Pascal is due out in 2016 and features stacked DRAM and a new connectivity protocol called NVLink.
Nvidia’s GPU Technology Conference kicked off Monday in San Jose with a bevy of announcements. While the 5K-ready GeForce GTX Titan Z stole the early thunder during the keynote, one of the other big announcements was Nvidia’s new GPU called Pascal.
Named after the mathematician Blaise Pascal, this GPU incorporates technology that seeks to address the GPU-CPU memory bandwidth bottleneck from PCI-Express that’s holding back GPGPU computing.
In order to bypass this memory bottleneck, Pascal will feature a new technology called NVLink as well as stacked DRAM. Jen-Hsun Huang said on stage at GTC that NVLink will offer five-to-12 times the speed of PCIe 3.0, for a total bandwidth throughput of 80GB/sec.
Other details on NVLink were rather sparse, but it’s assumed that Nvidia would license the technology to motherboard manufacturers. Obviously this would require motherboard makers to play along and offer the connector alongside regular PCIe 3.0 ports.
Huang also said that Pascal will feature stacked DRAM intergrated directly on the GPU die. Nvidia has talked up the virtues of stacking DRAM into high-density PCBs before via Through Silicon Vias (TSVs), but this is the first implementation — and not just announced — of the concept.
Nvidia is targeting a 2016 release for Pascal, and more information should be available later this year.
What happened to Volta and Parker?
The more interesting thing about Huang’s GTC keynote was not what was announced, but rather what was not.
Until GTC began, as per Nvidia’s roadmaps the successor to Maxwell was set to be Volta — due out sometime in 2015-2016. This was going to be the chip with stacked DRAM. But it appears it was never meant to be. The reasons for this are too numerous to speculate with any sense of certainty, but it looks like Nvidia ran into some technical trouble and had to do some serious redesigning.
Volta wasn’t the only project that did a disappearing act at GTC. Parker, the much-hyped future GPU at CES, was erased from history at GTC. On the Tegra side, Parker, which was supposed to feature a Project Dever CPU, a Maxwell GPU and the FinFET process also has been erased from history.
Nvidia hasn’t disclosed any challenges that might allude to these projects disappearance in SEC filings, so their fate is a real mystery. All we are left with is “Erista,” which Nvidia only annotated as having the Maxwell GPU on slides.
For Nvidia, this is a big problem as it continues the pattern of the company’s keynotes not reflecting reality. Every year Huang brags about the next big thing on stages around the world, but often these projects suffer from lack of manufacture wins and real-world performance that’s a far cry from advertised claims. Investors should be seriously concerned about the ability of the company’s management to deliver.