Nvidia unveils Maxwell: A supremely power-efficient GPU architecture
The GPU arms race is usually about frames per second: Which video card can churn out triple-A games such as Battlefield 4 at the fastest clip? So when AMD and Nvidia announce a new GPU architecture, they often lead with their top gun and fill in the lower-priced slots down the road. Nvidia is taking a different tack with its all-new Maxwell architecture.
Maxwell will eventually succeed the Kepler architecture that serves as the foundation for Nvidia’s most powerful video cards, including the almighty GeForce GTX Titans that power the Titan Supercomputer at the Oak Ridge National Laboratory in Tennessee. But the first Maxwell-based GPUs will appear in the modestly priced GeForce GTX 750 (suggested retail price of $119 with a 1GB frame buffer) and GeForce 750 Ti (suggested retail price of $149 with a 2GB frame buffer. SKUs with 1GB of memory will be available later this month for a suggested retail price of $139).
Like Kepler, Maxwell is a massively parallel architecture consisting of hundreds of CUDA cores (512 in the GeForce GTX 750; 640 in the GeForce 750 GTX 750 Ti). Maxwell is based on the same 28nm manufacturing process as Kepler, but Maxwell is many times more power-efficient. It seems Nvidia’s engineers discovered that in Kepler’s design, many CUDA cores are sometimes idle during any given compute operation. They continue to consume power, even though the tiny microprocessors aren’t doing anything useful, because the GPU’s control logic can’t turn off them individually.
Taking a number of design cues from Nvidia’s Tegra engineering team (Tegra is a System on a Chip, or SoC, designed for mobile devices), Maxwell’s designers tweaked the Kepler design by breaking its monolithic core logic into a number of independent components and having each of those control a small number of CUDA cores. Now when an individual processor is idle, the control logic managing it can power it down. Nvidia says this more precise control improves performance per core by 135 percent, and it doubles the GPU’s performance per watt.
Thanks to this new power-efficient design, a video card based on the GeForce GTX 750 Ti has a TDP (thermal design profile) of just 60 watts. Since a PCIe slot can provide up to 75 watts of power, a GeForce GTX 750 Ti doesn’t require a supplementary six-pin power cable. That means just about any budget desktop PC that currently relies on integrated graphics is now a candidate for a discrete GPU upgrade.
As an example, Nvida cites a generic desktop PC powered by an Intel Core i5 with an integrated Intel HD Graphics 4600 processor as being capable of delivering Batman: Arkham Origins at 8 frames per second (with resolution of 1920 by 1080 pixels, 4X antialiasing, and visual quality at medium). Drop a GeForce GTX 750 Ti in that box, and Nvidia says you’ll see frame rates jump to 53 fps.
A video card based on Nvidia’s reference design for the GeForce GTX 750 Ti will run its 640 CUDA cores at a base clock speed of 1020MHz (with a boost clock of 1085MHz). The chip’s memory interface is 128 bits wide. Video cards based on this GPU will have at least 2 dual-link DVI ports and one mini HDMI; DisplayPort 1.2 is optional. (Note: DisplayPort is a requirement for Nvidia’s G-Sync gaming display technology.)
A card based on Nvidia’s reference design for the GeForce GTX 750 will have the same outputs on its mounting bracket. It will also have the same base and boost clockspeeds as the GeForce GTX 750 Ti, but because it has fewer 128 fewer CUDA cores, its TDP is just 55 watts. Nvidia claims the GeForce GTX 750 will deliver three times the performance of its entry-level GPU, the GeForce GTX 450.
Third-party board manufacturers will have the freedom to overclock both GPUs, and we could see passively cooled cards featuring either GPU. These early iterations of Nvidia's Maxwell architecture should be just the ticket for small form-factor rigs—including many of the Steam Machine designs we saw at CES in January.