Nvidia Unveils PCI-Express Variant Of Tesla P100 At ISC 2016
Back in April, at GTC (Nvidia’s own GPU Technology Conference), Nvidia announced its Tesla P100 graphics card. At the time, it came with only an NVLink interconnect. Now, for the first day of ISC 2016, Nvidia announced the PCI-Express variant of the Tesla P100.
This card is based on the same Pascal GPU as the NVLink P100, packing 15.3 billion transistors made with the 16 nm FinFET manufacturing process and of course with HBM2 memory. The GP100 aboard is a huge GPU that measures 610 square mm.
However, the PCI-Express variant of the P100 operates a little slower than the previously-announced NVLink variant. Due to TDP restrictions in PCI-Express environments, Nvidia had to lower the power levels down to 250 W from 300 W, which it accomplished by lowering the GPU clocks from a boost clock of 1480 MHz down to 1300 MHz.
The result is that, according to Nvidia, performance measures 18.7 teraflops for half-precision (what much of Deep Learning is done at), 9.3 teraflops for single-precision, and 4.7 teraflops for double-precision -- a tad below the NVLink variant, but still far beyond what any other card can offer over PCI-Express.
In context, the NVLink P100 is therefore meant for large-scale applications, whereas the PCI-Express variant is meant to be used in workstations for smaller-scale workloads--which is still ideal for Deep Learning.
Whereas the NVlink P100 came with 16 GB of HBM2 memory only, the PCI-Express variant comes with either that, or for less memory-intensive applications, a 12 GB variant that delivers 540 GB/s memory bandwidth rather than 720 GB/s.
Availability for the cards in systems is slated for Q4 2016.
Follow Niels Broekhuijsen @NBroekhuijsen. Follow us @tomshardware, on Facebook and on Google+.
You'd Also Like
It is more about the environment. Sure they could throw a 300W GPU in there. Problem is that it would limit the TDP of other parts. Servers and HPC, where this will be going, require a precise design for the best cooling and functionality. They don't have custom liquid cooling or cases but a set design and air cooling.
It is also not stating a TDP limitation of PCIe but rather "Due to TDP restrictions in PCI-Express environments" which again refers to the limitations in cooling when you have tons of these in a single room doing HPC work.
The interesting thing is that this might be close to what the Titan for Pascal will be. Guess we will have to wait and see though because the Titan already has 12GB of VRAM and I would expect the new GPU to have more, 16GB but that is just me.
Plus, why do you assume these will only be used in servers? The aforementioned Xeon Phi's come in two variants - actively and passively air-cooled. The passive ones are for use in servers, while those with integrated blowers are aimed at workstations.
I've read that the GP100 has no graphics-specific blocks, meaning it cannot be used on a graphics card. We'll have to await a completely different chip.
Overwall what bothers me is that nvidia fights a battle with itself on both fields mainstream high end and profesional... And this is no good...
It may be silly but I assume that nVidia does it for a reason. Do you think they don't want to be able to sell people a 300W GPU/Coprocessor? I think they would if they could. People assume it is because they want people to go to the next step, that is probably true and would be true for any company but I also think they wouldn't leave a possible market open for Intel to scoop up. I can't find anything confirming any restrictions for PCIe to 250w TDP but again there has to be a reason why nVidia wouldn't try to sell to a market they could easily sell to.
And I am not saying only servers but the majority of Tesla and Xeon Phi cards are put into HPC farms and not work stations.
I also said this is probably close to what the next Titan will be, not that it is. Nothing is stopping nVidia from having a version of GP100 with the GPU blocks.
This will be useful for AI and Machine learning )))) it will improve the machine learning infrastructure !!!
Because there's not a big enough market of people willing to pay like $5k or $10k for a gaming card. So, they went for a less price-sensitive market, until they can bring down all the costs that make this thing so expensive.
A more cynical explanation would be that GTX 1080 already beats all of AMD's existing + soon-to-be-released cards, allowing them to hold back their big weapon.
I'm wondering if they're worried about power consumption to make the PCIe cards to functional in more server chassis. Thing about Server's.. They tend to be be purpose built. If you buy a server that's been designed specifically for GPGPUs, it'll have an extra beefy Power Supply(s) and extra power leads in the case for the cards.
But, there are actually cases where people want to add a GPU co-processor to a chassis that wasn't specifically designed for a GPU, or to one that wasn't designed for a 300W card (or cards).
Believe it or not, sometimes [ahem.. Higher Ed] cost is a huge driver of hardware selection. I've got A LOT of generic cases with Supermicro server boards in them [occasionally Desktop boards] in my Data Center because compute needs are effectively infinite in Higher Ed Research but funding is anything but.
Other thing to consider.. To the best of my knowledge, NVLink has still only been announced for Power systems. I could be wrong, but I'm not aware of NVLink being available for any Intel systems.. Because why would Intel do the R&D to adopt a proprietary bus for a niche market, especially when Intel is trying to compete in that market with Xeon Phi?
So.. With NVLink systems.. They're ALL custom designed to run whatever Nvidia spec'ed for these cards. With PCIe, you have to deal with what's out there.
Oh yea.. 16-lane PCIe slots are supposed to support up to 75W. The spec also only Officially supports the use of One 6-pin and One 8-pin aux power feed and guess what that totals up to? 300W. [some graphic cards have dual-8Pin but that's apparently not supported in the PCIe spec]
..Another thing about servers.. Especially HPC Servers.. You don't push the spec. You don't over-clock CPUs and you don't run your co-processor cards at 100% of the available power on the Bus. We don't like our servers crashing when they're 10 seconds into a simulation and we REALLY don't like them crashing a week into a simulation.
I can understand not wanting to max your PSU right to the hairy edge, but server gear should be built with enough margins to run it at spec.
I thought any serious HPC effort made liberal use of snapshotting to deal with the inevitable failures. I mean, as soon as you try to do anything at scale, the failures are virtually guaranteed. Hoping stuff doesn't crash is pretty much a plan for failure.
Not sure if this post is serious or not, but if it is:
This is not a gaming card. It doesn't even have video outputs.