Make your home smarter than the average home
Make your life smarter, simpler, and more convenient with IoT enabled TVs, speakers, fans, bulbs, locks and more.
Click here to know more
The GP104 is based off the GP100 which is a datacentre-class Tesla GPU. With the increasing hype around the competition’s offerings it was only natural for NVIDIA to step up its game. And that, they have rightfully done. Packing 2560 CUDA cores running at 1607 MHz, NVIDIA claims the GTX 1080 to be 1.5x more efficient in managing power compared to the GTX 980.
Like all previous architectures, the Pascal GP104 also breaks down into Graphics Processing Clusters (GPCs), Streaming Multiprocessors (SMs) and memory controllers. The GP104 consists of four such GPCs and each GPC has 5 SMs and each SM, in turn, has 128 CUDA cores. On the memory front, each GPC has 256 KB register memory, 96 KB shared memory and 48 KB L1 cache memory.
When we compare the GeForce GTX 1080’s GP104 and the GeForce GTX 980’s GM204, you’ll immediately spot the change in the GPCs. Each GPC was configured to hold 4 SMMs in the GM204 while the GP104 squeezes in an extra SM. Other than that you might as well be scratching your head to notice any other difference. This is the first time NVIDIA changed the basic structure of the GPC since the new architecture was released with Fermi’s GF100. And this is the reason for the increase in CUDA cores for the flagship GP104 compared to every flagship they’d shipped earlier. We’re discounting the CUDA Compute 3.5 upgrade brought in with Kepler. Now let’s take a closer look at the individual SMs.
The structure seems to be essentially the same and it is physically similar. Each SM has the same number of Texture Units i.e. eight. However, technology-wise we are looking at a unified memory architecture that has improved since its introduction in CUDA 6 i.e. Kepler. Pascal can now handle larger data sets which means they can store more data in memory, and this translates to better performance.
Especially, for use cases like deep-learning. And if you haven’t caught up on that, then everybody’s been harping about deep learning, both AMD and Intel have been talking quite a lot about this aspect. Now with CUDA 8, not only can the GPU and CPU interact with the same memory pool but the Pascal architecture can interface with HBM2 memory for the Datacentre cards and GDDR5X memory for desktop cards. So they’ve essentially bridged the CPU-GPU divide.
Also, NVIDIA claims that CUDA 8 can compile twice as fast compared to the previous generation. But this has more to do with the folks who’re interested in the parallel processing part of these cards.
The GeForce GTX 1080 now comes with 8 GB of GDDR5X memory. That’s right, the GP104 doesn’t come with HBM like everyone else predicted. Instead, NVIDIA opted to go with GDDR5X which is a one-off thing that Micron came out with wasn’t adopted by JEDEC as a standard until January this year, which was a close call. So what exactly is the ‘X’ for? To summarise what we’re going to rabble on about in the next few lines, it all comes down to higher bandwidth. High binned GDDR5 chips can give you 7 Gbps of bandwidth, however, NVIDIA seems to consider that as inadequate for the GTX 1080.
The key difference with GDDR5X is the 16n prefetch architecture compared to the 8n prefetch architecture of GDDR5. This enables 2^16 = 65536/1024 = 64 Bytes per array of read or write operation. GDDR5 does exactly half, i.e. 32 Bytes per array of read or write. However, this doesn’t translate into a 100% boost in bandwidth, there are a lot more factors that come into play and the very best binned chips can hit 10-12 Gbps. The command protocol, i.e. the way memory talks to anything else digitally remains the same. So it’s essentially GDDR5 on steroids.
There are a lot more other benefits to going with GDDR5X but we’re going to stick to the relevant ones here so that brings into focus power efficiency. GDDR5X operates on 1.35 volts compared to 1.5 volts that GDDR5 uses, this 10% reduction doesn’t seem to be all that much but aside from reduced power consumption this also translates to less heat being generated by the memory chips. However, GDDR5X seems to be a minor blip in the memory horizon and will soon be replaced with GDDR6 which was initially slated to appear by mid-2016 but now seems could be out towards the end of this year or early next year.
Ansel: NVIDIA’s new API is all about in-game photography. Once integrated into your video game it allows you to freely move the camera around so you can take better screenshots ingame. And speaking of photography, Ansel comes with its own set of filters and pushes the GPU to capture HDR images that are up to 8 Gigapixels in size. You can literally zoom in on your favourite character and count their nose hairs.
NVIDIA VRWorks Audio: With VR being a primary focus for the new gen graphics cards, NVIDIA want’s to give you full immersion so they’re now including a new set of APIs that work seamlessly with the graphics card to apply path-tracing to audio. So if you are in game and you hear someone walking towards you in a dark dingy alley, you’ll be able to hear the footsteps coming from the source as well as all the reverberations bouncing off the walls.
Simultaneous Multi-projection: Aimed at the multi-monitor junkies, simultaneous multi-projection ensures that individual monitors from a multi-monitor setup retain perspective. So if you keep two monitors at right angles set it so ingame, you’ll be looking at two distinct scenes. Game developers have to ensure this is implemented as generally textures that are not within the view-cone are disabled.
FastSync: Another sync! Well, you gotta hear them out. V-sync basically limits the game engine to the refresh rate of the monitor ensuring that frame tearing doesn’t happen. But this leads to an increase in latency which is a bane in competitive gaming. Then came G-sync which basically is variable V-sync that goes up or down along with the monitor’s refresh rate which is also variable. Now comes FastSync, why? To combat that latency issue we spoke of. FastSync decouples the render pipeline from display pipeline, so there is now a buffer between the two pipelines. But this helps them render frames fully before sending it out to the display pipeline. This works very well with games that run at excessively high frame rates. A buffer does mean increased latency since anything new adds to the latency, but it’s a minor blip. NVIDIA claims that for CS:GO running at 200 FPS, FastSync will give your full frames with only an 8 millisecond increase in latency.
The Founder’s Edition is a reference card which was till now not available in India but that changed with the GTX 1080. Reference cards are generally built with durable materials. The shroud itself is machined from magnesium-alloy and the cooling fan seems like a dual-ball bearing BLDC fan but we haven’t ripped it apart to confirm. And NVIDIA has added a more geometric design with the GTX 1080. In our benchmarks, we saw it hit 80 degrees in 4K gaming which is a little too much. Hopefully, AIBs will come up with a better cooling solution because the unibody heatsink and the vapour chamber are clearly not enough for the flagship when it comes to 4K gaming. Gaming on 1080p isn’t that taxing and we hardly saw temperatures go past 63 degrees celsius.
Needless to say that flagships pretty much ace every test. We ran the GTX 1080 with the new Hitman, Rise of the Tomb Raider, Witcher 3 and other games. We noticed an increment of 22-30 per cent FPS scores compared to the GTX 980 Ti. Even in synthetic benchmarks like 3DMark FireStrike Ultra, it managed to maintain a similar lead. We didn’t have much time to overclock the card since we got out samples way late and we only had it for a really short time. So in a little quick session with the card we managed to eke out 231 extra MHz over the 1607 MHz stock frequency with a power target of 110%. The card did manage to boot on higher clocks but we couldn’t see a complete 3DMark FireStrike Ultra run on anything higher than 231 MHz.
TThe card is exorbitantly priced in India and it is quite unfair. For `63, 250 this card is way overpriced and it’s easier to have a friend get one for you from the US for roughly Rs. 47,000. At the time of launch third-party cards weren’t available and ecommerce sites had the Founder’s Edition priced for upwards of Rs.71,000. Moreover, with the GTX 1070 out, it hardly seems like the GTX 1080 justifies the price. Hopefully, the prices will come down with base-level AIB cards which are supposed to be $100 cheaper overseas. Which would mean that these AIB cards should cost around Rs.56,000 in India. Another thing to note is that AMD is coming out with the RX 480 around June 29th and certain online benchmarks which can’t be confirmed have pegged 2x RX 480 in crossfire it to be better than a GTX 1080. If this CrossFire configuration costs less than Rs.63,250 then NVIDIA’s in for a good fight.
Base clock: 1607MHz
Memory clock: 1733MHz
Stream processors: 2560
Texture Units: 160
Manufacturing process: 16nm, PCIe 3.0, 7680 x 4320 digital resolution support, 8 GB GDDR5X Memory
OpenGL support: 4.5
Power Connectors: 8Pin
Warranty: 3 years
Phone: +91 022 4376 4567
While not dishing out lethal doses of sarcasm, this curious creature can often be found tinkering with tech, playing 'vidya' games or exploring the darkest corners of the Internets. #PCMasterRace https://www.linkedin.com/in/mithunmohandas/
Digit caters to the largest community of tech buyers, users and enthusiasts in India. The all new Digit in continues the legacy of Thinkdigit.com as one of the largest portals in India committed to technology users and buyers. Digit is also one of the most trusted names when it comes to technology reviews and buying advice and is home to the Digit Test Lab, India's most proficient center for testing and reviewing technology products.
We are about leadership-the 9.9 kind! Building a leading media company out of India.And,grooming new leaders for this promising industry.