Video Cards and the Ever-Increasing Memory Power Consumption Adventure

[bp9801]

Video cards and memory go hand in hand, with the more power a card has, the more memory it carries with it. Memory keeps on increasing as the power of the cards reach new heights, and even more so with HBM jumping into battle. However, there is one thing that's an issue regardless of GDDR or HBM: power consumption of the memory. Video card TDP typically tops out at 250W for NVIDIA and 275W for AMD, but jumps up higher on the dual-GPU monsters like the TITAN Z (375W) and R9 295X2 (500W). The memory TDP ranges from 8W to 32W on the NVIDIA side and 10W to 36W on the AMD side, with those dual-GPU cards again jumping things to 60W on the AMD side (curiously NVIDIA's TITAN Z still is at 32W memory TDP). The percentage totals of memory TDP to total TDP is anywhere from 8.4 to 14.5% for NVIDIA and 5.3 to 15.3% for AMD. It may not seem like much, but when the bandwidth starts to increase we can run into power problems and huge demands.

Right now, the bandwidth for NVIDIA tops out in the 336GB/s range, with AMD hitting 512GB/s with HBM. NVIDIA plans to move to HBM2 in its next line of cards, bringing even more power (and higher bandwidth) than it has now. When the bandwidth jumps into the thousand and beyond GB/s area, things can get quite dicey when it comes to memory power consumption. We could see cards with the memory alone pulling in 100W of power, shoving the card's total TDP beyond the 250-275W mark. HBM is more economical than GDDR5, as evidenced by the Fury cards featuring a 15W memory TDP despite the 512GB/s bandwidth. Scaling up from there, an HBM card pulling 120W of memory TDP would feature a staggering 4200GB/s bandwidth. Things don't always scale linearly, so if we put the bandwidth around 20-25GB/s per watt, we're looking at a (still impressive) 2400GB/s bandwidth. Of course this is all based on current trends and efficiency, so things can change over time as move further into the future.

So, where does this leave us? Well, NVIDIA and AMD can work on power efficiency with HBM to improve its capabilities. Intel, Rambus, and Micron are all working on new technologies, too, with new standards of their own that can really kickstart the memory power consumption numbers. Right now bandwidth and the memory power consumption will hit a wall, where either the bandwidth reaches a certain level and companies refuse to take it higher to save on power, or we see high-end video cards pushed to the absolute limit and feature TDPs in the 400+W range. We can also see huge advances in memory power consumption, either with HBM2 and beyond or in a new technology.

Source: WCCFtech

Back to original news post