Post

Ahead of the Computex trade show a bit later this month AMD is getting the ball rolling early with the announcement of a new series of Ryzen AI Max processors. Joining the existing Ryzen AI Max 300 series of chips, AMD Ryzen AI Max PRO 400 family will be continuing the Strix Halo legacy with a minor performance bump and a major memory capacity bump – which will see the chips paired with up to 192GB of RAM. Systems based on the new Ryzen AI Max chips will begin shipping in the third quarter of this year.

Diving right in to the hardware, the AMD Ryzen AI Max PRO 400 family is an additional set of chip SKUs based around AMD’s Strix Halo silicon (or as AMD calls the chips in this family, Gorgon Halo). This means we are looking at a high-performance SoC combining up to 16 Zen 5 CPU cores with AMD’s powerful RDNA 3.5 architecture integrated GPU, and feed by a sizable 256-bit memory bus. The combination of the high-performance iGPU and LPDDR5X-based memory subsystem has made Strix Halo a chip to contend with in the AI space over the last year, and AMD is looking to get a bit more out of the silicon over the coming year with a mid-generation refresh.

Compared to their Ryzen AI Max 300 counterparts, the Ryzen AI Max PRO 400 chips are a rather minor spec bump in terms of chip clockspeeds and core configurations – in fact most of the SKUs are unchanged in this regard. Only the top-end flagship SKU, the Ryzen AI Max+ PRO 495, offers meaningfully different specs in regards to computational throughput, getting a small boost to clockspeeds for its CPU, GPU, and NPU. All together, these bumps bring it up to 5.2GHz on the CPU while the GPU gets the new Radeon 8065S branding.

Instead, the big change with the new chips is with memory support: AMD has dialed up memory clockspeeds a bit and greatly increased the the total amount of supported memory, which was one of the defining features of the Ryzen AI Max family to begin with. While the 300 series chips supported a maximum of 128GB of LPDDR5X-8000 memory, the new 400 series chips are boosting that by 50%, bringing the total to 192GB of LPDDR5X-8533 memory. With the recent advent of 24GB(192Gbit) LPDDR5X memory chips, AMD is quickly taking advantage of the tech to boost the Ryzen AI Max family’s already spacious memory capacity to a new level, while boosting memory bandwidth by about 7% to 273GB/sec.

With the chips’ popularity in systems for local AI inference, AMD’s value proposition for a higher memory capacity is rather straightforward: a larger memory pool allows for these systems to be loaded up with even larger models. The chips can be configured to allocate as much as 160GB of their memory pool to the GPU (leaving the last 32GB for the CPU), which is enough space to load up a 300B FP4 parameter model – a first for any single SoC system that is not a Mac Studio (and likely not the last).

Users who are truly limited by memory capacity should find the extra 48GB of GPU memory a significant boon. Though based on what we have seen with the Strix Halo hardware up to this point, this change is not likely to move the needle on performance. As AMD is not meaningfully boosting the compute throughput or memory bandwidth of the chips, existing bottlenecks such as pre-fill performance are going to remain where they are.