AMD puts 192GB of shared memory at the centre of its local AI pitch

AMD is pitching its upcoming Ryzen AI Max+ PRO 495, known as Gorgon Halo, as a way to run larger AI models locally by giving the system up to 192GB of shared memory. The most eye-catching benchmark claims are workload-dependent, but the memory capacity points to a real trade-off in local AI hardware: fitting a model can matter as much as raw GPU speed.

AMD Watch analysis

What happened

In a report published on 7 October, Igor’sLAB describes AMD’s new figures for a system with 192GB of memory, including up to 160GB configurable as graphics memory. AMD says the chip can run models with more than 300 billion parameters locally, but explicitly with 4-bit quantisation. This is not a claim that an uncompressed model of that size will fit.

AMD’s tests compare the system with an Intel Core Ultra X9 388H machine with 64GB of RAM and Arc B390 graphics. Across selected image, video, audio and 3D workloads, AMD claims results ranging from about 1.1 times to 32.2 times as fast. Read Igor’sLAB’s report.

Why it matters

A model that cannot fit in memory is not made usable by a faster GPU. Gorgon Halo’s shared memory gives AMD a potential advantage for running large quantised models and other demanding AI workloads on a local workstation, rather than renting cloud compute for every experiment.

But the biggest benchmark number is not a universal measure of GPU speed. Igor’sLAB notes that memory capacity and whether a workload fits can drive the gap; its report also says comparable measurements for Nvidia’s coming RTX Spark systems were not yet available.

Our read

The useful headline is 192GB, not 32.2 times. AMD is making a clear case for capacity over conventional graphics-card muscle, and that could matter to developers trying to run bigger models locally. The next test is less glamorous and more important: whether ROCm and the software people actually use make good on the hardware’s promise.

What to watch

• Independent comparisons with Nvidia’s RTX Spark systems.
• Which models and applications run well on Gorgon Halo in practice.
• Whether ROCm and framework support make the large shared-memory pool easy to use.

Discussion spark: For local AI, would you choose more memory to fit larger models, or a faster GPU with a more mature software ecosystem?

Independent WittyWires tracker for public updates about AMD. Not affiliated with or endorsed by AMD; this is not an official account.