Sam Naffziger, AMD senior vice president, corporate fellow and product technology architect, took part in a wide-ranging interview with VentureBeat recently. Interestingly, AMD sees high-performance GPUs with TDPs as high as 700W emerging before we reach 2025. However, Naffziger’s specialism is in leading efficiency technologies, and with the success of 25x20 behind him, he is currently working hard on the ambitious 30x25 efficiency project.
Naffziger has been at AMD for 16 years, and has been embedded within the graphics division since 2017. He had some very interesting things to say about boosting the efficiency of Radeon GPUs via developments in the RDNA architecture, as well as the hand-in-glove development of the CDNA architecture for the data center.
The above chart isn’t specifically discussed in the interview, but making GPUs more efficient is. With this lack of clear context, we are seeing the chart as an indicative warning about the industry by AMD, rather than a prediction of the direction of travel for AMD GPUs. Naffziger reveals some of the innovative ways in which AMD has avoided seeing its GPU TDPs balloon, and from the interview it sounds like AMD still has some tricks up its sleeves for RDNA 3 and RDNA 4.
Much has been written about AMD’s transition from RDNA 1 to RDNA 2; it happened with the introduction of the Radeon RX 6000 series, of course. In the VentureBeat interview, Naffziger provides some interesting background to this generational change. He says that the “doubling of performance and a 50% gain in performance-per-watt” was largely due to taking things learned from CPU design to get the 2.5GHz+ clocks at modest voltages. AMD also took a bet on reducing bus widths and using big caches – specifically Infinity Cache – which was turned out to be “a great fit for graphics.”
For RDNA 3, Naffziger asserts that “We’re not going to let our momentum slow at all in the efficiency gains.” This means AMD is targeting a 50% performance-per-watt improvement. AMD will continue to use its CPU know-how in optimizing the potential of its GPUs, and this will include chiplet designs. RDNA 3 GPUs will leverage 5nm chiplets later this year. Naffziger notes that Intel is already big in chiplets + GPUs, as evidenced by Ponte Vecchio, but Nvidia has shown no signs of a jump yet. Talking about competitors, the AMD fellow added “Our competitors either have good CPUs or good GPUs, but nobody has both, at least not yet.”
For more information about the upcoming Radeon RX 7000 series GPUs, please check our RDNA 3 deep dive article based on information from the AMD financial analyst day this June.
