NVIDIA's GTX 1080 & GTX 1070 Detailed
re fresh off the unveiling of not one but two of NVIDIA’s next generation GPUs, the GTX 1080 and GTX 1070 and it looks like everyone has a lot to be excited about. As Jen-Hsun Huang was on stage we all got to see how their Pascal architecture, which was initially announced about a year ago and detailed a bit more at this year’s GTC, has evolved from a predominantly compute-focused design into something that can effectively accelerate next generation games. While there weren’t a lot of details for anyone watching the live stream (more information will be available to the public soon), there are some things I wanted to go over since what was announced seemed to be pretty earth shattering.
Let’s start with the obvious first: while NVIDIA has been talking about the Pascal-based P100 for some time now, the Geforce lineup won’t be receiving the big-daddy core quite yet. Its 15 billion transistor core that packs in an astounding 3584 CUDA cores is extremely complex to produce. Pair that up with a new 16nm FinFET manufacturing process alongside cutting edge HBM2 memory and the result is (for now) limited production quantities that will be shifted towards leading Tesla series customers.
The GeForce segment on the other hand will receive a cut down 7.2 billion transistor Pascal core which rolls in all the benefits associated with this new architecture but does so without –hopefully- sacrificing core yields or throwing pricing to the wolves. It will still use an efficient 16nm process note but instead of including expensive HBM modules, NVIDIA opted for Micron’s new GDDR5X memory for the GTX 1080. The GTX 1070 meanwhile retains traditional GDDR5 modules. This is all being done in an attempt to maximize product availability, get these cards into buyers’ hands and avoid costly new-generation technologies that may have availability issues of their own.
NVIDIA’s choice of GDDR5X is an interesting one since we’ve all been riding the HBM bandwagon for quite some time now. Unfortunately, while HBM does have tangible benefits its volume production has been beset by poor yields, it is still quite costly and actually incorporating it into a die package involves additional complications. It looks like NVIDIA simply didn’t think the cost and other speedbumps were worth the chance. Nonetheless, GDDR5X is supposed to offer improved throughput in comparison to GDDR5 with frequencies hitting 10Gbps and higher.
In a discussion with Ryan Smith from Anandtech, he mentioned that unlike GDDR5 and the previous GDDR3 technologies, GDDR5X is solely produced by a single manufacturer, that being Micron. That's actually an interesting point since sole sourcing has in the past proven to be challenging if there are any hiccups at the one supplier. Hopefully this won’t cause supply problems as NVIDIA attempts to ramp up production to meet what will likely be huge demand from the buying public.
Moving onto the specs and we actually have an interesting mashup. The basic GTX 1080 will come with 2560 CUDA cores, clock speeds of 1607 / 1733MHz for the Base and Boost respectively and 8GB of GDDR5X operating at 10Gbps across a 256-bit interface. All of this happens within a TDP that’s about 15W higher than the outgoing GTX 980. Even though these Pascal cores are supposed to be highly overclockable there will also be a so-called $699 Founders Edition that features what’s assumed to be a binned core that offers even more headroom.
UPDATE: The Founders Edition is simply the reference design which runs at REFERENCE clocks but uses NVIDIA's premium shroud along with a blower design and low height PCB components. The partners meanwhile will be expected to sell "custom cooled" cards that hit the lower $599 price point.
The GTX 1070 on the other hand hasn’t been discussed all that much but we do know it will have 8GB of standard GDDR5 memory. We’ll have more on that card as its launch day approaches.
Pricing will likely be a hotly debated subject since even though the audience tonight absolutely loved the $599 and $379 costs for the GTX 1080 and GTX 1070, that does represent a not-too-insignificant $50 premium over what the GTX 980 and GTX 970 went for when they were launched in 2014. Now these new generation GPUs arguably have an awesome feature set but so did the Maxwell cards back in their prime.
If there was anything during the presentation that got people out of their seats to cheer, it was the GTX 1080’s performance relative to NVIDIA’s current lineup. In pure gaming situations it is supposed to be faster than the stupidly expensive TITAN X while offering better framerates than a pair of GTX 980 cards in SLI. Essentially, this means we are seeing a potential doubling of performance in just two years while power consumption has remained relatively level. However, some things like which games, resolutions and settings these results were based off of wasn’t discussed.
Actually looking at another NVIDIA-provided chart, things start to look a bit clearer. Once again Pascal dominates in the role of an optimal VR companion but moving onto traditional titles shows a slightly different story. While performance is still phenomenal, instead of an over-100% uplift versus the GTX 980 it’s more like between 65% and 75%. For those keeping track at home, the GTX 780 to GTX 980 would have netted you a relatively meager 25% boost and you’d have to look towards the GTX 680 to GTX 980 transition to get something close (roughly 60%). Essentially, it looks like NVIDIA skipped forward a whole generation based on previous experiences. The question is whether or not DX12 compatibility has been improved but that’s another story for another time.
Moving onto VR environments and the GTX 1080 naturally shines due to its inclusion of a feature called Single Pass Stereo. It is extremely important to take this chart in context and not mix it up with standard gaming scenarios. Nonetheless, these kinds of visuals are carefully crafted affairs that strive to make products look good.
Another area NVIDIA wanted to highlight with the GTX 1080 is how far they’ve come towards making their design one of the most efficient around. Whereas typical power delivery systems always impart a relatively large amount of power loss into the equation, this new generation behaves very differently. Not only is it much more efficient but it also holds those values at a steadier pace throughout the GPU’s power range. In effect this means less power is needed through external sources since more of it is being directed where it needs to go.
Last but not least there’s the new GeForce GTX SLI HB Bridge. While it may first look like a face-lifted bridge connector, it is actually a new High Bandwidth interconnect which doubles the transfer bandwidth of a traditional SLI bridge. Due to the latent processing power of Pascal, it was only a matter of time before NVIDIA needed to introduce a updated multi card connector.
The physical card itself is 10.5” long and quite reminiscent of previous NVIDIA high end cards but there are a few additional angled edges on the shroud. As for I/O connectivity, there’s a single 8-pin power connector, three DisplayPort 1.4 outputs and single connectors for HDMI 2.0b and dual link DVI. That means the maximum resolution would be 7680x4320 at 60Hz and even though 4k HDR wasn’t mentioned, we’d assume that is rolled into the spec as well.
All in I’d say the response to NVIDIA’s new cards has been extremely positive given the fact their competition hasn’t dropped anything particularly concrete about their upcoming solutions. While the hustle and bustle of a grand reveal may hide many truths that eventually come out after more testing, there seems to be a lot to be hopeful about.
It’s also obvious that NVIDIA is willing and able to get these cards into the hands of gamers ASAP since the GTX 1080 will become available on May 27th while you should be able to purchase the GTX 1070 on June 10th. As per their last few launches, NVIDIA is expecting there to be product on the shelves, ready for purchase at that time.
|Latest Reviews in Video Cards|