NVIDIA is shaping up to be the most consistent chipmaker in the industry when it comes to per-generation performance and energy efficiency gains. Over the past three generations, spread across the past four years, the company dedicated vast R&D resources to developing increasingly more energy efficient GPUs. The 2012 "Kepler," followed by the 2014 "Maxwell" architectures, provided performance and energy efficiency leaps that do the word "leap" justice. At its recent GeForce 1080 "Pascal" unveiling, the company claimed to shatter all previous performance-to-wattage records, promising another leap in performance and efficiency.
The GeForce "Pascal" architecture debuted this April at NVIDIA's GTC event, driving the company's Tesla P100 HPC processor based on the massive GP100 silicon. The architecture's consumer-graphics debut was earlier this month, in the form of the GeForce GTX 1080, reviewed today, and the GeForce GTX 1070. The GTX 1080 will hit shelves on May 27th, with the GTX 1070 following closely on June 10th.
The GeForce GTX 1080 logically succeeds the GeForce GTX 980 and is priced not too far apart from its predecessor. It will launch at $599 where the GTX 980 debuted at $549. NVIDIA, however, is positioning it as a premium part since it's supposedly faster than any currently available single-GPU card. The company has hence come up with a clever marketing tactic.
Historically, reference-design cards set the baseline launch price for SKUs. This time around, NVIDIA is making its partners sell the reference-design card at $699 (higher than the launch prices of the GTX 980 Ti, GTX 780 Ti, etc.). The company is sub-branding this SKU as the "Founders Edition." Since this also happens to be the only GTX 1080 card designed entirely by NVIDIA, it's the card we are reviewing today. Custom-design cards will start at $599, but this is really a "suggested" price by NVIDIA to its partners. Adding meaty cooling solutions and custom VRM designs to the mix could easily push prices way above the $599 mark and perhaps even close to the $699 "Founders Edition" price, which could have partners take that as the baseline instead.
The GeForce GTX 1080 is based on NVIDIA's "Pascal" architecture. This architecture sees the streaming multiprocessors (SMs), the indivisible subunits of an NVIDIA GPU, get even more dedicated components, which increases their performance. NVIDIA claims to have "meticulously" designed the GPU architecture to be as energy efficient as possible given the silicon fab node and is leveraging the 16 nm FinFET node at TSMC for "Pascal."
The GTX 1080 features more CUDA cores than its predecessor – 2560 vs. 2048. It features even more TMUs (160 vs. 128) and, at 8 GB, double the memory. Memory technology sees a major update with NVIDIA's adoption of the GDDR5X memory standard. The memory is clocked at a staggering 10 GHz effective, at which speed the GPU has 320 GB/s of memory bandwidth over a 256-bit wide memory interface. Such bandwidths were only possible with 384-bit or 512-bit GDDR5 memory interfaces. First-generation HBM being restricted to 4 GB, HBM2 not being readily available, and HBM-class stacked memory being more expensive to deploy (using on-chip silicon substrates) could have contributed to NVIDIA's decision to go with GDDR5X for the GeForce GTX 1080.
In this review, we're testing a GeForce GTX 1080 reference-design graphics card, which NVIDIA is marketing as the "Founders Edition" at $699. This will be the price we will use in our primary price/performance calculation, although we did add a $599 price/performance data point as a reference.