GTC 2014: Nvidia reveals dual-GPU Titan Z, new Pascal GPU offers colossal memory bandwidth

Nvidia’s Jen-Hsun Huang kicked off the GPU Technology Conference (GTC) this week — and first on the list of announcements was a massive new GPU project. According to Nvidia, two of the greatest problems facing modern graphics cards is the memory bandwidth problem between both the CPU and GPU, and the GPU to onboard memory. These issues have made it difficult to incorporate the GPU as a truly equal computational partner — the GPU may have a firehose connection to its own memory banks, but it connects to the rest of the system through a PCI-Express-sized straw. Nvidia wants to solve both problems, and Jen-Hsun believes its new chips (codenamed Pascal) will do so.

NV-Link and Stacked RAM

First, there’s the CPU-GPU link issue. Nvidia has developed what it calls NVLink, a new technology that leverages the PCI-Express standard to increase bandwidth between the CPU-GPU by 12-15x.



While this is an enormous jump over current PCI-Express 3.0, it’s less of a leap over next-generation PCIe 4.0. Given shipping dates, Nvidia’s improvements for NVLink should still offer two-to-six more times performance than the then-current PCIe 4.0. Since Nvidia isn’t in the chipset manufacturing business anymore, we assume that Nvidia will offer NVLink certification on certain chipsets.




 Update: A bit of digging appears to show a link between IBM’s CAPI (Coherent Accelerator Processor Initiative) and NVLink. CAPI is encapsulated by PCI-Express and appears to use the same bus architecture but with higher bandwidth capabilities and lower overall latencies. Nvidia’s blog post refers to IBM specifically incorporating support for NVLink into its next-generation Power chips, which implies that this is an enterprise-level technology. “Although future NVIDIA GPUs will continue to support PCIe, NVLink technology will be used for connecting GPUs to NVLink-enabled CPUs as well as providing high-bandwidth connections directly between multiple GPUs. Also, despite its very high bandwidth, NVLink is substantially more energy efficient per bit transferred than PCIe.”

It’s not clear if this can be integrated into next-generation motherboards or if it’ll remain a Power-Nvidia corner case. The other big announcement of the morning was a new GPU with integrated TSVs and stacked DRAM, codenamed Pascal. But wait — haven’t we heard this before?

Five to 12 times PCIe 3.0 bandwidth, two to four times the current memory bandwidth.

Turns out we have. Last year, Nvidia was talking up the Maxwell successor, codenamed Volta.


This year, Nvidia is talking up the next-generation Maxwell successor, codenamed Pascal.



NVLink is a new technology that wasn’t on the previous roadmap, but there’s no mention of Volta. The gap between Maxwell and Pascal is larger than what we saw last year — assuming that’s not a marketing decision, it implies that Nvidia ran into trouble with its roadmap for next-generation GPUs and had to cancel the version they’d planned in favor of a new ground-up redesign that will bring new features to market in 2016.

Titan Z: Dual-GPU Titan

Nvidia’s other major new GPU announcement was an unveil of a dual-GPU Titan card, dubbed Titan Z. This is a dual GK110 card with two full implementations of the GPU (2,880 cores each, 5,760 cores total), 12GB of RAM, and up to 8TFLOPS of theoretical FLOPS performance. Nvidia hasn’t revealed clock speeds yet, but the performance should be 1.5 to two times better than the existing GTX 780 Ti.



The one thing Nvidia did reveal is the price: a cool $3,000. That’s substantially more expensive than a pair of GTX 780 Ti’s running in SLI; we’ve got a feeling that Nvidia’s decision to raise prices here reflects a desire to charge more for Titan’s double-precision floating point and additional RAM.