AMD is welcoming NVIDIA's entry into the large-local-memory AI accelerator space, but the company insists its upcoming Gorgon Halo APU still holds the edge. The chip, also known as Strix Halo, targets AI workloads that demand massive unified memory. AMD executives argue that Gorgon Halo's memory capacity and compute capabilities surpass what NVIDIA's RTX Spark offers.
Gorgon Halo combines Zen5 CPU and RDNA 3.5 GPU
Gorgon Halo is a high-end APU designed for laptops and compact AI systems. It combines a 16-core, 32-thread Zen5 CPU with a Radeon 8065S GPU based on the RDNA 3.5 architecture. The GPU packs 40 compute units, making it one of the most powerful integrated graphics solutions on the market.
Specifications
- CPU Cores/Threads: 16 cores / 32 threads
- GPU Architecture: Radeon 8065S, 40 CU RDNA 3.5
- Max Unified Memory: 192 GB
- Max GPU Dedicated Memory: 168 GB
- PassMark Single-Core: 4293
The APU supports up to 192 GB of unified memory, with up to 168 GB allocable as dedicated GPU memory. This exceeds the 128 GB limit of NVIDIA's RTX Spark. According to AMD, large local memory is critical for agentic AI workloads that require holding large models and data locally.
ROCm software ecosystem eases CUDA migration
PassMark benchmark scores for the Ryzen AI MAX+ 495, the top SKU, show a single-core score of 4293 and a multi-core score of 57525. These figures indicate strong CPU performance for both general and AI tasks. AMD also highlights that its ROCm software ecosystem has become more developer-friendly, reducing the cost of migrating from CUDA.
Gorgon Halo is expected to launch globally in Q3 2026. Pricing has not been announced. AMD executives have expressed confidence in the product, with one stating that not buying a Strix Halo laptop would be a pity.
AMD's Rahul Tikoo noted that for nearly two years, AMD was almost the only player in the large-local-memory AI accelerator field. The company welcomes NVIDIA's RTX Spark as validation of the market. Gorgon Halo aims to maintain AMD's lead by offering higher memory capacity and a mature software stack.



Discussion
0 comments
Log in to join the thread with a thoughtful take, question, or correction.