Apple M5 vs AMD Ryzen AI Max+ 395

Agent access

1. ai-native userPoint an agent at llms.txt or agent-oriented docs

   weight 2 · round to AMD Ryzen AI Max+ 395
   Apple M5—none0/10

   Apple M5 is a hardware chip, not a documentation site or SaaS product; the probe explicitly shows no llms.txt exists at developer.apple.com (404), and there is no evidence of agent-oriented docs. Since this axis is plausible for a product with developer documentation but no such artifact exists, it is 'none' rather than 'na'.

   • [probe] “PROBE llms.txt: HTTP 404 at https://developer.apple.com/llms.txt”

   proof ↗⚑ flag

   AMD Ryzen AI Max+ 395✓fullprobed7/10

   A direct probe confirms that AMD's Ryzen AI developer docs serve a working llms.txt file (HTTP 200 with structured content) at ryzenai.docs.amd.com, which an AI agent could be pointed at directly. Missing for 10: broader agent-oriented documentation structure beyond the single llms.txt file, and independent/community confirmation of agents actually using it.

   • [probe] “PROBE llms.txt: HTTP 200 at https://ryzenai.docs.amd.com/llms.txt # Ryzen AI Software > Note: ROCm documentation is split across multiple p…”
   • [probe] “PROBE runtime (recorded 2026-09-15): the Ryzen AI Software documentation at ryzenai.docs.amd.com answered a keyless curl naming Ryzen AI — t…”

   proof ↗⚑ flag

2. ai-native userRun the product headlessly / in CI for automation

   weight 2 · round to AMD Ryzen AI Max+ 395
   Apple M5—none0/10

   The axis applies to this product kind (peer products hold positive or none verdicts on this story), so lack of evidence for an applicable capability is "none", never "na". (na/none harmonized at arena bring-up — see pipeline/scripts/na-harmonize.ts.)

   ⚑ flag

   AMD Ryzen AI Max+ 395~partialclaimed4/10

   The Ryzen AI software stack supports headless Linux server OSes (Ubuntu, RHEL) and exposes CLI/API-driven inference via ONNX Runtime (C++/Python APIs) and the Lemonade SDK for llama.cpp/OGA, which are automatable without a GUI. However, there is no explicit documentation of CI pipelines, containerized/Docker deployment, or automated build/test workflows for this hardware. Missing for 10: explicit CI/automation examples, containerization support, and independent confirmation of headless scripted runs.

   • [claimed-docs] “The AI model is deployed using the ONNX Runtime with either C++ or Python APIs.”
   • [claimed-docs] “the Lemonade SDK, which is multi-vendor open-source software that provides everything necessary for quickly getting started with LLMs on OGA…”
   • [claimed-docs] “Windows 11 - 64-Bit Edition , RHEL x86 64-Bit , Ubuntu x86 64-Bit”

   proof ↗⚑ flag

3. ai-native userDrive the product through a documented public API

   weight 3 · round drawn
   Apple M5~partialprobed4/10

   Apple documents public developer APIs (Metal 4 Tensor APIs for Neural Accelerators, Core ML, Foundation Models framework) that let developers program the M5's AI hardware, but this is a hardware/developer-framework API, not an agentic API meant for an 'AI-native user' to drive the product directly. missing for 10: no evidence of an agent-facing/programmatic control API for the chip itself, no llms.txt or agent-oriented API docs (llms.txt probe returned 404), no independent corroboration of AI agents actually invoking these APIs.

   • [claimed-docs] “Developers can also build solutions for their apps by directly programming the Neural Accelerators using Tensor APIs in Metal 4.”
   • [claimed-docs] “Applications using built-in Apple frameworks and APIs — like Core ML, Metal Performance Shaders, and Metal 4 — can automatically see immedia…”
   • [claimed-docs] “Also, developers using Apple’s Foundation Models framework will get faster performance.”
   • [claimed-docs] “Accelerate the training of machine learning models in third-party frameworks right on your Mac with PyTorch backends.”
   • [probe] “PROBE llms.txt: HTTP 404 at https://developer.apple.com/llms.txt”

   proof ↗⚑ flag

   AMD Ryzen AI Max+ 395~partialprobed4/10

   AMD documents software-level APIs (ONNX Runtime C++/Python APIs, Vitis AI EP, Lemonade SDK) for driving AI workloads on the NPU, but there is no public REST/OpenAPI-style programmatic interface for 'driving the product' as an agentic system — probes for OpenAPI/swagger specs all 404. missing for 10: a documented public REST/agent-facing API or SDK entry point beyond ML framework runtimes, independent confirmation of API usage for agentic control.

   • [claimed-docs] “The AI model is deployed using the ONNX Runtime with either C++ or Python APIs.”
   • [claimed-docs] “the Lemonade SDK, which is multi-vendor open-source software that provides everything necessary for quickly getting started with LLMs on OGA…”
   • [probe] “PROBE openapi: all candidate paths 404 (https://ryzenai.docs.amd.com/openapi.json, https://ryzenai.docs.amd.com/swagger.json, https://ryzena…”

   proof ↗⚑ flag

4. ai-native userBuild against official SDKs

   weight 2 · round to AMD Ryzen AI Max+ 395
   Apple M5~partialcommunity7/10

   Apple documents official SDKs (Metal 4 Tensor APIs/Neural Accelerators, Core ML, Metal Performance Shaders, Foundation Models framework, PyTorch-Metal backend) that AI-native developers can build against on M5 hardware, giving clear first-party API surfaces. However a community comment flags that Apple's ML libraries are 'insular and disconnected from the rest of the industry,' a real caveat on ecosystem openness rather than a functional failure. Missing for 10: independent hands-on developer reports building real AI apps with these SDKs, and clearer documentation/tutorials beyond marketing pages.

   • [claimed-docs] “Developers can also build solutions for their apps by directly programming the Neural Accelerators using Tensor APIs in Metal 4.”
   • [claimed-docs] “Applications using built-in Apple frameworks and APIs — like Core ML, Metal Performance Shaders, and Metal 4 — can automatically see immedia…”
   • [claimed-docs] “Also, developers using Apple’s Foundation Models framework will get faster performance.”
   • [claimed-docs] “Accelerate the training of machine learning models in third-party frameworks right on your Mac with PyTorch backends.”
   • [claimed-docs] “It supports Metal 4 and GPU Neural Accelerators for maximum performance, and can scale training across multiple Macs with RDMA over Thunderb…”
   • [claimed-docs] “Tap-to-segment lets you isolate objects within images, while OCR, barcode scanning, and your own custom tools can be passed directly to Appl…”
   • [community] “I appreciate Apple propping up the GPU performance of their SoC but it feels a bit pointless when all the libraries they provide are so insu…”

   proof ↗⚑ flag

   AMD Ryzen AI Max+ 395✓fullprobed7/10

   AMD provides official Ryzen AI SDK docs covering ONNX Runtime with C++/Python APIs, the Vitis AI EP, AMD Quark quantization toolkit, and the Lemonade SDK for LLMs, all live and crawlable at ryzenai.docs.amd.com, giving AI-native developers concrete official SDKs to build against. Missing for 10: independent hands-on developer reports validating SDK usability/completeness, and no OpenAPI/formal API reference confirmed (probe found 404s for openapi endpoints).

   • [claimed-docs] “This allows developers to build and deploy models trained in PyTorch or TensorFlow and run them directly on laptops powered by Ryzen AI usin…”
   • [claimed-docs] “AMD Quark is a comprehensive cross-platform deep learning toolkit designed to simplify and enhance the quantization of deep learning models.”
   • [claimed-docs] “The AI model is deployed using the ONNX Runtime with either C++ or Python APIs.”
   • [claimed-docs] “the Lemonade SDK, which is multi-vendor open-source software that provides everything necessary for quickly getting started with LLMs on OGA…”
   • [probe] “PROBE llms.txt: HTTP 200 at https://ryzenai.docs.amd.com/llms.txt # Ryzen AI Software > Note: ROCm documentation is split across multiple p…”
   • [probe] “PROBE runtime (recorded 2026-09-15): the Ryzen AI Software documentation at ryzenai.docs.amd.com answered a keyless curl naming Ryzen AI — t…”

   proof ↗⚑ flag

Agentic features

1. ai-native userOperate the product with natural-language commands

   weight 2 · round drawn
   Apple M5—none0/10

   The axis applies to this product kind (peer products hold positive or none verdicts on this story), so lack of evidence for an applicable capability is "none", never "na". (na/none harmonized at arena bring-up — see pipeline/scripts/na-harmonize.ts.)

   ⚑ flag

   AMD Ryzen AI Max+ 395—none0/10

   The axis applies to this product kind (peer products hold positive or none verdicts on this story), so lack of evidence for an applicable capability is "none", never "na". (na/none harmonized at arena bring-up — see pipeline/scripts/na-harmonize.ts.)

   ⚑ flag

Heavy compute

1. developerCompile big codebases and run heavy parallel jobs fast — documented core counts and boost behavior, corroborated by independent multi-core benchmarks

   weight 3 · round to Apple M5
   Apple M5~partialprobed5/10

   Apple documents GPU/Neural Engine core counts and memory bandwidth for M5 (apple-m5-docs-15) and independent Geekbench multi-core benchmarks show real gains (~15% multi-core uplift, comm-12/13) plus a multithreaded CPU performance claim (comm-8) and a throughput comparison (comm-14), giving some corroboration for parallel workload speed. However, Apple's own materials do not publish CPU core count, clock speed, or boost behavior for M5 (confirmed by apple-m5-probe-rt-1 noting no clock/TDP spec sheet exists), and no benchmarks specifically target compiling large codebases or dev toolchains. Missing for 10: documented CPU core count and boost/turbo clock specifics, compiler/build-workload benchmarks, and Mac-specific (not just iPad) independent multi-core corroboration.

   • [claimed-docs] “10-core GPU * Neural Accelerators * Hardware-accelerated ray tracing * 16-core Neural Engine * 153GB/s memory bandwidth”
   • [community] “It delivers up to four times the peak GPU compute performance compared with M4, provides 30% higher graphics performance, and offers 15% fas…”
   • [community] “iPad M5 vs M4 (leaked unbox video): Single-Core Score 4133 vs 3748 (110.3%); Multi-Core Score 15437 vs 13324 (115.9%). Same max clock speed,…”
   • [community] “Looks like an improvement over the M4 iPad of: Single Core ~12% (3679 vs 4133), Multi Core ~15% (13420 vs 15437), in line historically with …”
   • [community] “MicroGPT-C in pure C hits 10M TPS on Apple M5, compared to about 7M TPS on a 5-year-old AMD Ryzen 5 5600H.”
   • [probe] “PROBE runtime (recorded 2026-09-15): a keyless curl of Apple's M5 announcement — the chip's canonical public spec source — returned the page…”

   proof ↗⚑ flag

   AMD Ryzen AI Max+ 395—none0/10

   No evidence pack item documents core counts, boost clock behavior, or independent multi-core/compile benchmarks; the AMD docs and product page instead focus on AI/NPU software stack, TDP, and I/O specs. Community comments (comm-2) even push back on judging this chip by CPU-only benchmarks, but no concrete multi-core compile benchmark or core/boost spec is cited either way.

   ⚑ flag

Single thread

1. power-userEveryday interactive work feels instant — leading single-thread performance shown in independent benchmarks, not just a peak-GHz number on a slide

   weight 2 · round to Apple M5
   Apple M5~partialcommunity6/10

   Independent leaked Geekbench benchmarks show M5's single-core score improving ~10-12% over M4 (e.g., 4133 vs 3748), giving real hands-on single-thread numbers rather than just Apple's peak-GHz marketing. However, there's no independent comparison against competing power-user chips (AMD/Intel) to substantiate a 'leading' claim, and Apple's own docs focus on GPU/AI throughput rather than single-thread specs. Missing for 10: cross-vendor single-thread benchmark comparisons, sustained/interactive-workload latency testing, and independent reviewer verification beyond a leaked unboxing.

   • [community] “iPad M5 vs M4 (leaked unbox video): Single-Core Score 4133 vs 3748 (110.3%); Multi-Core Score 15437 vs 13324 (115.9%). Same max clock speed,…”
   • [community] “Looks like an improvement over the M4 iPad of: Single Core ~12% (3679 vs 4133), Multi Core ~15% (13420 vs 15437), in line historically with …”
   • [community] “It delivers up to four times the peak GPU compute performance compared with M4, provides 30% higher graphics performance, and offers 15% fas…”

   proof ↗⚑ flag

   AMD Ryzen AI Max+ 395—none0/10

   No independent single-thread CPU benchmark data appears anywhere in the evidence pack; the only performance-related community comments concern GPU/LLM token throughput and memory bandwidth, and one explicitly notes the CPU is not the reason to choose this part ('If you're just going to use the CPU obviously the 395 is not what you want'). There's no vendor or third-party single-thread benchmark citation to support the story.

   • [community] “Terrible benchmarks. They should probably compare the 395 GPU against the 9950 CPU or against a 7600/9060. If you're just going to use the C…”
   • [community] “Own a framework desktop 128gb. GPU bandwidth limits the amount of tokens/sec that you get. I've mostly been running QWEN-3.6-35B-A3B at Q8 a…”
   • [community] “With 128 GB of unified memory, large models can fit into memory, but memory bandwidth quickly becomes the limiting factor. With long context…”

   proof ↗⚑ flag

Toolchain

1. developerThis architecture is a first-class development target — mature compilers, official optimization guidance, and an OS and tooling ecosystem that treats it as tier one

   weight 3 · round to Apple M5
   Apple M5✓fullcommunity8/10

   Apple documents mature first-party tooling for the M-series/ARM64 target: Metal 4 with Tensor APIs and a dedicated Metal debugger, Core ML and Metal Performance Shaders auto-acceleration, a PyTorch GPU backend, and the Foundation Models framework getting native speedups — all signs of tier-one OS/toolchain integration on Apple Silicon. Community commentary corroborates real performance gains (10M TPS in a C benchmark, GPU/CPU uplifts) though one comment criticizes Apple's ML libraries as 'insular and disconnected from the rest of the industry,' a minor ecosystem caveat. Missing for 10: independent compiler-maturity benchmarks (LLVM/clang codegen quality vs x86), explicit official optimization guides beyond marketing copy, and broader third-party tooling parity evidence.

   • [claimed-docs] “Developers can also build solutions for their apps by directly programming the Neural Accelerators using Tensor APIs in Metal 4.”
   • [claimed-docs] “Applications using built-in Apple frameworks and APIs — like Core ML, Metal Performance Shaders, and Metal 4 — can automatically see immedia…”
   • [claimed-docs] “Also, developers using Apple’s Foundation Models framework will get faster performance.”
   • [claimed-docs] “Inspect, debug, and optimize your entire rendering pipeline with Metal debugger — from mesh shading to ray tracing and machine learning.”
   • [claimed-docs] “Accelerate the training of machine learning models in third-party frameworks right on your Mac with PyTorch backends.”
   • [claimed-docs] “It supports Metal 4 and GPU Neural Accelerators for maximum performance, and can scale training across multiple Macs with RDMA over Thunderb…”
   • [community] “I appreciate Apple propping up the GPU performance of their SoC but it feels a bit pointless when all the libraries they provide are so insu…”
   • [community] “MicroGPT-C in pure C hits 10M TPS on Apple M5, compared to about 7M TPS on a 5-year-old AMD Ryzen 5 5600H.”

   proof ↗⚑ flag

   AMD Ryzen AI Max+ 395~partialcommunity5/10

   AMD provides official Ryzen AI docs, ONNX Runtime + Vitis AI EP, the Quark quantization toolkit, and the Lemonade SDK, plus OS support for Windows 11, RHEL, and Ubuntu, indicating a real first-party tooling stack. However, community reports show developers relying on third-party runtimes (llama.cpp/Vulkan, Ollama) rather than a mature native compiler/optimization path, and cite memory-bandwidth bottlenecks and confusing benchmarks/naming as friction points, suggesting the ecosystem is still maturing rather than unambiguously tier-one. Missing for 10: evidence of mature first-party compiler toolchains, independent benchmarks confirming optimization guidance efficacy, and confirmation that NPU/GPU stack is treated as tier-one by major ML frameworks beyond ONNX.

   • [claimed-docs] “This allows developers to build and deploy models trained in PyTorch or TensorFlow and run them directly on laptops powered by Ryzen AI usin…”
   • [claimed-docs] “AMD Quark is a comprehensive cross-platform deep learning toolkit designed to simplify and enhance the quantization of deep learning models.”
   • [claimed-docs] “the Lemonade SDK, which is multi-vendor open-source software that provides everything necessary for quickly getting started with LLMs on OGA…”
   • [claimed-docs] “Windows 11 - 64-Bit Edition , RHEL x86 64-Bit , Ubuntu x86 64-Bit”
   • [community] “Own a framework desktop 128gb. GPU bandwidth limits the amount of tokens/sec that you get. I've mostly been running QWEN-3.6-35B-A3B at Q8 a…”
   • [community] “I'm running qwen3.6, gpt-oss and embeddinggemma with Ollama on Ryzen 7 8700G + 96 GB DDR5 with 12-14 tokens/sec... On Strix Halo it will run…”
   • [community] “With 128 GB of unified memory, large models can fit into memory, but memory bandwidth quickly becomes the limiting factor. With long context…”

   proof ↗⚑ flag

Virtualization

1. developerVMs and containers run well on this silicon — documented virtualization support and mainstream hypervisor and Docker workflows

   weight 2 · round drawn
   Apple M5—none0/10

   The evidence pack covers M5's AI/ML acceleration, GPU, Metal APIs, and general hardware specs, but contains no mention of virtualization support, hypervisor frameworks (e.g. Apple Hypervisor.framework, Parallels, UTM, VMware Fusion), or Docker/container workflows on M5. Missing for 10: documented virtualization framework support, hypervisor compatibility claims, Docker Desktop/container runtime performance data, and any developer or community confirmation of VM/container workflows on M5 silicon.

   ⚑ flag

   AMD Ryzen AI Max+ 395—none0/10

   The evidence pack lists supported OSes (Windows 11, RHEL, Ubuntu) but contains no documentation of virtualization features (SVM/AMD-V, IOMMU), hypervisor compatibility (KVM/Hyper-V/VMware), or Docker/container workflows on this chip. Virtualization support is a fair question for a CPU/SoC but no evidence confirms it here.

   • [claimed-docs] “Windows 11 - 64-Bit Edition , RHEL x86 64-Bit , Ubuntu x86 64-Bit”

   proof ↗⚑ flag

Gaming fps

1. gamerThis chip drives high frame rates in real games — vendor gaming claims (cache, boost behavior, integrated GPU class) corroborated by independent game benchmarks

   weight 3 · round to Apple M5
   Apple M5~partialcommunity3/10

   Apple's own materials claim gaming-relevant GPU gains (second-gen dynamic caching, 'smoother gameplay,' hardware ray tracing, up to 4x GPU compute vs M4), but the evidence pack contains no independent frame-rate or game-specific benchmarks — community discussion focuses on CPU Geekbench scores, LLM throughput, and memory bandwidth, not actual game FPS testing. Missing for 10: independent game benchmark results (FPS/frame-time comparisons), third-party reviewer corroboration of gaming claims, real-world title testing beyond synthetic CPU scores.

   • [claimed-docs] “Combined with rearchitected second-generation dynamic caching, the GPU provides smoother gameplay, more realistic visuals in 3D applications…”
   • [claimed-docs] “10-core GPU * Neural Accelerators * Hardware-accelerated ray tracing * 16-core Neural Engine * 153GB/s memory bandwidth”
   • [community] “It delivers up to four times the peak GPU compute performance compared with M4, provides 30% higher graphics performance, and offers 15% fas…”
   • [community] “iPad M5 vs M4 (leaked unbox video): Single-Core Score 4133 vs 3748 (110.3%); Multi-Core Score 15437 vs 13324 (115.9%). Same max clock speed,…”

   proof ↗⚑ flag

   AMD Ryzen AI Max+ 395—none0/10

   The evidence pack contains no vendor gaming claims (cache/boost/iGPU class positioning) and no independent game benchmark data or FPS figures; community comments are about AI/LLM token throughput and general benchmark methodology complaints, not real-game performance.

   • [community] “Terrible benchmarks. They should probably compare the 395 GPU against the 9950 CPU or against a 7600/9060. If you're just going to use the C…”
   • [community] “It would have been interesting to see the graphics performance and AI performance of the 395 compared to discrete gpus... how does it stack …”

   proof ↗⚑ flag

Media engines

1. creatorHardware media engines carry my editing and streaming — documented hardware encode and decode (AV1, HEVC, ProRes-class) on the chip itself

   weight 2 · round drawn
   Apple M5—none0/10

   The evidence pack contains no mention of a dedicated hardware media engine, nor any reference to AV1, HEVC, or ProRes encode/decode capabilities on M5 — it only covers GPU compute, Neural Accelerators, ML frameworks, and RAM/bandwidth debates. This is a fair axis for a media-focused Apple chip, but none of the provided docs or community posts document hardware video encode/decode support, so it cannot be credited as full/partial/disputed.

   ⚑ flag

   AMD Ryzen AI Max+ 395—none0/10

   Evidence pack contains no documentation of hardware media/video encode-decode engines (AV1, HEVC, ProRes-class) on this chip; all citations focus on AI/NPU software stack, CPU/GPU specs, TDP, and community discussion of LLM performance. Missing for 10: any mention of media/video codec engine, encode/decode capability, or streaming/editing hardware acceleration documentation.

   ⚑ flag

Npu sdk

1. ai-native userThe chip's AI acceleration is exposed to developers — an NPU or neural engine with a published TOPS figure (precision stated) and an official SDK or runtime that ships today

   weight 3 · round to AMD Ryzen AI Max+ 395
   Apple M5~partialprobed5/10

   Apple documents a 16-core Neural Engine plus GPU Neural Accelerators and ships real developer runtimes today (Core ML, Metal Performance Shaders, Metal 4 Tensor APIs, PyTorch backend), satisfying the 'official SDK/runtime ships today' half of the story. However, no published TOPS figure (with precision stated) for the Neural Engine or Neural Accelerators appears anywhere in the evidence pack — Apple's newsroom and spec pages list core counts and memory bandwidth but omit any TOPS metric. Missing for 10: a published TOPS number with stated precision (e.g., INT8/FP16) for the M5 Neural Engine or Neural Accelerators, and independent verification of that figure.

   • [claimed-docs] “Developers can also build solutions for their apps by directly programming the Neural Accelerators using Tensor APIs in Metal 4.”
   • [claimed-docs] “Applications using built-in Apple frameworks and APIs — like Core ML, Metal Performance Shaders, and Metal 4 — can automatically see immedia…”
   • [claimed-docs] “Accelerate the training of machine learning models in third-party frameworks right on your Mac with PyTorch backends.”
   • [claimed-docs] “10-core GPU * Neural Accelerators * Hardware-accelerated ray tracing * 16-core Neural Engine * 153GB/s memory bandwidth”
   • [probe] “PROBE runtime (recorded 2026-09-15): a keyless curl of Apple's M5 announcement — the chip's canonical public spec source — returned the page…”

   proof ↗⚑ flag

   AMD Ryzen AI Max+ 395~partialprobed6/10

   Strong evidence of an official, shipping SDK/runtime stack (Ryzen AI Software docs, ONNX Runtime + Vitis AI Execution Provider, AMD Quark quantization toolkit, Lemonade SDK for LLMs) confirming real developer-facing NPU acceleration tooling. However, the evidence pack contains no published TOPS figure or precision spec for the NPU itself, which the story explicitly requires. Missing for 10: a documented TOPS number with precision (e.g., INT8/INT4) for the XDNA2 NPU, and independent benchmark corroboration of that figure.

   • [claimed-docs] “This allows developers to build and deploy models trained in PyTorch or TensorFlow and run them directly on laptops powered by Ryzen AI usin…”
   • [claimed-docs] “AMD Quark is a comprehensive cross-platform deep learning toolkit designed to simplify and enhance the quantization of deep learning models.”
   • [claimed-docs] “The AI model is deployed using the ONNX Runtime with either C++ or Python APIs.”
   • [claimed-docs] “the Lemonade SDK, which is multi-vendor open-source software that provides everything necessary for quickly getting started with LLMs on OGA…”
   • [probe] “PROBE runtime (recorded 2026-09-15): the Ryzen AI Software documentation at ryzenai.docs.amd.com answered a keyless curl naming Ryzen AI — t…”

   proof ↗⚑ flag

Runtime support

1. developerMainstream local-AI runtimes target this silicon — llama.cpp, MLX, ONNX Runtime, or the vendor's own AI software stack document support for its CPU, GPU, or NPU

   weight 2 · round to AMD Ryzen AI Max+ 395
   Apple M5~partialcommunity6/10

   Apple documents its own AI stack (Core ML, Metal Performance Shaders, Metal 4 Tensor APIs, PyTorch backend via Metal, GPU Neural Accelerators) explicitly supporting M5's CPU/GPU/NPU, and community evidence confirms real-world LLM/diffusion workloads (webAI, Draw Things, Qwen models) running locally on the chip. However, no evidence explicitly ties llama.cpp, MLX, or ONNX Runtime by name to M5-specific support, so mainstream cross-vendor runtime targeting is only inferred, not documented. missing for 10: explicit llama.cpp/MLX/ONNX Runtime documentation naming M5 support, independent benchmark confirming these runtimes exploit M5's Neural Accelerators.

   • [claimed-docs] “Developers can also build solutions for their apps by directly programming the Neural Accelerators using Tensor APIs in Metal 4.”
   • [claimed-docs] “Applications using built-in Apple frameworks and APIs — like Core ML, Metal Performance Shaders, and Metal 4 — can automatically see immedia…”
   • [claimed-docs] “the new 14-inch MacBook Pro and iPad Pro benefit from dramatically accelerated processing for AI-driven workflows, such as running diffusion…”
   • [claimed-docs] “Accelerate the training of machine learning models in third-party frameworks right on your Mac with PyTorch backends.”
   • [claimed-docs] “It supports Metal 4 and GPU Neural Accelerators for maximum performance, and can scale training across multiple Macs with RDMA over Thunderb…”
   • [community] “153 GB/s is not bad at all for a base model; the Nvidia DGX Spark has only 273 GB/s memory bandwidth despite being billed as a desktop 'AI s…”
   • [community] “It can run larger models quite slowly but lacks matmul acceleration (included in the M5) useful for context/prompt performance at inference.…”

   proof ↗⚑ flag

   AMD Ryzen AI Max+ 395✓fullcommunity8/10

   AMD's own Ryzen AI docs confirm ONNX Runtime with Vitis AI Execution Provider support, and explicitly mention the Lemonade SDK enabling llama.cpp on this platform; community hands-on reports independently confirm llama.cpp (Vulkan) and Ollama running well on Strix Halo silicon. missing for 10: explicit MLX support (Apple-specific, not applicable here but leaves a gap in the story's named runtimes), and clearer first-party NPU-specific runtime benchmarks beyond GPU/CPU token-rate anecdotes.

   • [claimed-docs] “This allows developers to build and deploy models trained in PyTorch or TensorFlow and run them directly on laptops powered by Ryzen AI usin…”
   • [claimed-docs] “The AI model is deployed using the ONNX Runtime with either C++ or Python APIs.”
   • [claimed-docs] “the Lemonade SDK, which is multi-vendor open-source software that provides everything necessary for quickly getting started with LLMs on OGA…”
   • [community] “Own a framework desktop 128gb. GPU bandwidth limits the amount of tokens/sec that you get. I've mostly been running QWEN-3.6-35B-A3B at Q8 a…”
   • [community] “I'm running qwen3.6, gpt-oss and embeddinggemma with Ollama on Ryzen 7 8700G + 96 GB DDR5 with 12-14 tokens/sec... On Strix Halo it will run…”

   proof ↗⚑ flag

Llm memory

1. ai-native userRun a large local LLM (70B-class, quantized) on this platform — enough addressable memory and published memory bandwidth to make local inference practical

   weight 3 · round to AMD Ryzen AI Max+ 395
   Apple M5!disputedcontradicted4/10

   Apple markets M5's unified memory as enabling 'larger AI models completely on device' (apple-m5-docs-5), but the shipping base M5 specs show only 32GB max RAM and 153GB/s bandwidth (apple-m5-docs-15), and community hands-on commentary explicitly states this is 'not enough to run viable open source LLM models properly' and that 32GB is 'an even bigger problem' for real workloads (apple-m5-comm-1, apple-m5-comm-4, apple-m5-comm-9). Higher-memory Pro/Max variants needed for 70B-class quantized models are not yet available. Missing for 10: published Pro/Max M5 specs with sufficient memory/bandwidth, and independent benchmarks of actual 70B-class quantized inference throughput.

   • [claimed-docs] “The unified memory architecture enables the entire chip to access a large single pool of memory, which allows MacBook Pro, iPad Pro, and App…”
   • [claimed-docs] “10-core GPU * Neural Accelerators * Hardware-accelerated ray tracing * 16-core Neural Engine * 153GB/s memory bandwidth”
   • [community] “This is only the base model, no upgrades yet for the Pro/Max version. The memory bandwidth is 153GB/s which is not enough to run viable open…”
   • [community] “The memory capacity to me is an even bigger problem, at 32GB max.”
   • [community] “32GB RAM limit on current M5 models. Now wait for M5 Max.”
   • [community] “It can run larger models quite slowly but lacks matmul acceleration (included in the M5) useful for context/prompt performance at inference.…”

   proof ↗⚑ flag

   AMD Ryzen AI Max+ 395~partialcommunity6/10

   Community hands-on reports confirm the platform (128GB unified memory) can actually run large quantized models like Qwen3.5-122B-A10B at Q4 and similar 35B+ models via llama.cpp, showing real-world feasibility of 70B-class local inference. However, users consistently note memory bandwidth is the limiting factor for tokens/sec, and no official AMD-published memory bandwidth spec appears in the evidence pack. Missing for 10: an official AMD-published memory bandwidth figure, first-party benchmarks for 70B-class models, and confirmation that performance is 'practical' (not just possible) at long context lengths.

   • [community] “Own a framework desktop 128gb. GPU bandwidth limits the amount of tokens/sec that you get. I've mostly been running QWEN-3.6-35B-A3B at Q8 a…”
   • [community] “I'm running qwen3.6, gpt-oss and embeddinggemma with Ollama on Ryzen 7 8700G + 96 GB DDR5 with 12-14 tokens/sec... On Strix Halo it will run…”
   • [community] “With 128 GB of unified memory, large models can fit into memory, but memory bandwidth quickly becomes the limiting factor. With long context…”

   proof ↗⚑ flag

Memory spec

1. developerSize memory-bound workloads from the vendor's own numbers — published memory type, capacity ceiling, and bandwidth (or spec detail complete enough to derive it)

   weight 2 · round to Apple M5
   Apple M5~partialcommunity5/10

   Apple's spec page explicitly states 153GB/s memory bandwidth for the base M5 and describes a unified memory architecture, giving a developer a concrete bandwidth figure and general memory type concept, but it lacks a stated memory technology (e.g., LPDDR generation) and does not publish a capacity ceiling — the 32GB max cited comes only from community discussion, not vendor docs. missing for 10: vendor-stated memory technology/type, vendor-published max capacity configuration, independent corroboration of the bandwidth figure.

   • [claimed-docs] “10-core GPU * Neural Accelerators * Hardware-accelerated ray tracing * 16-core Neural Engine * 153GB/s memory bandwidth”
   • [claimed-docs] “The unified memory architecture enables the entire chip to access a large single pool of memory, which allows MacBook Pro, iPad Pro, and App…”
   • [community] “The memory capacity to me is an even bigger problem, at 32GB max.”
   • [community] “32GB RAM limit on current M5 models. Now wait for M5 Max.”

   proof ↗⚑ flag

   AMD Ryzen AI Max+ 395—none0/10

   The evidence pack includes AMD's spec page items but none cite memory type, capacity ceiling, or bandwidth figures (only TDP, NVMe, OS, and display specs are shown); community posts discuss bandwidth being a bottleneck qualitatively but no vendor numbers are cited to let a developer size workloads.

   • [claimed-docs] “AMD Configurable TDP (cTDP) 45-120W”
   • [claimed-docs] “Max Displays 4”
   • [community] “Own a framework desktop 128gb. GPU bandwidth limits the amount of tokens/sec that you get. I've mostly been running QWEN-3.6-35B-A3B at Q8 a…”
   • [community] “With 128 GB of unified memory, large models can fit into memory, but memory bandwidth quickly becomes the limiting factor. With long context…”

   proof ↗⚑ flag

Platform io

1. developerThe platform has documented I/O headroom — PCIe generation and lanes, fast storage, and external connectivity specs I can plan a build or dock setup around

   weight 2 · round to AMD Ryzen AI Max+ 395
   Apple M5—none0/10

   The evidence pack contains only display-resolution specs and memory bandwidth figures, with no documented PCIe generation/lane count, SSD/storage throughput specs, or Thunderbolt/USB port specifications that a developer could plan a dock or expansion setup around.

   • [claimed-docs] “Up to 24 hours video streaming”
   • [claimed-docs] “One display up to a native resolution of 8K at 60Hz or 5K at 120Hz or 4K at 240Hz”
   • [claimed-docs] “10-core GPU * Neural Accelerators * Hardware-accelerated ray tracing * 16-core Neural Engine * 153GB/s memory bandwidth”

   proof ↗⚑ flag

   AMD Ryzen AI Max+ 395~partialclaimed2/10

   AMD's product page lists some I/O-adjacent specs (NVMe boot/RAID support, max 4 displays, OS support) but there is no documented PCIe generation or lane count, and no external connectivity specs (USB/Thunderbolt/dock) to plan a build around. missing for 10: PCIe generation, PCIe lane count, USB/Thunderbolt/external port specs, dock compatibility documentation.

   • [claimed-docs] “NVMe Support Boot , RAID0 , RAID1”
   • [claimed-docs] “Max Displays 4”
   • [claimed-docs] “Windows 11 - 64-Bit Edition , RHEL x86 64-Bit , Ubuntu x86 64-Bit”

   proof ↗⚑ flag

Socket longevity

1. power-userUpgrade the CPU without replacing the platform — a documented socket with a stated multi-generation support commitment

   weight 2 · round drawn
   Apple M5—none0/10

   The axis applies to this product kind (peer products hold positive or none verdicts on this story), so lack of evidence for an applicable capability is "none", never "na". (na/none harmonized at arena bring-up — see pipeline/scripts/na-harmonize.ts.)

   ⚑ flag

   AMD Ryzen AI Max+ 395—none0/10

   The evidence pack contains no mention of a socket, upgrade path, or multi-generation platform commitment for the Ryzen AI Max+ 395; all specs shown are TDP, display, storage, and OS support with no socket/upgrade language. Community citations also reference it in fixed-configuration devices (e.g., Framework Desktop, mini-PCs) rather than a socketed upgrade path. missing for 10: any documented socket name, generational upgrade commitment, or evidence of user-replaceable CPU in a motherboard.

   • [claimed-docs] “AMD Configurable TDP (cTDP) 45-120W”
   • [claimed-docs] “Max Displays 4”
   • [community] “Own a framework desktop 128gb. GPU bandwidth limits the amount of tokens/sec that you get. I've mostly been running QWEN-3.6-35B-A3B at Q8 a…”

   proof ↗⚑ flag

Mobile endurance

1. power-userThis chip powers thin, quiet, all-day-battery machines — shipping in fanless or low-power designs with credible battery-life evidence

   weight 2 · round to Apple M5
   Apple M5✓fullcommunity8/10

   Apple's M5 ships in fanless iPad Pro and thin, fanless-adjacent MacBook Pro designs with documented all-day battery claims (up to 24 hours video streaming) and community benchmarks confirming real-world performance/efficiency gains over M4. Community discussion corroborates strong CPU efficiency and performance-per-watt improvements, consistent with Apple's power-efficiency-first chip design philosophy. missing for 10: independent third-party battery-life testing (e.g., reviewer runtime benchmarks) and explicit fanless-design confirmation for the specific M5 MacBook Pro SKU.

   • [claimed-docs] “Up to 24 hours video streaming”
   • [claimed-docs] “10-core GPU * Neural Accelerators * Hardware-accelerated ray tracing * 16-core Neural Engine * 153GB/s memory bandwidth”
   • [community] “It delivers up to four times the peak GPU compute performance compared with M4, provides 30% higher graphics performance, and offers 15% fas…”
   • [community] “iPad M5 vs M4 (leaked unbox video): Single-Core Score 4133 vs 3748 (110.3%); Multi-Core Score 15437 vs 13324 (115.9%). Same max clock speed,…”
   • [community] “Looks like an improvement over the M4 iPad of: Single Core ~12% (3679 vs 4133), Multi Core ~15% (13420 vs 15437), in line historically with …”

   proof ↗⚑ flag

   AMD Ryzen AI Max+ 395—none0/10

   The evidence pack shows a configurable TDP range of 45–120W and OS/platform specs, but contains no mention of fanless designs, thin-and-light chassis, or battery-life benchmarks; a community comment even complains AMD hasn't focused on reducing power usage. This is a fair axis for a laptop-class chip, but no supporting evidence exists.

   • [claimed-docs] “AMD Configurable TDP (cTDP) 45-120W”
   • [community] “The performance is incredible... Now when will AMD put some real effort in to reducing power usage?”

   proof ↗⚑ flag

Perf per watt

1. creatorLong renders and exports don't throttle away — documented power envelopes and independent testing showing strong sustained performance per watt

   weight 3 · round to AMD Ryzen AI Max+ 395
   Apple M5—none0/10

   No documented power envelope (TDP/wattage) or independent sustained-load/thermal-throttling testing is present; evidence covers battery-video-hours, peak GPU/CPU speedup claims, and short benchmark scores, none of which address sustained render/export performance-per-watt over time.

   ⚑ flag

   AMD Ryzen AI Max+ 395~partialcommunity3/10

   AMD documents a configurable TDP range (45-120W) for the chip, but there is no independent benchmark or hands-on testing demonstrating sustained performance-per-watt or absence of thermal throttling during long renders/exports; community comments focus on LLM token throughput and even note frustration that AMD hasn't addressed power usage. missing for 10: independent sustained-load/thermal throttling benchmarks, creator-workload (render/export) power efficiency data, cooling/chassis-specific real-world tests.

   • [claimed-docs] “AMD Configurable TDP (cTDP) 45-120W”
   • [community] “The performance is incredible... Now when will AMD put some real effort in to reducing power usage?”
   • [community] “Terrible benchmarks. They should probably compare the 395 GPU against the 9950 CPU or against a 7600/9060. If you're just going to use the C…”

   proof ↗⚑ flag

Spec disclosure

1. power-userComparison-shop from a real spec sheet — the vendor publishes clocks, power, memory support, and AI TOPS with test conditions, instead of marketing adjectives

   weight 2 · round to AMD Ryzen AI Max+ 395
   Apple M5—none0/10

   Apple's own M5 materials list core counts and unified memory bandwidth (153GB/s) but omit clock speeds, TDP/power figures, and AI TOPS numbers with test conditions — instead using relative marketing comparisons like 'up to 4x GPU compute' or '30% higher graphics performance' (apple-m5-comm-8). The runtime probe explicitly confirms Apple publishes no clock-speed, TDP, or ARK-style spec sheet, with the newsroom post serving as the only 'spec disclosure' (apple-m5-probe-rt-1).

   • [claimed-docs] “10-core GPU * Neural Accelerators * Hardware-accelerated ray tracing * 16-core Neural Engine * 153GB/s memory bandwidth”
   • [community] “It delivers up to four times the peak GPU compute performance compared with M4, provides 30% higher graphics performance, and offers 15% fas…”
   • [probe] “PROBE runtime (recorded 2026-09-15): a keyless curl of Apple's M5 announcement — the chip's canonical public spec source — returned the page…”
   • [probe] “PROBE llms.txt: HTTP 404 at https://developer.apple.com/llms.txt”

   proof ↗⚑ flag

   AMD Ryzen AI Max+ 395~partialcommunity5/10

   AMD's product page does publish concrete spec-sheet items (cTDP 45-120W, NVMe RAID support, OS support, display counts, voltage offset support) rather than pure marketing prose, and the Ryzen AI docs give technical detail on the software stack. However, the pack contains no explicit clock-speed figures or AI TOPS number with stated test conditions, and community commentary explicitly complains that AMD's benchmark comparisons are misleading/incomplete rather than rigorously documented. missing for 10: published boost/base clock figures, an explicit AI TOPS figure with test-methodology footnotes, and independent verification that the spec sheet's numbers hold up under real-world testing.

   • [claimed-docs] “AMD Configurable TDP (cTDP) 45-120W”
   • [claimed-docs] “Curve Optimizer Voltage Offsets Yes”
   • [claimed-docs] “NVMe Support Boot , RAID0 , RAID1”
   • [claimed-docs] “Windows 11 - 64-Bit Edition , RHEL x86 64-Bit , Ubuntu x86 64-Bit”
   • [claimed-docs] “Max Displays 4”
   • [community] “Terrible benchmarks. They should probably compare the 395 GPU against the 9950 CPU or against a 7600/9060. If you're just going to use the C…”
   • [community] “It would have been interesting to see the graphics performance and AI performance of the 395 compared to discrete gpus... how does it stack …”

   proof ↗⚑ flag