AMD Ryzen AI Halo offers up to 192GB of memory

AMD lists up to 192GB of unified LPDDR5X-8533 memory for the updated Halo and says users can allocate up to 160GB of system memory to graphics.[2] The company says the system supports local language models with 300 billion or more parameters at 4-bit quantization.[2] That qualification matters when assessing model fit: AMD’s statement ties the parameter figure to a specific quantization level, rather than claiming the same capacity for every model configuration. AMD also says developers can run multiple large agents at once with separate KV caches.[2]

AMD identifies the Ryzen AI Max+ Pro 495 as the Halo’s flagship APU.[2] The chip has 16 CPU cores and 32 threads, a maximum boost clock of 5.2GHz, and an NPU rated for up to 55 TOPS; AMD rates overall AI performance at up to 132 TOPS.[2] HotHardware reports an MSRP starting at $7,299, while AMD’s announcement identifies Micro Center as a retailer where the Halo is available.[2][3] Those figures give buyers a price and channel to check against their workload requirements.

AMD says the Halo achieved higher relative throughput than an Apple M5 Pro comparison system in its tested generative AI workloads.[2] AMD’s published comparisons reported up to 34.3 times higher throughput than an Apple MacBook Pro with M5 Pro and up to 32.2 times higher than an Intel Core Ultra X9 388H system.[4] These are AMD-reported results, not a guarantee for every workload: AMD’s comparison included workloads the Apple system could not run, and TechTimes says several workloads in the 18-workload comparison were marked as not running on both comparison systems.[2][4]

The comparison claims are not the only performance data in the evidence. TechTimes reports that independent benchmarks found the Halo generating tokens roughly 30% more slowly than NVIDIA’s DGX Spark, which has 128GB of memory.[4] The measures and test conditions behind that result are not specified in the evidence here, so it should not be treated as a direct reconciliation of AMD’s throughput figures. AMD’s stated rationale for agentic AI is that these workflows need CPU performance, accelerated compute, and enough memory for models, applications, and working data.[2]