A year ago at a conference NVIDIA told about a hypothetical architecture Volt (Volta), which should be expected in 2016-2017 and which must first try to imagine 3D-memory HBM (High Bandwidth Memory). At today's conference, NVIDIA chief tactfully silent about solutions generation Volt, while the company's plans for 2016 appeared Pascal graphics cards with 3D-memory interface and new interprocessor NVLink. It can be assumed, that Volt will be postponed on a scale of priorities at a later date and will boast some innovative changes in architecture, while Pascal will be a combination of Maxwell, 3D-memory NVLink.
In yesterday's news, my colleague touched on the main points related to stack memory on the processor substrate and Pascal interface NVLink.
As for the 3D-memory, NVIDIA has gone down the same path as AMD company . Actually, no other way. Memory HBM preparing for standardization committee JEDEC. Samples of high-density stack package chips using TSVs-compounds presented today all the leading manufacturers of memory, but still the same company Micron is busy promoting cubic HMC-memory, the South Korean company SK Hynix sent all their forces to promote a standard HBM-memory.
Due to the broadband interface HBM (h1024) data transfer rate between memory and the GPU Pascal on the substrate crystal can exceed three times the bandwidth, which we now see between GPU Kepler and GDDR5. Compared with the GPU Maxwell memory exchange speed by switching to 3D-memory can be doubled. As you can see, Pascal optionally have a new architecture to capture the imagination. It is important to note that the consumption of GPU will remain unchanged, which suggests a significant increase in energy efficiency.
Regarding interface NVLink. Overall NVLink coined for ligaments GPU NVIDIA composed of supercomputers for GPGPU-acceleration calculations. Prototype module Pascal, the company revealed that uses something like a standard server connector type mezzanine. However, in the future, the company expects to extend the use of NVLink also among discrete gaming graphics.
Recall last summer NVIDIA and IBM entered into a strategic agreement on the establishment of a common computing ecosystem. It was also about licensing elements IBM Power architecture, NVIDIA and Google.
So, NVLink. According to the company, this kind of productive alternative to PCI Express 3.0, which can speed 5-12 times surpass the regular interface. So far, however, the advantage NVLink looks modest. Eight lanes PCI Express 3.0 can collectively provide the exchange rate at 8 GB / s (8 Gb / s per lane). Related Products :
|