AMD Siren Halo Offers Excellent News for Gaming Laptops, But Sets a Trap Against Apple and Nvidia’s RTX Spark

AMD Siren Halo Offers Excellent News for Gaming Laptops, But Sets a Trap Against Apple and Nvidia’s RTX Spark
3 min read

AMD’s next high-end laptop processor could integrate its own dedicated GDDR7 video memory, according to new leaks. This evolution would allow for gaming without a separate graphics card, but could undermine the unified memory that makes Ryzen AI Max chips appealing for running local AI models.

For nearly two years, AMD has sold its Ryzen AI Max processors on a simple principle: a single large pool of RAM, shared between the central processing unit (CPU) and the graphics chip (GPU). Its future mobile generation, however, may deviate from this scheme.

According to these leaks, the architecture dubbed Siren Halo would feature dedicated GDDR7 video memory, the same as that used in GeForce RTX 50 cards, in addition to standard system memory. The tip comes from leaker Gotou_3rd, who stated on X “Siren Halo has dedicated GDDR7 VRAM”. According to information corroborated by other sources, the chip would combine Zen 6 CPU cores, an RDNA 5 graphics part (from the “AT2” branch), and support for LPDDR6, the future low-power memory soldered onto motherboards.

Strix Halo (Ryzen AI Max 300) inaugurated this format in early 2025, before Gorgon Halo (Ryzen AI Max 400) pushed the same Zen 5 and RDNA 3.5 architectures up to 192 GB of unified memory. The Medusa generation (Zen 6) is on AMD’s official roadmap for 2027, and Siren Halo would be its successor for the Medusa Halo variant, though there is no official confirmation at this stage.

Bandwidth Multiplied by 2.5

Technically, the addition of GDDR7 would target the main bottleneck of these all-in-one chips: memory bandwidth, which determines the speed at which the graphics circuit retrieves its data. This throughput currently peaks at around 273 GB/s on Strix and Gorgon Halo, despite a 256-bit bus with LPDDR5X-8533.

With a 192-bit GDDR7 bus clocked at 28 Gbit/s, it would theoretically jump to 672 GB/s, nearly 2.5 times more. This calculation is based on the characteristics attributed to “Magnus,” the custom chip AMD is reportedly developing for the future Xbox (Project Helix) and which would share the AT2 graphics block with Siren Halo. The leaks credit this console processor with 68 compute units and a capacity of up to 48 GB of GDDR7, with no guarantee that the laptop version will adopt this configuration. Such bandwidth would allow AMD to surpass the Apple M5 Max, whose 40-core graphics version peaks at 614 GB/s.

A Trap for Local AI

Yet Ryzen AI Max chips appeal to tech enthusiasts and researchers precisely for their unified memory: the GPU can draw up to 96 or 128 GB at once. This block allows loading large language models (LLMs) with over 120 to 300 billion parameters locally, which consumer graphics cards, limited to 24 or 32 GB of VRAM, cannot accommodate.

However, Siren Halo would separate fast but limited-capacity GDDR7 (16, 24, or 48 GB) from slower LPDDR6, which risks eliminating this benefit. During inference, i.e., when a trained model generates its responses, speed is dictated by the slowest link: as soon as a model spills over from GDDR7 to system RAM, throughput plummets. In a context of sharply rising memory prices, soldering two different technologies onto the same board could also significantly increase the final cost of machines.

Competitors Taking the Opposite Path

This supposed reversal is all the more puzzling as AMD’s competitors are converging on unified memory. Since October 16, Nvidia has been selling its first laptops based on the RTX Spark platform (ARM N1X chip) in the United States, such as the Surface Laptop Ultra or the Asus ProArt P16, equipped with 128 GB of unified memory. Apple relies on the 128 GB unified memory of its MacBook Pro M4 Max and M5 Max, while Intel remains absent from this segment after abandoning the idea of succeeding the embedded memory of Lunar Lake.

One hope remains, however: that GDDR7 does not function as a sealed compartment, but as an ultra-fast cache, or a hybrid pool, supporting the unified memory, similar to 3D V-Cache on desktop processors. With the Medusa generation not expected before CES in January 2027 at the earliest, Siren Halo will only appear much later, and AMD is sharing its development with that of the future Xbox to amortize significant design costs.

S’abonner
Notification pour
guest

0 Commentaires
Le plus ancien
Le plus récent Le plus populaire
0
Nous aimerions avoir votre avis, veuillez laisser un commentaire.x