AMD's RDNA 4 architecture-based Radeon RX 9000 series has been on the market for a year now, and in terms of the red factory's overall strategy, this generation is positioned at the mid-range, avoiding competition with NVIDIA's high-end. The flagship model, the RX 9070 XT, is positioned to compete with NVIDIA's GeForce RTX 5070 Ti, a card that belongs to NVIDIA's mid-range echelon, but is widely regarded as one of the most well-balanced graphics cards available today due to its performance and efficiency. The choice of product line for this generation suggests that AMD is avoiding a head-on collision with its rivals in the desktop GPU market for the time being, preferring to stabilize the mainstream market. However, signals from within the company suggest that the next generation of architectures may be a more groundbreaking attempt.

According to Laks Pappu, Senior Fellow and Principal System-on-Chip Architect at AMD, his work spans data center GPUs, cloud gaming products, and the Navi4x and Navi5x series of Radeon architectures. He mentions that he's "building the next generation of competitive 2.5D/3.5D chipsets and monolithic GPUs based on a variety of packaging technologies," a statement that could almost be taken as an indication that AMD is exploring the possibilities of co-existing multi-chip modules and monolithic designs. Pappu joins AMD from Intel in 2022, after a 25-year career at Intel that included solo projects like DG1, Alchemist, and Battlemage, during which time he also proposed multi-block GPUs. While Intel's dual-GPU Battlemage was more focused on AI than consumer graphics, this experience clearly provided the foundation for his next step at AMD. In the world of GPU design, there is a clear difference in thinking between consumer and data center. Datacenter GPUs are often disaggregated designs, such as AMD's Instinct MI300 or NVIDIA's Blackwell, which are packaged in multi-chip packages for scale and performance. But consumer GPUs require extremely low latency and high-bandwidth communications, and thousands of parallel threads must be precisely synchronized, making multi-chip designs extremely challenging. Cross-chip communication latency, increased power consumption, consistency maintenance, and how to disguise a multi-chip GPU as a single device at the driver and software level are all real challenges. In contrast, while CPUs can tolerate some latency, the graphics workloads of GPUs are more sensitive. As a result, multi-chip GPUs have so far been used primarily in HPC and AI gas pedals, and rarely in gaming graphics cards. However, as the size and cost of single-chip GPUs continues to rise, this barrier may be being re-evaluated. Building large single-chip GPUs is not only expensive to manufacture, but wafer yields are also declining, while multi-chip solutions can improve yields to some extent and allow for flexible product configurations. AMD itself was one of the pioneers of the chiplet architecture, pioneering the use of multi-chip modularity in Ryzen and EPYC processors, and experimenting with it in the Navi 31 in the Radeon RX 7900 series. AMD itself was one of the pioneers of the chiplet architecture, pioneering the use of multi-chip modularity in Ryzen and EPYC processors, and experimenting with it in the Radeon RX 7900 series with the Navi 31: a single core graphics chip paired with six cache/memory chips, a sort of "proving ground" for GPU decomposition. This allows AMD to derive a variety of products with different positioning from a single architecture, and if we can solve the problem of multi-chip logic splitting and software layer integration in the future, the multi-block design of consumer GPUs will no longer be unimaginable.

Pappu's background suggests that he may be leading the Navi 5x architecture from the ground up. The typical GPU development cycle is about a year for architecture definition and planning, and 12 to 18 months for physical implementation, followed by flow and validation. Considering that the RDNA 4 architecture will have been finalized by the time he joins AMD in 2022, it is likely that RDNA 5 will be the first consumer GPU architecture over which he has full control. By August 2025, RDNA 5 should have completed the architecture and RTL phases and be in the flow or early silicon validation phase. At this rate, RDNA 5 will likely launch in late 2026 or early 2027, in line with AMD's two-year, one-generation cadence. The value of multi-chip GPUs is that they may become the natural choice to continue performance growth as gaming graphics cards approach manufacturing limits, and NVIDIA's high-end GPUs, which are already in the tens of billions of transistors, continue to escalate in cost and power consumption, could be the first to deliver true multi-chip GPUs to the consumer market, if they can solve the issues of latency and power efficiency in a 2.5D or 3.5D package. Until then, the Navi 31 design may be an initial test, but the Navi 5x could be a real game-changer. From a market perspective, AMD has chosen to hold onto its mainstream positioning in the Radeon RX 9000 series, avoiding a direct confrontation with NVIDIA's RTX 5090 and 5080 in the high-end segment. However, this strategy has caused AMD's share of the desktop graphics card market to continue to decline, and NVIDIA's lead has further expanded. To change the situation, AMD needs to come up with sufficiently differentiated technical solutions in its next-generation products. Whether it's the multi-chip design of RDNA 5, breakthroughs in packaging and interconnect technology, or optimization at the software level, the ultimate goal is to bring Radeon back into the arena of head-to-head competition with its rivals. AMD's next-generation graphics products are at a critical juncture. If RDNA 5 succeeds and introduces a multi-chip design, it will not only potentially improve the economics of GPU manufacturing, but also bring a new way of increasing performance to the consumer market. In the coming months, the outside world will likely see more early information about Navi 5x as engineering sample testing progresses, and this will determine whether AMD can turn the tables on the graphics competition in 2026-2027. For industry watchers and gamers alike, this will be an exciting step in the evolution of GPUs.