
80%VerifiedApple M系列芯片采用统一内存架构（UMA），CPU和GPU共享同一块高带宽内存池
80%VerifiedM1's unified memory architecture allows the CPU and GPU to physically share the same memory, eliminating the need for explicit data copying and requiring only the passing of pointers or virtual addresses.
75%VerifiedApple Silicon's Neural Engine shares a Unified Memory Architecture with the CPU and GPU, eliminating the bottleneck of data transfer between different memory pools in traditional heterogeneous computing.
75%VerifiedWhen selecting hardware for local large model inference, memory capacity should be prioritized first, followed by architecture and memory bandwidth.
70%VerifiedThe AMD Ryzen AI Max+ 395 features a 16-core N5-process CPU paired with 40 RDNA 3.5 compute units and supports up to 128 GB of LPDDR5X unified memory.
65%UnverifiedApple Silicon M-series unified memory architecture allows CPU and GPU to share the same physical memory, eliminating data copy overhead beneficial for LLM inference
90%UnverifiedThe NVIDIA Jetson series adopts a SoC unified memory architecture in which the CPU and GPU share the same physical memory.
70%Partially Verified统一内存架构使CPU、GPU与NPU共享同一块内存池，避免数据在处理单元间反复搬运的开销
65%UnverifiedMini PCs equipped with a unified memory architecture can run large models with 70 billion parameters (70B), whereas high-end discrete graphics cards cannot fully load them due to insufficient VRAM.
60%UnverifiedUMA's memory bandwidth leads on mobile, but still lags behind the HBM bandwidth of data-center-grade GPUs, which becomes a bottleneck in high-concurrency inference or training scenarios.
50%UnverifiedUMA's memory capacity is tied to the chip package, and users cannot upgrade it themselves.
50%UnverifiedThe unified memory architecture lets the CPU, GPU and Neural Engine share the same high-bandwidth memory, which is particularly beneficial for workloads such as video editing, 3D rendering and AI inference.
50%UnverifiedThe unified memory architecture eliminates the data-copy overhead between CPU and GPU in traditional PCs.
50%UnverifiedThe ratio between nominal memory and model-usable memory is not a universal formula; how much is taken varies by machine, system, drivers, and usage, and there is no all-purpose rule of "30% off."
50%UnverifiedThe Windows system, background programs, GPU reservations, and so on first take a portion of memory from the unified memory pool, resulting in a gap between nominal memory and model-usable memory.
50%UnverifiedThe unified memory architecture (UMA) lets the CPU and GPU share the same high-bandwidth memory pool, and the GPU can directly address system memory.
50%UnverifiedThe cost of the unified memory architecture is lower memory bandwidth than the dedicated VRAM of high-end discrete GPUs, so throughput is usually below that of data-center GPUs.
50%UnverifiedFor local inference scenarios, high-bandwidth memory (such as Apple's unified memory architecture, or the HBM of high-end GPUs) is more important than multi-core CPUs.
50%UnverifiedThe benefit of a unified memory architecture is that data does not need to be copied back and forth between processors, giving lower latency and higher energy efficiency, which especially benefits graphics and machine learning tasks; but it also means memory capacity cannot be upgraded later, and the configuration at purchase is the configuration for life.
50%