Updated October 25, 2026: This report now separates AMD’s published Ryzen AI Halo specifications and measurements from results that still require independent, workload-matched testing.
AMD’s Ryzen AI Halo Developer Platform with the Ryzen AI Max+ PRO 495 is designed around one unusual advantage: 192GB of shared LPDDR5X memory in a compact x86 system. AMD says up to 160GB can be allocated as graphics memory and positions the platform for 300-billion-parameter and larger language models at 4-bit quantization.
That is a capacity claim, not a universal speed result. A model can fit and still respond too slowly for an interactive workflow. The useful buying question is therefore not merely “Can it load?” but “Does the exact model, context, runtime and task meet my waiting-time and output-quality requirements?”
What AMD’s specifications establish
| Published item | What it supports | What it does not prove |
|---|---|---|
| 192GB LPDDR5X unified memory | A substantially larger shared capacity pool than ordinary consumer GPUs | That all 192GB is available to one model or GPU workload |
| Up to 160GB assignable as graphics memory | AMD’s advertised upper graphics-memory allocation | Whole-workload fit after the OS, runtime, cache and other applications |
| 273GB/s published memory bandwidth | The memory-system specification for the 192GB developer platform | Tokens per second for an arbitrary model or runtime |
| 16 Zen 5 CPU cores, 40 RDNA 3.5 GPU compute units and up to 55 TOPS NPU | The compute resources of the Ryzen AI Max+ PRO 495 platform | A direct performance ratio against CUDA GPUs or Apple systems |
| 120W platform TDP; Linux or Windows 11 | The listed platform power and operating-system options | Complete-system power, acoustics or support in every local-AI application |
What AMD has measured
AMD reports one relevant large-language-model scenario for a preproduction Ryzen AI Halo developer platform: 36 output tokens per second and 446 input tokens per second at 128K context. AMD identifies the system as a Ryzen AI Max+ platform using LM Studio 0.4.12, Vulkan llama.cpp and Adrenalin driver 26.10.08.
That result is useful only within its disclosed boundary. It is a manufacturer measurement, not a HighVRAM test, and it does not establish performance for every 300B model, quantization, prompt, batch size, operating system or frontend. AMD’s public note does not turn 36 tokens per second into a general platform rating.
Why the earlier estimates were removed
The earlier article projected several Ryzen AI Halo and Mac output rates from memory bandwidth. Those estimates were too weak for a purchase decision. Bandwidth can constrain inference, but observed speed also depends on model architecture, active parameters, quantization, context, cache, batching, kernels, power limits, thermals and software maturity. Results measured on the older 128GB Ryzen AI Max+ 395 do not automatically predict the 192GB Ryzen AI Max+ PRO 495.
The defensible comparison is narrower: the 192GB platform offers more potential shared-memory capacity than the 128GB predecessor, while any speed comparison must hold the workload and measurement boundary constant.
What 300B support really means
AMD says the platform can support 300B and larger language models at 4-bit quantization. Parameter count alone does not determine fit. The exact checkpoint files, quantization method, model architecture, context, KV cache, runtime workspace, operating-system use and other applications all consume memory. “Up to 160GB” graphics allocation also leaves less than the full 192GB pool for GPU-addressed work.
For a buyer, the 300B statement should be treated as a vendor-supported capacity class. It is not proof that every 300B checkpoint fits, that one-million-token context is practical, or that generation is interactive.
What to verify before buying
- Exact platform: confirm the 192GB Ryzen AI Max+ PRO 495 configuration rather than assuming every “Ryzen AI Max” system has the same memory.
- Memory allocation: verify the supported VGM setting, operating-system requirements and remaining memory for the CPU and applications.
- Software chain: confirm the model format, ROCm or Vulkan path, driver, runtime, frontend and required kernels.
- Complete workload: record the exact files, context, cache, batch or concurrency, modality and background applications.
- Measured outcome: capture load success, time to first token, generated tokens per second, peak shared-memory use, power mode, temperature and output correctness.
Bottom line
Ryzen AI Halo 192GB is primarily a capacity-first local-AI platform. Its documented advantage is one large shared-memory pool on an x86 system with Linux and Windows support. Its published bandwidth and compute specifications do not establish that it is faster or better value than a smaller CUDA system or an Apple unified-memory computer.
Choose it when the intended workload is verified on the software stack and crosses the capacity available on a less expensive alternative. Use the local AI hardware recommender to estimate memory, compare neighboring systems in the hardware guide, and build a reproducible benchmark record before treating a vendor demo as a general result.
Primary sources
- AMD Ryzen AI Halo developer-platform overview — platform positioning, specifications and vendor performance disclosures.
- AMD: Meet the next Ryzen AI Halo — 192GB launch details, graphics-memory allocation and 300B-class claim.
- AMD Ryzen AI Halo user guide — 128GB and 192GB system specifications and configuration guidance.
- AMD Ryzen AI Max+ PRO 495 product page — processor, graphics, memory, bandwidth, power and system specifications.