SAN JOSE, Calif. – Nvidia will give 3-D chip stacks a nudge forward with its next-generation graphics processor, just not quite as big a one as it originally suggested. That was one of several takeaways from the company's annual GPU Technology Conference here.
Nvidia also updated its high-end consumer graphics card, rolled out a general-purpose GPU appliance for business and teamed up with VMWare to enable a GPU cloud service. In addition, it tipped NVLink, a new chip-to-chip interface that it and IBM will use.
A walk across the show floor showed virtual reality gear, gesture recognition software, car computers, and plenty of server GPU servers running on Nvidia chips. (See pictures from the keynote and show floor on subsequent pages.)
The big headline from the event was Pascal, Nvidia's next-generation core graphics processor. The chip will debut in 2016 using a 2.5-D chip stack with the GPU and a DRAM stack side-by-side on a silicon interposer based on TSMC's CoWoS technology.
Huang confirmed in a brief hallway chat that Pascal will use a 2.5-D stack based on TSMC's CoWoS technology.
The move marks a significant step forward for 3-D stacks, but not as big a one as Nvidia CEO Jen-Hsun Huang described just a year ago here. At that time he announced Volta, a 2015 processor suggesting it could use Micron's Hybrid Memory Cube as part of a 3-D stack.
Pascal may still tap Micron's HMC but it now has a new name and is at least a year further away. What's more it is not yet the full kind of 3-D vertical stack the industry has been working towards for years but still faces technical and business hurdles.
Huang suggested Pascal's use of the 2.5-D stacking and the new NVLink will help break through the memory bottleneck of current systems using PC Express.
"The last time I looked the only one seriously shipping 3-D ICs was Xilinx, so if Nvidia gets behind the technology, it will help more people believe it is viable," said Nathan Brookwood, principal of market watcher Insight64 (Saratoga, Calif.).
The 2.5-D version of Pascal is only expected to be used in high-performance computing clusters, a market generally of less than 100,000 units/year, said Jim McGregor, principal of Tirias Research.
Next page: IBM, Nvidia create NVLink