facebook rss twitter

Nvidia reveals more about Pascal at Japanese GTC

by Mark Tyson on 18 November 2015, 13:25

Tags: NVIDIA (NASDAQ:NVDA)

Quick Link: HEXUS.net/qacwet

Add to My Vault: x

A recent report on VR World provides some new information about Nvidia's Pascal GPU design and the plans the chipmaker has for the technology. At the Japanese GTC (GPU Technology Conference) last month Marc Hamilton, ‎VP of Solutions Architecture and Engineering at Nvidia, closed his GTC keynote with a talk about the future GPUs coming from Nvidia.

Hamilton said that Pascal will be shipping in volume in 2016. He touched upon the most important and unique new features of Pascal, saying it will provide; mixed precision (INT8, FP16 and FP32) and double precision calculations, on board 3D memory and the new NVLink interconnect between GPUs and memory.

3D memory will improve both memory capacity and bandwidth significantly. With a faster GPU and 3D memory the existing interconnect technology was seen to be 'outdated', so the 5x faster 3D Link interconnect was developed. Nvidia will also support unified memory in its software, providing a single address space for CPUs and GPUs.

Filling in further background detail, VR World confirms that the Pascal GPU will be manufactured by TSMC on the 16nm FinFET process. Pascal will be able to support up to 32GB of HBM2 memory and provide up to 1TB/s of bandwidth, while internally enjoying 2TB/s bandwidth.

The next generation Nvidia GPU will also be made in multi-GPU packaging in all-new Tesla server accelerators. NVLink will enable dual-GPU cards with up to 80GB/s bi-directional interconnects, this should replace PLX PCIe Gen3 bridge chips that can only support 16GB/s (8GB/s per GPU). Newer APIs like DirectX 12 and Vulkan should benefit greatly from the improved memory management and access speeds as we move towards the mainstream 4K PC gaming era.

In other Nvidia Pascal GPU news, The Register reports that the US NOAA agency, involved in weather prediction and related fields, intends to build a new computer boosted by 760 Nvidia Pascal GPUs. The new machine will improve the NOAA Global Forecast System (GFS) grid reference point resolution from 28km to 3km. Increased resolution, more clarity about how weather systems are acting, should simply result in better forecasts.



HEXUS Forums :: 12 Comments

Login with Forum Account

Don't have an account? Register today!
NV..LINK…3D MEMORY…yawning…GDDR5 is still very fast the only problem is the under performing cores. Lets see if the next fastest CUDA core will beat the Fiji core with its 4096 Shading units that currently beats any core made by man in floating point math performance
lumireleon
NV..LINK…3D MEMORY…yawning…GDDR5 is still very fast the only problem is the under performing cores.
That's not what AMD found.

Lets see if the next fastest CUDA core will beat the Fiji core with its 4096 Shading units that currently beats any core made by man in floating point math performance
You might want to revise that statement (2.91TFlops DP for Tesla K80, 2.62TFlops DP for Firepro S9170. The best fiji can manage is 0.5)
What a piece of fluff :(

Wish there was some real information there.
lumireleon
NV..LINK…3D MEMORY…yawning…GDDR5 is still very fast the only problem is the under performing cores. Lets see if the next fastest CUDA core will beat the Fiji core with its 4096 Shading units that currently beats any core made by man in floating point math performance

Not sure if you're being serious or if you truly think the only job a GPU does is floating point maths?
Also if GDDR5 is still fast enough then how do explain the advantage that AMD gain at higher resolutions when using HBM (technically the same thing as 3D Memory) on their Fury cards?

On a more serious note, does anyone know if NVLink is going to be a consumer grade interconnect or is it only going to be limited to HPC and SLI?
Corky34
Not sure if you're being serious or if you truly think the only job a GPU does is floating point maths?
Also if GDDR5 is still fast enough then how do explain the advantage that AMD gain at higher resolutions when using HBM (technically the same thing as 3D Memory) on their Fury cards?

On a more serious note, does anyone know if NVLink is going to be a consumer grade interconnect or is it only going to be limited to HPC and SLI?
Are you thinking in terms of an external dock/thunderbolt competitor? At the moment I get the impression it's not all that different to AMD/MS's existing DMA/move engine type stuff.