AMD could develop an RDNA 3 architecture GPU to compete with the RTX 4090 but gave up after measuring cost and power consumption
Recently, AMD Radeon Technology Division Senior Vice President of Engineering and Development, David Wang, and AMD Computing and Graphics Business Division Executive Vice President, Rick Bergman, gave an interview to ITMedia. They answered questions about RDNA 3 architecture GPUs and ray tracing development and shared their future GPU strategies.
Last November, AMD released the Radeon RX 7900 series graphics cards based on the RDNA 3 architecture and equipped with Navi 31. Although the new generation of graphics cards has better performance, it clearly cannot compete with Nvidia’s GeForce RTX 4090. Many players were disappointed and did not understand why AMD did not develop more powerful GPUs.
Rick Bergman stated that AMD could indeed develop an RDNA 3 architecture GPU that competes with the AD102 specification, but after considering the multiple factors such as the power consumption, heat output, and high cost of the entire card, they chose to abandon this strategy. In AMD’s view, a $999 price point is already the upper limit for the majority of high-end PC gamers to purchase a graphics card. Another reason for adopting this strategy is to enable new graphics cards to be used with existing conventional power supplies and cases without requiring installation in oversized cases.
Some players have also pointed out that although Ryzen is designed for gamers, there is also Ryzen Threadripper for the highest-end users, and it seems that AMD has not considered cost and power consumption issues in that specific area. Rick Bergman believes that in his view, the insane performance of the Ryzen Threadripper is no longer a CPU for ordinary users and gamers. Furthermore, after the introduction of 3D V-Cache technology to mainstream Ryzen, it already meets the needs of high-end PC gamers.
Rick Bergman believes that the performance range, development cost, and manufacturing cost of high-end GPUs are very different from the past. Currently, AMD focuses on planning GPUs with a balance between performance and cost at a $1000 price point. David Wang stated that in fact, from the development of the CDNA series architecture, AMD knows that it is not impossible to develop GPUs that target the ultra-high-end market. It’s just that this type of GPU is not suitable for ordinary consumers in AMD’s view.
Regarding future technology, David Wang said that although future GPUs need to implement the small-chip architecture, AMD has decided to delay the implementation of true multi-chip design for RDNA 3 architecture GPUs. Navi 31 is AMD’s first consumer-level GPU with MCM multi-chip packaging, equipped with one GCD and six MCDs, and is still a single integrated package, not like Ryzen or EPYC, which can be customized by increasing or decreasing the number of CCDs according to performance needs.
In David Wang’s view, small chip architecture can truly demonstrate its value for high-performance chips while reducing manufacturing costs. Since GPUs have numerous computing cores, if a CPU-like connection method is used, it is likely to be unable to guarantee the reliable transmission of electrical signals. Therefore, this approach was rejected in the current RDNA 3 architecture.
Rick Bergman added that in the semiconductor field, the trade-off between “performance” and “manufacturing cost” must always be considered before making decisions. AMD’s challenges in multi-chip packaging will continue, and the future can be expected.
David Wang believes that we are currently in an era where traditional rasterization and ray tracing coexist. As we transition, AMD will undoubtedly focus more on improving the ray tracing performance of GPUs and accelerate the pace of improvement. The processing cost of ray tracing technology is very high, and with current processing capabilities, it is impossible to perform real-time ray tracing processing on all elements of game graphics. In the foreseeable future, traditional rasterization and ray tracing.
