GTC: GTX Titan Z, Shield and Future Plans
At the GPU Technology Conference, (GTC) Nvidia had many things to show. There were new product announcements, earning reports and plans for the future. Here are some interesting things from their keynote.
Nvidia has announced a new top end card – GeForce GTX Titan Z. It will have two full GK110 chips with a total of 12 GB VRAM (6 GB per GPU). Other specifications were not named, but it is reasonable to expect a full FP64 performance, just like in other GTX Titan cards. While clocks were not mentioned and are likely to be lowered, the TDP of the card will definitely be well beyond 300W, possibly similar to the monstrous 365W of GTX 590. The card is very large and its cooler is 3 slots high. Finally, do not expect a good value offer, as Titan Z will set you back by $3000. With such two for the price of three deal, it is very hard to imagine scenarios where the card is worth its price. Pushing even more performance into tiny cases, like Tiki for outrageous price may be one of few justified uses for it. GeForce GTX Titan Z will launch next month.
There also are some news about more affordable hardware – Nvidia Shield. Nvidia drops Shield price to $200 for a limited time once again. The portable console also receives an OS update to Android 4.4.2 “KitKat” with additional tweaks to some functions. Streaming functionality receives a very important update, allowing the use beyond local network. Remote GameStream allows you to stream games from your home PC via internet to the Shield. Depending on your connection, the quality and latency may be close to Onlive or Gaikai levels, but it avoids the ownership problem of cloud gaming platforms. In addition, unless you often travel far away, streaming games within the same city should offer relatively low latency. If you have several PCs running, you can select which one to use for streaming. Finally, Shield’s console mode, where you connect shield to a TV via HDMI, get a usability boost by offering Bluetooth keyboard and mouse support.
Nvidia’s roadmap for the future has changed since last year. Nvidia originally planned to introduce unified GPU-CPU virtual memory in Maxwell. We have already seen first Maxwell GPUs on 28 nm with impressive power consumption, but without unified memory. That feature is now expected to be present on Maxwell’s successor – Pascal GPUs in 2016. Maxwell’s main feature in the current roadmap is DirectX 12 support. There was some more information about DX12 hardware since the initial report. There will be some new DX12 features not supported on existing hardware after all, but Microsoft did not mention what exactly those will be. The DX feature set of GTX 750 and 750 Ti had been similar to Kepler so far, so it will be interesting to know how exactly DX12 is supported in Maxwell.
After Maxwell, there are some changes. Last year the Volta GPUs with stacked RAM were mentioned. Now, there is another architecture before it, codenamed “Pascal”, which will use the stacked memory. Thus, in 2016 Pascal should introduce unified virtual memory, 3D memory on GPU and a new feature called NVLink. The VRAM configuration used in Pascal will be based on JEDEC’s High Bandwidth Memory standard. The prototype card shown on keynote uses all the memory this way, without splitting it into slow and fast. The bandwidth targets of the new memory design were not mentioned due to the unfinished cost/performance balance.
Another important addition to Pascal will be NVLink – interconnect bus, which should provide much higher bandwidth than PCI-E 3.0. Each block of NVLink should provide 20 GigaTransfers/s, compared to 8 GT/s in PCI-E 3.0. More blocks can be used to provide additional bandwidth between devices. The connections will be directly between GPUs without using some common hub. To achieve those speeds, Nvidia plans to use mezzanine connectors. These are similar to sockets CPU use with many contacts spread over a wide area. This would also mean putting the graphics card parallel to the motherboard into its own special socket. The NVLink still requires a PCI-E bus, so the overall design will not be simple. There are also plans to add NVLink connection to the CPU. While these plans sound scary, do not expect to be forced to use them any time soon. NVLink is similar to Intel’s and AMD’s server connecting solutions and will be more relevant for high performance compute servers. For our gaming rigs at home PCI-E should remain more than enough for the foreseeable future.