NVIDIA Planning To Ditch Maxwell GPUs For HPC Purposes Due To Lack of DP Hardware – Will Update Tesla Line With Pascal in 2016, Volta Arriving in 2017
Some major reports on NVIDIA's GPU development process has arrived from 4Gamer (via Kitguru) which shows what NVIDIA plans to do with their current generation Maxwell plus the next generation Pascal and Volta GPU cores. The details come from Kenichi Hayashi (NVIDIA platform business headquarters Director) who revealed the development roadmap for NVIDIA during the upcoming years at a conference held in Tokyo, Japan.
NVIDIA Maxwell GPUs Ditched For High Performance Computing In Tesla Parts
Do note first that this news comes straight from NVIDIA's official representative so its highly likely that this news is true and NVIDIA would possibly provide an update on this matter during their CES conference when their event official kicks off early next month. Now according to 4Gamer's interview with Mr.Kenichi, it's said that NVIDIA won't opt to use Maxwell for their Tesla GPU accelerators which are aimed towards the HPC market. While the addition of Maxwell GPUs in the Quadro line of cards would be seen over the upcoming months, those aiming for super computing won't get an updated Maxwell core and might have to stick with current Kepler based Tesla offerings.
Just over a month back, we saw NVIDIA update their Tesla lineup with the Kepler GK210 based Tesla K80 Accelerator which featured 24 GB GDDR5 memory and around 8.5 TFlops of single precision performance with its 4992 CUDA Cores. The card is a dual core Tesla offering and remains to be seen if NVIDIA will now launch a K80X to replace it in the future during the 2015 timeline in response to AMD's own FirePro dual chip offering which is yet to be launched.
The reason behind not deploying Maxwell in Tesla Accelerators is said to be the lack of FP64 Floating Point Units and additional double precision hardware. And the reason behind not including DP FPUs in Maxwell might have to do with the extremely efficient design that NVIDIA was aiming for. This however means that NVIDIA's upcoming Maxwell core, the GM200 which is the flagship core of the series might just remain the GeForce only offering unlike Kepler which was aimed for the HPC market first with the Titan supercomputer and launched a year later after the arrival of the initial Kepler cores as the GeForce GTX Titan. While this may sound negative for the HPC market since they'll have to wait for another year and a half to get the next GPU based Tesla part, it's good news for the consumer market. it should also be noted that the advantage of owning a consumer level card with full DP support such as the Titan and Titan Black will be lost.
Since the DP FP64 FPU hardware blocks will be removed from the top-tier cards that are rumored to arrive next year, they will include several more FP32 FPUs to make use of the remaining core space and that means better performance for the GeForce parts since games have little to do with Double precision performance. So with no Tesla parts based on Maxwell, HPC and the supercomputing market will have to do with either the current GPU parts or wait till next year when Pascal arrives. There's also a slight bit of an update on Pascal and Volta too.
NVIDIA Ramps Up Pascal and Volta Production - Next Generation Volta Arrives in 2017 To HPC FIrst
We have already heard about Volta coming in 2017 during the announcement of the two new supercomputers, the Summit from Oak Ridge National Laboratory and Sierra from Lawrence Livermore National Laboratory. This news just makes it much more official since its coming from Mr.Kenichi. While Pascal would replace Maxwell as the next generation node based on the 16FF+ process node, Volta will be launched in just a years lifespan of Pascal to update the next supercomputers with an enormous update in compute performance. While it is easy to believe that Pascal will stick with the 2 years lifespan of every new GPU at the consumer level, the HPC side will see an update one year earlier just after Pascal launches. In 2017, NVIDIA will ship both Laboratories with their new Tesla accelerators based on the Volta GPU architecture that would be based presumably on the 10nm FinFET node.
Rated at a peak performance of 150-300 PFLOPS, the Summit supercomputer will be based on more than 3400 compute nodes with each node consistin of several next generation IBM POWER9 CPUs and NVIDIA Volta architecture based Tesla GPUs. Each node will deliver around 40 TFLOPs of compute performance and is said to be enough to outperform an entire rack of top of the line Haswell based x86 CPU servers. The most impressive feature about Summit will be that while it will consume 10% more power than the Titan supercomputer, it will also deliver an impeccable 5 times more system to application power than the previous fastest supercomputer. While Titan was rated at 25-30 PETAFLOPs, Sierra will be >100 PFlops while Summit will be 150-300 PFlops at any given circumstance.
With Summit come a few other enabling technologies such as NVLINK which is expected to debut around 2016. The Pascal GPU would be the first to introduce NVLINK which is the next generation Unified Virtual Memory link with Gen 2.0 Cache coherency features and 5 – 12 times the bandwidth of a regular PCIe connection. This will solve many of the bandwidth issues that high performance GPUs currently face.
First technology we’ll announce today is an important invention called NVLink. It’s a chip-to-chip communication channel. The programming model is PCI Express but enables unified memory and moves 5-12 times faster than PCIe. “This is a big leap in solving this bottleneck,” Jen-Hsun says. NVIDIA
While Pascal will aim to add 3D Stacked Memory designs with on-board and on-chip level integration, Volta will further refine this technology and support the new NVLINK tech which is the next generation Unified Virtual Memory link with Gen 2.0 Cache coherency features and 5 – 12 times the bandwidth of a regular PCIe connection. This will solve many of the bandwidth issues that high performance GPUs currently face. These supercomputers will alone amount to a major revenue generation for NVIDIA however the GeForce department still amounts to a majority of NVIDIA sales, what's going to happen to GeForce parts during the 2015-2018 timeline.
Well, this would be mostly speculation but I believe that 2015 will see the arrival of several new GeForce and Quadro parts based on the Maxwell chip architecture. Currently we know only about the GeForce GTX 960 which is allegedly launching on 22nd January and aiming for the sweet price point of $200 US. Aside from that is all just dust but we can make a pattern from here. The GM200 will arrive later in 2015 prior to a few GM204 Quadro parts. In 2016, we will see NVIDIA update the Telsa market as quickly as possible while launching a few GP*04 Pascal parts. The flagship parts would arrive in 2017 most definitely for consumers however they will stick the long run while NVIDIA makes a quick shift towards Volta accelerators the same year. It is unknown when Volta makes way to consumers, the date could probably be set in 2018 however we can't yet confirm since there's no official information on that. Here's hoping that NVIDIA provides an update on their GPU road map at CES 2015 in the upcoming week.