Nvidia has released their latest Nvidia Titan X graphics card and together with its release, there are independent benchmarks and more details about the card’s capabilities. While named in a confusing way, similar to GeForce GTX Titan X (Maxwell), the new Titan X drops GeForce name, but is often referred as Titan X (Pascal) or less formally “Titan XP” for clarity.
Titan X is yet another fastest graphics card from Nvidia this year. Based on GP102 GPU, containing 11 billion transistors made on TSMC’s 16 nm process, it’s the largest GPU available for gamers (471 mm2). It is not as large as GP100 (610 mm2), only available in Tesla P100 compute card at the moment, but still considerably larger than GP104 (314 mm2) in GTX 1080. GP102 in Titan X is slightly cut down and has 3584 CUDA cores, running at 1417 MHz with 1531 MHz boost, leading to 11 TFLOPS in FP32 performance. On the memory side, GP102 has 96 ROP units connected over 384-bit bus to 12 GB of 10 Gbps GDDR5X, providing an impressive 480 GB/s bandwidth. Thus, based on specifications alone, there is 22% computational performance advantage and 50% memory size and bandwidth advantage over GTX 1080. Titan X is rated at 250 W TDP, which is the same as Nvidia’s previous Titan cards.
However, it is much more interesting how big the difference is in the actual games when they are not limited by CPU. As Pascal cards overclock dynamically based on power and temperature, there is often a headroom above the guaranteed boost clock. Any game scenario limited by memory bandwidth benefits even more from the much higher ROPs and bandwidth on the Titan X. At 2560×1440, Titan X usually is 30% ahead of GTX 1080, while at 4K (3840×2160) the gap goes closer to 40% as memory bandwidth becomes more important. The performance of Titan X is high enough to comfortably play such games as Battlefield 4, Hitman (2016), The Witcher 3, DOOM (2016), GTA V at 4K with average FPS exceeding 60 FPS at their highest graphics quality presets. Compared to the single GPU cards that had top performance before this summer, last 28 nm monsters GTX 980 Ti and Fury X, Titan X (Pascal) enjoys 80% performance advantage on average. While those 28 nm GPUs were already hitting the production size limit, there is clearly even more gaming performance that can be achieved during next few years on 16 nm as the process matures. Even though there are games where AMD enjoys considerable advantage with Vulkan or DX12 at the moment, the sheer power difference between Titan X (Pascal) and Fury X still leaves Nvidia’s newest monster at least 40% ahead even in AMD’s best case scenario.
In terms of what makes Titan special, Titan X (Pascal) seems to have the least compared to its predecessors. All previous Titans tripled the VRAM amount, compared to G*104 part, while Titan X only gets 50% increase corresponding to the bus width increase. Just like Titan X (Maxwell), there is no big benefit to FP64 (double-precision) compute performance. In the same way as GM200, GP102 simply cut those parts of the chip. While GP100 has FP16 (half-precision) at 200% compared FP32, FP16 performance in GP102 is just 1.5% of FP32 performance. Essentially FP16 and FP64 in Titan X are just for code compatibility, just like in other Pascal GeForce cards. The original Titan (Kepler GK110) had 1500 GFLOPS in FP64 performance, while Titan X (Pascal) only has 343 GFLOPS FP64. The only way to get a generation improvement in double performance from Nvidia is an extremely expensive (over $5000) Tesla P100, offering 4700 GFLOPS in FP64. Meanwhile, Nvidia’s latest productivity carrot is 44 TOPS INT8 (44 trillion operations per second with 8-bit integers) on Titan X (Pascal). While such performance is useful for the neural networks, it is not yet clear if the rest of Pascal GPUs have crippled INT8 performance, or just proportionally reduced compared to Titan X.
Titan X shares a lot of common features with other Pascal cards. GPU Boost 3.0 supports granular overclocking, allowing setting different frequencies for different voltages before hitting temperature and power limits. It can also be overclocked using automatic tools, but those are not yet able to handle more advanced overclocking failures like system crashes or freezes. Only 2-way SLI is supported in most games and the new high bandwidth SLI bridge utilising both SLI fingers is recommended for the best performance. 3-way and 4-way SLI is only supported for setting new records in the benchmarks. The new Pascal VR optimisations should benefit Titan X, when developers implement them in their games. On media side there is support for 4K 60 FPS HEVC (h.265) video decoding and encoding as well as HDR display support. Visually it has similar shape as GTX 1080 Founders edition, with angled design, but painted black.
Overall Titan X is a graphics card providing a top level performance for an enormous price. Sold only directly from Nvidia at $1200 and allowing a single customer to only buy 2 cards. Which is more expensive than $1000 launch prices for the older single-GPU Titans and with less features to sweeten the deal. If you have a lot of money and want the top performance as soon as possible, the latest Titan X has its appeal. For the most gamers looking to improve their performance with less damage to the wallet the waiting game seems more worthwhile. As soon as AMD manages to top GTX 1080 performance, Nvidia is very likely to respond with “GTX 1080 Ti” with very similar performance to the Titan X, but at more manageable price.