TweakTown NewsRefine News by Category:
GTC 2015 - One of the things we didn't quite have the time to talk about yesterday was the 3D Memory and HBM 2.0 side of the upcoming Pascal architecture form NVIDIA. This architecture should find itself onto a new VGA card sometime in 2016, rocking a huge 32GB of VRAM.
NVIDIA briefed the crowd at its GPU Technology Conference here in San Jose, California where they displayed a slide with the amount of VRAM per GPU architecture. We can see that Kepler in 2012, which the last flagship card based on Kepler was the GeForce GTX 780 and GTX Titan Black featuring 3GB and 6GB of RAM respectively, while the Maxwell architecture provides between 4GB (on the GTX 980) and 12GB (on the GTX Titan X) respectively. The slide teases that Pascal will feature 32GB of RAM, and Volta will rock up to 72GB of RAM in 2018.
The way NVIDIA will do this is thanks to SK Hynix's High Bandwidth Memory (HBM) which allows for four-layer stacks, which is also known as 4-Hi. This will come in 1GB and 2GB varieties, but eight-layer stacks will eventually arrive, which should see a huge increase in the amount of framebuffer on the next generation of GPUs.
Not only will Pascal deliver more VRAM on the card, but it will have magnitudes more memory bandwidth. The Maxwell-based GeForce GTX Titan X has 336GB/sec of memory bandwidth from its 384-bit memory bus on its GDDR5 RAM, but the Pascal architecture will be capable of a huge 750GB/sec or more. This will be using a variety of technologies to achieve this lofty height of memory bandwidth, including mixed precision, 3D Memory and NVLink.
We should expect NVIDIA to talk more about Pascal later in the year, or GTC 2016 this time next year.
GTC 2015 - NVIDIA has just unveiled the new Linux-based Digits Devbox which is not a mass production platform, but a PC that is built as the "world's fastest deskside deep learning system" priced at a huge $15,000.
Inside, you'll find four of the new Maxwell-powered GeForce GTX Titan X cards, which will super accelerate deep learning with its insane GPU capabilities. The system is built-to-order, where you have a direct contact at NVIDIA once you purchase the system to help you out. This isn't a gaming PC, but we're sure this system could probably run Crysis.
GTC 2015 - NVIDIA has just started providing many more details on its next generation GPU architecture, known as Pascal. Pascal is going to be a giant leap from Maxwell, where it will be "10x faster than Maxwell" according to co-founder and CEO of NVIDIA, Jen-Hsun Huang.
Pascal will feature NVLink and 3D Memory which will provide some impressive performance and bandwidth, where we can expect 3x the memory bandwidth of Maxwell, which is quite the leak. On top of that, we have 400% more mixed precision, but the Pascal architecture itself is a huge leap for NVIDIA, even compared to the already impressive Maxwell architecture.
We have NVLink which has 5x the bandwidth of PCI Express, which is quite the jump in terms of available bandwidth. NVLink will provide the ability of having 8 GPUs, up from the 4-way limit of SLI right now.
GTC 2015 - NVIDIA has just made its new GM200-powered GeForce GTX Titan X official at its GPU Technology Conference, but the company is really pushing the deep learning side of things.
The new Maxwell-based GeForce GTX Titan X is capable of training AlexNet for deep learning, at magnitudes faster than previous GPUs like the Titan, Titan Black and even a 16-core Xeon CPU from Intel. The Xeon CPU with 16 cores takes 43 days to train AlexNet, but Titan X drags this down to just three days. Even the Titan Black sits at five days, while the Titan is at 6 days.
GTC 2015 - NVIDIA has made its GeForce GTX Titan X official at the GPU Technology Conference in San Jose, based on the full GM200 core. The GeForce GTX Titan X packs quite the performance punch with 8 billion transistors and 12GB of VRAM.
The new GeForce GTX Titan X is the most powerful GPU from NVIDIA yet, with the full GM200 core compared to the 'cut down' GM204 that makes the GTX 980 tick. The new Titan X is a powerhouse VGA card, with 30-50% additional performance of the GTX 980, and in some cases, much more. That's without taking into consideration the 12GB of framebuffer, too.
We have 3072 CUDA cores, 7 TFlops of performance and the huge 12GB framebuffer. We will have more details on the Titan X shortly. NVIDIA has also announced a price of just $999.
Hours before NVIDIA takes the veil off of its GM200-based GeForce GTX Titan X at GTC 2015, there are leaked details on specifications and performance for AMD's upcoming Fiji-based Radeon R9 390X.
According to the latest leaks, the new Radeon R9 390X with HBM memory will be up to 60% faster than the Radeon R9 290X with its 4GB of GDDR5 RAM. The leaked specs see the Fiji XT model featuring 4096 stream processors, a 4096-bit memory bus with up to 8GB of HBM memory, something that will provide 600GB/sec+ of memory bandwidth.
Comparing this to the R9 290X, we have 8.6 TFlops of performance on the R9 390X, which is a huge leap on the 5.6 TFlops that the R9 290X is capable of. HBM is said to be "leaps ahead" when it comes to video memory technology, where we have "drastically increased power efficiency" compared to GDDR5, and so much more. The R9 390X has full support for DirectX 12 and Mantle, with "new dual-GPU modes with upcoming drivers" being teased, as well as an "immersive VR experience with AMD LiquidVR".
Then we have the interesting model: the Radeon R9 390X WCE, which will come with watercooling by default, something that reviewers like us will receive. We don't know if this model will make it to the public, but the R9 390X WCE is "designed for 4K & VR gaming" according to the leaked slides.
We've heard that the Radeon R9 390X will arrive in both a 4GB and 8GB version, but the latest rumor is that AMD will release a Water-Cooled Edition as the Radeon R9 390X WCE.
The news is coming from a leaked slide which shows the R9 390X featuring 4096 shader units, DirectX 12 support, "optimized for 4K gaming & beyond", and it will arrive in a "special enthusiast water-cooled edition" with up to 8GB of ultra-high bandwidth HBM video memory. The more we hear about the R9 390X, the more we want to see it.
NVIDIA's GPU Technology Conference is right around the corner, kicking off on Tuesday, but we've already been introduced to the GM200-based GeForce GTX Titan X, and now the Maxwell-based Quadro M6000... courtesy of electronic music star Deadmau5.
Deadmau5 posted the image to his Instagram account, with two NVIDIA Quadro M6000 cards sitting in his rendered machine. We can't see the PCIe power connectors, but considering how power efficient the Maxwell architecture is, we can expect dual 6-pin power connectors at most. Comparing the Quadro K6000 and M6000 against each other, the M6000 should be quite the jump for users.
The new M6000 should feature 24 SMMs compared to the 15 found on the K6000, while we have 3072 GPU cores, compared to the 2880 Kepler-based cores found in the K6000. 96 ROPs with 192 TMUs on the M6000 versus the 48 ROPs and 240 TMUs on the K6000, which will result in 6.07 TFlops of performance versus the K6000 and its 5.2 TFlops.
We should hear more about this during the week at NVIDIA's GTC 2015 event, which we will be covering live.
We've only been introduced to the GeForce GTX Titan X, but now we're hearing about the GeForce GTX 980 Ti which will also reportedly be based on the GM200 core, but cut down. Not only that, in some leaked benchmarks both of those cards are being compared against the upcoming AMD Radeon R9 390X.
According to the benchmarks, the new Radeon R9 390X looks like quite the force to be reckoned with at 4K, where it beats the R9 290X by 49%. It even beats the new Titan X and the cut down GM200 or GTX 980 Ti. Keep in mind that the above benchmarks are not in FPS, but in relative % to the Radeon R9 290X. We also have an unreleased GTX 96* GPU (GTX 965 or GTX 960 Ti) that you'll find lower on that list.
The same can be said for 2560x1600, where the Fiji-based R9 390X continues to dominate with a 40% jump on the R9 290X. It beats the Titan X by 3%, and the GM200-based GTX 980 Ti by just over 10%. But what about one of the more important aspects of the new architecture: power consumption? This is where, according to the leaked benchmarks, it looks like AMD has done something incredible: increased the horsepower of the 390X over the 290X, but kept to the same TDP.
NVIDIA has already teased its GeForce GTX Titan X which should arrive next week at the company's GPU Technology Conference next week, but now we're hearing about the new Fiji-based Radeon R9 390X, and that it will come in two variants: one with 4GB of VRAM, the other with 8GB.
Because AMD will be using High Bandwidth Memory (HBM) on these cards, the 8GB version will not have the numbers to go mainstream compared to its 4GB counterpart. The 8GB version will require eight 8Gbit chips (1GB per chip) that will provide a huge 1024MB/sec of bandwidth. This card will be quite the force to be reckoned with, but considering the GeForce GTX 980 already performs super well at 4K with its 256-bit memory bus, the 1024-bit memory bus and this next-gen HBM RAM should usher in some 4K numbers we haven't seen before.
We are expecting the new Fiji-based R9 390X to be unveiled at Computex in June.