Yes folks, they did it again. But this time it`s even harder to understand than few month earlier with Titan Xp (Pascal). What is the problem ? Power consumption ? No. Availability ? Nah (only two per customer though – LOL). Compute power ? Definitely not, in fact Volta – brand new architecture – is the future. Yes, of course it`s the price. Previous Titan Xp was 1200$ and it was bonkers. This time.. I dont`t even believe in it at the time of writing – it`s 2999$.
Now I understand new production process, the use of futuristic HBM2 memory modules, the Nvidias “fools tax” for the possibility of calling yourself an owner of THE MOST POWERFUL PC GPU EVER CREATED but come on ! Almost 3000 bucks for a GPU is a brand new level of insolence. Especially that when it comes to gaming, GTX 1080Ti is not far behind. Some times the differences are marginal. Lack of SLI capabilities, almost identical vapor chamber cooler than few previous generations (don`t be fooled by the gold colour), only reference model available are only the few drawbacks. But i guess that is the price for having something fresh nad brand new – at least in Nvidias mind.
But let`s dig a little deeper, there should be something new inside since it costs over twice the price of the previous Titan. Please take a quick look at this cards comparison :
| GPU | Nvidia Titan Xp | Nvidia Titan V |
|---|---|---|
| Microarchitecture | Pascal | Volta |
| Fabrication process | 16nm | 12nm |
| GPU type and speed (boost) | GP102-450 – 1405 MHz (1582 MHz) | GV100-400 – 1200 MHz (1455 MHz) |
| RAM type and speed | 12GB GDDR5X – 11448 MHz | 12GB HBM2 – 1700 MHz |
| RAM bandwidth | 547.7 GB/s | 652.8 GB/s |
| Bus width | 384 bit | 3072 bit |
| CUDA cores | 3840 | 5120 |
| TENSOR cores | 0 | 640 |
| Texture units | 240 | 320 |
| ROPs | 96 | 96 |
| Streaming Multiprocessors | 30 | 80 |
| Transistors (bilion) | 12 | 21,1 |
| TFLOPS | 12 | 110 |
As you can see – at a first glance – the new card is slower. But is it, really? One value is strangely different from the others and it`s the memory bus width. Volta has such a wide bus because of the HBM2 memory interface. As you cen see it is much slower than the GDDR5X, but this is greatly compensated by the bus width what results in overall higher performance (over 100 GB/s more bandwidth). New ram type is more than that. Chips one the face of the PCB are gone now, HMB2 is “integrated” in the core. It`s a separate chip but is produced with the gpu on the die. But memory is not the whole story. The main reason for the Voltas ludicrous price is the use of a deep learning processors called Tensor cores. Nvidia claims that Titan V is even nine times faster than previous model in applications like artificial intelligence or simulation processing. Tensor cores are specially designed to fulfill those type of tasks. And this the main reason gaming benchmarks are not sky high, in fact they only benefit form the extended core setup. GV100-400 has more cuda cores, texture units and streaming processors after all. That raw processing power of 110 TeraFLOPS gives me shivers on the whole body. Mind you, owning this monster does not automatically ensures that your beefed up rig will finally run Crisis 😀 Application needs to be designed and programmed to make use of Tensor cores, and none of modern games are coded this way.
With all that being said, I have to admit I simply adore it. Needles to say – I cant`t stop thinking about owning one 😀 I mean c`mon, just look at it !
Copyright of all pictures – NVIDIA.
Leave a Reply