Will the Real Titan Xp Please Step Forward?
WHEN NVIDIA RELEASED the Pascal version of the Titan X last August, it created an overlap on the Titan X name. That caused confusion, so instead of calling the new card the Titan X (Pascal), the unofficial name became Titan XP.
Now we know why Nvidia didn’t use that name in the first place, because following the launch of the GTX 1080 Ti, which boasts slightly higher performance than the Titan X (Pascal), we have the new Titan Xp. And unlike all the earlier cards, the Titan Xp comes with a fully armed and operational GP102 GPU.
As for specs, the Titan Xp isn’t just a full GP102 processor with 3,840 CUDA cores. It has 12GB of GDDR5X memory, and the memory runs at an impressive 11.4Gb/s. The GPU clock is also higher than the Titan X, matching the 1080 Ti. Combined with the extra 1GB of memory and an extra 32 bits on the memory interface, the Titan Xp has 13 percent more total memory bandwidth and 7 percent more computational power than the 1080 Ti. It also retains the $1,200 price point of the Titan X (Pascal).
I suspected something like this would happen ever since the Titan X (Pascal) first launched, because it had two SM clusters disabled. This is in clear contrast to the Titan X (Maxwell), which has a fully enabled GM200 GPU, looking more like the Titan (Kepler), with its partially disabled GK110 core. Nvidia left room for the Xp by releasing the GTX 1080 Ti with 11GB of memory and a 352-bit bus, still with a partially disabled GP102 chip. We end up, then, with a repeat of the 780, Titan, 780 Ti, Titan Black rollout of the Kepler days—without the added bonus of fast FP64 support on Titan.
The more interesting aspect of this pattern is that it points to the slowing down of Moore’s Law. It would be awesome to see a completely new graphics architecture every year, but that’s not sustainable, and we don’t get new process nodes every other year either. The smart business approach is a slow rollout of new parts to keep things fresh, which is exactly what Nvidia has done with Pascal (and Maxwell and Kepler).
The GP100 was announced for Tesla P100 in April 2016 (and the Quadro GP100 as of February 2017); it probably won’t ever be a consumer chip, because of the cost of HBM2 and the added FP64 support. While technically second, the GP104 launched in May/June with the GTX 1080/1070, and was the first widely available Pascal implementation. The modestly priced GP106 (GTX 1060 6GB/3GB) then followed in July/August, which also happened to coincide with the GP102 showing up in the professional Quadro P6000. The GP107 (1050/1050 Ti) was released in October, rounding out the budget end of the 10-series parts, and now the GP102 is in the 1080 Ti and Titan Xp.
In other words, Nvidia stretched what would have once been a top-tobottom launch into a rollout that has spanned the past year. That should carry us to the next-generation Volta architecture, which will still use a 16nm FinFET process. Count on another year for the full Volta rollout (Titan TV coming in spring 2019— you read it here first!), and Nvidia should successfully fill the void as we wait for 10nm and 7nm to come online.
There’s a second aspect to the Titan Xp and 1080 Ti worth mentioning, and that’s the pending arrival of AMD’s RX Vega. We should have that in our labs soon enough, but based on early information, I don’t expect Vega to be able to beat the 1080 Ti, let alone the Titan Xp. Nvidia is very good at staying on top, and 1080 Ti and Titan Xp are supposed to do just that. I’d love to see an upset, but AMD’s RX 500 cards are slight tweaks to the existing Polaris offerings, and Vega will need to see AMD through the next year. It’s the graphics card equivalent to Intel’s ticktock model: high-end GPUs in the spring, midrange in the fall, and new architectures every two years.
The smart business approach is a slow rollout of new parts to keep things fresh.