Mixed-precision, 32GB of 3D memory and NVLink
In addition to unveiling the new GTX Titan X graphics card based on GM200 Maxwell GPU, Nvidia also revealed a bit more details regarding its next-gen Pascal GPU architecture during GPU Technology Conference 2015.
While we already heard a bit about Pascal GPU during the GTC 2014 show last year, Nvidia has now revealed a bit more details regarding it, including the fact that it will operate with a mixed-precision mode.
While current Nvidia GPU Maxwell architecture works mostly with FP32 compute, while Tegra X1 also got FP16 mode, it appears that Pascal GPU will be able to combine int8, FP16 and FP32 compute operation performance, or simply combine single- and double-precision data points, retaining accuracy but at the same time improve performance.
According to rough estimates, or as Jen-Hsun Huang called them, "CEO math", Pascal should provide up to four times the throughput in mixed-precision workloads and be up to 10 times faster than Maxwell. It will also have significant memory bandwidth compared to Maxwell, or about 3x the bandwidth according to Jen-Hsun.
The new Pascal GPU will also feature 3D memory, something that Nvidia has already revealed at last year GTC. According to Nvidia, next-gen Pascal GPU will be able to use up to 32GB of RAM and have that 3x the memory bandwidth when compared to Maxwell. While it did not reveal any precise details regarding the memory, Nvidia will most likely use 2.5D HBM memory, the same one that we will see on upcoming AMD Fiji GPU, which also means that memory will connected with an interposer layer. While this is significantly better than what we have now, earlier rumors suggest that HBM gen2 memory will lack that same interposer and offer much better performance and bandwidth so hopefully Nvidia eventually use it.
The new Pascal GPU will also features Nvidia's NVLink, which will allow to link a lot more GPUs than it is possible today. What is also important is a promise that these GPUs will be able to communicate much faster than with SLI and bring a possibility to connect up to 64 GPUs together.
According to current Nvidia GPU roadmap, Pascal GPU will be coming in 2016 while Volta GPU is again back on the roadmap and will be coming in 2018.
Courtesy of Anandtech.com.
Courtesy of Anandtech.com.
Source: Nvidia.com.
While we already heard a bit about Pascal GPU during the GTC 2014 show last year, Nvidia has now revealed a bit more details regarding it, including the fact that it will operate with a mixed-precision mode.
While current Nvidia GPU Maxwell architecture works mostly with FP32 compute, while Tegra X1 also got FP16 mode, it appears that Pascal GPU will be able to combine int8, FP16 and FP32 compute operation performance, or simply combine single- and double-precision data points, retaining accuracy but at the same time improve performance.
According to rough estimates, or as Jen-Hsun Huang called them, "CEO math", Pascal should provide up to four times the throughput in mixed-precision workloads and be up to 10 times faster than Maxwell. It will also have significant memory bandwidth compared to Maxwell, or about 3x the bandwidth according to Jen-Hsun.
The new Pascal GPU will also feature 3D memory, something that Nvidia has already revealed at last year GTC. According to Nvidia, next-gen Pascal GPU will be able to use up to 32GB of RAM and have that 3x the memory bandwidth when compared to Maxwell. While it did not reveal any precise details regarding the memory, Nvidia will most likely use 2.5D HBM memory, the same one that we will see on upcoming AMD Fiji GPU, which also means that memory will connected with an interposer layer. While this is significantly better than what we have now, earlier rumors suggest that HBM gen2 memory will lack that same interposer and offer much better performance and bandwidth so hopefully Nvidia eventually use it.
The new Pascal GPU will also features Nvidia's NVLink, which will allow to link a lot more GPUs than it is possible today. What is also important is a promise that these GPUs will be able to communicate much faster than with SLI and bring a possibility to connect up to 64 GPUs together.
According to current Nvidia GPU roadmap, Pascal GPU will be coming in 2016 while Volta GPU is again back on the roadmap and will be coming in 2018.
Courtesy of Anandtech.com.
Courtesy of Anandtech.com.
Source: Nvidia.com.
News by Luca Rocchi and Marc Büchel - German Translation by Paul Görnhardt - Italian Translation by Francesco Daghini
Previous article - Next article
comments powered by Disqus
comments powered by Disqus