After announcing the Titan Z dual-GPU graphics card back in March, NVIDIA (NVDA +1.90%) has now released the $3,000 behemoth. While the card is marketed as the ultimate gaming graphics card, the price tag makes it completely uncompetitive for gaming, especially compared to Advanced Micro Devices' (AMD +0.73%) recent Radeon 295X2, and this leads to the question of what NVIDIA is trying to accomplish. The real target appears to be the GPU computing market, and if that's the case, then the Titan Z actually makes plenty of sense.
Source: NVIDIA
An expensive gaming behemoth
The Titan Z is certainly one of the fastest graphics cards for gaming, and early benchmarks have the card holding its own against the Radeon 295X2 for gaming at 4K resolutions. But the 295X2 costs $1,500, half the price of the Titan Z, while two Titan Blacks, of which the Titan Z is comprised, costs a combined $2,000. The Titan Z carries a 50% premium over simply buying two Titan Blacks, and since the Titan Z actually runs at a lower clock frequency compared to a single Titan Black, that premium ultimately delivers lower performance.
This performance deficit is balanced out by a significant reduction in power usage. The Titan Z uses 375 watts at peak load, compared to 500 watts for both the 295X2 and two Titan Blacks. For gaming, this doesn't matter all that much, especially since price is no object for those considering spending thousands of dollars on a graphics card. The 295X2 needs a higher-end power supply and will cost more to run, but that's peanuts compared to the price tags of these cards. For gaming, the Titan Z makes little sense.
But for computing, it makes a lot of sense
For certain applications, graphics cards can provide an enormous boost in performance compared to CPUs. Scientific simulations are one such area, and NVIDIA's Tesla line of GPUs are aimed at that market. These cards are extremely expensive, with the top-of-the-line Tesla K40 retailing for more than $5,000 -- although, there are likely volume discounts. NVIDIA has been successful in getting its GPUs into the world's fastest supercomputers, with 17 of the top 100 supercomputers in the world powered by NVIDIA hardware as of November of last year. In contrast, AMD GPUs power just one of these supercomputers.
GPUs used for computing have different requirements than those used for gaming. While single-precision floating-point math is adequate for gaming applications, double-precision floating-point math is often required for serious computing work. It's no surprise, then, that gaming GPUs focus on maximizing single-precision performance, while those used for computing need strong double-precision performance as well.
Based on benchmarks for the Titan Black, the Titan Z is going to have off-the-charts double-precision performance for what is being marketed as a gaming card. Benchmarks based on Folding @ Home, which simulates protein folding, put the Titan Black's double-precision performance at more than double the AMD Radeon 290X and nearly 55% higher than the AMD Radeon 295X2. Even if scaling is poor using the Titan Z, it should still more than double the double-precision performance of the 295X2.
It's clear that the Titan line of GPUs from NVIDIA is aimed at the computing market, because there's no reason to have these levels of double-precision performance otherwise. In fact, the Titan Z has 2.66 TFLOPS of theoretical double-precision performance, nearly twice the 1.43 TFLOPS of the Tesla K40 and the 1.48TFLOPS of AMD's flagship FirePro S10000.
The Titan Z does lack some features of the Tesla, namely ECC memory. This type of memory is capable of detecting and correcting memory corruption errors, and it's widely used in applications such as scientific and financial computing, where data corruption is unacceptable. ECC memory adds a premium, and because the Titan Z lacks it, there's little threat that Tesla GPUs will be supplanted. However, for computing applications where ECC memory isn't important, the Titan Z is a bargain.
The bottom line
While the Titan Z isn't competitive in the gaming GPU market, its blazing-fast, double-precision computing performance makes it a formidable entry in the GPU computing market. The marketing around the product is strange, given that it's in the GeForce line of gaming GPUs when that's clearly not the target market; it seems that the line between high-end gaming GPUs and computing GPUs is blurring. While NVIDIA has produced a laughably priced high-end gaming GPU with the Titan Z, it has also produced a bargain-priced supercomputer.
