Products You May Like
For the past few months there’s been a constant drip-drip-drip of rumors about the upcoming GPUs from AMD and Nvidia. The most surprising element of those rumors, at least on the Nvidia side of the fence, has been a dramatic increase in power consumption. Nvidia’s next-gen Ada Lovelace RTX 40-series were originally reported to require almost double the power of its current Ampere cards. Those numbers have since come down a bit, but the latest rumors still hint at a 600W flagship GPU.
That’s hard to even fathom. On the AMD side of the rumor mill though, things have been much more quiet. We know the company is moving to a chiplet-based GPU design for the first time, but will AMD be cranking up power levels too? According to an AMD engineer in a recent interview, yes, it will. By how much, he doesn’t say, only that this is the new trend.
The insight was revealed by AMD’s Sam Naffziger in an interview with Tom’s Hardware. He’s been with the company 16 years and is a Senior Vice President, Corporate Fellow, and Product Technology Architect. He has been working on the company’s chiplet designs for years now, which were formerly reserved for its Zen, Epyc, and Threadripper CPUs. With RDNA3, that design will now be applied to GPUs for the first time. Naffziger explains that it’s not a binary question of whether to make a chiplet power efficient or more powerful. Though those two goals are at odds with each other, Naffziger explains that improving efficiency gives AMD more headroom to maximize performance too.
This slide depicts the steadily decreasing rate of node advancement and silicon area scaling. (Image: AMD). The limited scaling depicted for analog and SRAM is part of an overall trend of diminishing marginal returns. AMD and Intel have adopted chiplet-based designs partly in response to this trend.
When Naffziger was asked whether RDNA3 would follow Nvidia’s rumored tactic of increasing power levels, he replied that AMD really has no choice but to follow the same path. “Performance is king, but even if our designs are more power-efficient, that doesn’t mean you don’t push power levels up if the competition is doing the same thing. It’s just that they’ll have to push them a lot higher than we will.” He was also asked what is the main impetus for graphics cards needing a lot more power all of a sudden, and he said it’s due to twin forces of soaring compute demand and shrinking gains in node advances.
“It’s really the fundamentals of physics that are driving this,” he said. “The demand for gaming and compute performance is, if anything, just accelerating, and at the same time, the underlying process technology is slowing down pretty dramatically — and the improvement rate. So the power levels are just going to keep going up. Now, we’ve got a multi-year roadmap of very significant efficiency improvements to offset that curve, but the trend is there.”
AMD hasn’t revealed much about RDNA3 just yet, but it has put a few numbers in writing. One of those is that it expects a greater than 50 percent improvement in performance-per-watt. That’s a huge leap, but it’s what it achieved in previous generational transitions. That was going from Polaris/Vega to RDNA, then to RDNA2. However, those were back in the heady days of GCN 14nm for Polaris, and TSMC 7nm for RDNA. AMD will be tapping TSMC’s 5nm node for RDNA3, but little else is known about power requirements just yet.
Twitter user AMDGPU’s mockup of the Navi 31 package. (Image: @AMDGPU_)
All that’s been alleged so far is that the company’s RDNA3 GPUs will feature a seven-chiplet design. The company has announced it’s “rearchitected” its compute unit. We also know it will offer a “next-gen” version of its Infinity Cache. Still, Naffziger was clear that although it will be increasing power levels, it probably won’t be going to same extremes as Nvidia. This tracks with its recent bragging about its superior efficiency compared to the company’s Ampere GPUs. Obviously, this is a mantle AMD is keen to hold onto in the future.
Now Read: