Graphics Cards - Page 233

GTC 2016 - On day two of NVIDIA's GPU Technology Conference, we walked around the exhibition halls with the super-awesome Dimitry from Hardware Canucks and ran into the SK Hynix booth. What did we see? HBM2. HBM1 was there side-by-side with a 1GB chip, while the HBM2 chip was a much larger 4GB chip.

Last year at GTC 2015, we did the same thing and walked into HBM1 - because at the time it was only becoming available, and was still months away from AMD using it on their Fiji-based Radeon R9 Fury X. Fast forward a few months, and HBM1 made its way onto the Fury X.

NVIDIA announced its new GP100-based Tesla P100 video card yesterday, which is the first video card to feature the Pascal architecture, and HBM2. Seeing HBM2 at GTC 2016 was great, as we could see how small it was in person, but we weren't allowed to get close to it - which is why the image isn't exactly up close. If we get closer later in the day, we'll take another photo and update the article.

GTC 2016 - NVIDIA just unveiled its next-generation Pascal architecture at GTC 2016, showcasing its new Tesla P100 video card - with an insane amount of power. The Tesla P100 has 16GB of HBM2, with an insane amount of horsepower behind it.

Well, NVIDIA also announced its new DGX-1 - the world's first Deep Learning supercomputer. DGX-1 packs 8x Tesla P100 video cards, each with 16GB of HBM2. DGX-1 has 7TB of SSDs, 2 x Intel Xeon processors, 170 TFlops of performance, Quad IP 100Gbps and dual 10GbE connectivity. It uses just 3200W of power (3.2kW) which might sound like a lot to you and I, but this is a supercomputer... not a traditional rackmount PC.

NVIDIA's new Pascal-based DGX-1 is priced at $129,000.

GTC 2016 - NVIDIA's GPU Technology Conference kicked off today, where the company talked about its work with AI and deep learning were a large focus, as well as VR being a large part of NVIDIA's opening keynote at GTC 2016. But the star of the show was the new Tesla P100.

We have the new Tesla P100, which is "the most advanced hyper scale datacenter GPU ever built". The Pascal chip itself is 600mm2 and built on the incredible 16nm FinFET process with the Tesla P100 packing a how-did-they-do-it 150 billion transistors (150,000,000,000), with the super-fast HBM2 memory included. The new Pascal-based Tesla P100 is in volume production right now, which is incredibly exciting news as it means HBM2 is ready right now - and NVIDIA has it.

Now, the 150 billion transistors sounds absolutely mind-blowing - as its 1775% more than the 8 billion that the GM200 includes (the GPU that powers the GTX 980 Ti and Titan X). We chatted with Rob and Fudo from Fudzilla, who say that the Tesla P100's 150 billion transistors, is made up from 17 billion on the GPU - with the rest of the transistors coming from the HBM2. This makes much more sense.

AMD has today released its Radeon Software Crimson Edition 16.4.1 drivers.

With it comes up to 35% better performance in Quantum Break with the R9 Fury X, a fix for framerate capping issues in some DirectX 12 apps, and a fix for flickering in Hitman when using DirectX 11 mode and high shadow settings; support for HTC Vive and Oculus Rift is listed again for some reason (it was introduced last week in the 16.3.2 drivers).

If you're experiencing any other issues, AMD has some workarounds while they work on fixes:

GTC 2016 - World Exclusive - We've just had some breaking news delivered to us, on NVIDIA's next-generation video card. We don't have specifics to share with you right now, but we do know that it's going to be what we expected: a freakin' monster.

I expected NVIDIA's next-gen video card to be something quite special, thanks to it being based on the Holy Trinity of GPU upgrades - the shift from the 28nm to 16nm FinFET process, the new exciting Pascal architecture and the insanely fast HBM2 memory. Well, according to our exclusive source - NVIDIA's new video card was right on the expectations that I was able to summarise about in the last few months.

We should expect the unannounced video cards to be a beast, with an option between 8/16GB of VRAM or 16/32GB of VRAM. The performance numbers have not been unveiled, and neither has the new naming or branding for the card - but we do know it's going to be insanely powerful. I'm expecting it to be ~1.5x the power of the GeForce GTX Titan X, and the AMD Radeon R9 Fury X (which runs HBM1 @ 512GB/sec).

GTC 2016 - Before the GPU Technology Conference has even kicked off, the new NVIDIA Quadro M5500 has been unveiled, powered by the GM204 GPU. The GM204 is the same GPU that powers the GTX 980, and GTX 970.

The new Quadro M5500 features 8GB of GDDR5 and is the fastest VR ready, workstation-class video card that ships with 8GB of VRAM. The Quadro M5500 features 2048 CUDA cores, 128 TMUs and 64 ROPs, with its GPU boosting up to 1140MHz, just lower than the 1218MHz on the notebook GTX 980 video card.

NVIDIA's new Quadro M5500 is poised for the VR development and workstation market, where it should find a new home in high-end, workstation-class notebooks. These workstation notebooks sport insane Intel Xeon processors, with up to 64GB of DDR4 RAM.

The new Steam hardware survey has some interesting numbers to look at, with the March survey seeing Windows 10 64-bit being the most popular OS on Steam with a huge 37% of users running Microsoft's latest OS.

As for video cards, NVIDIA takes a huge chunk of the Steam gaming pie with 56% of gamers running a GeForce video card of some sort, leaving AMD/ATI with 25%. 17% of Steam gamers are still rocking the integrated GPU on their Intel processor, too.

When it comes to CPUs, Intel continues to take the lion's share of Steam gamers with their CPUs - with 76% of Steam gamers rocking an Intel processor of some sort. 47% of gamers use a dual-core CPU, while 45% run a quad-core processor. Just 4.6% of gamers have a CPU with more than four cores, and 2% are still running a single-core CPU.

Just as I land in San Jose to cover NVIDIA's GPU Technology Conference - where we should be introduced (in detail this time) to their next-gen Pascal architecture, and hopefully a new video card or so - I read about the reported end of GM200 production.

The GM200 might not sound like much, but it's the GPU that powers the GeForce GTX 980 Ti and the Titan X. NVIDIA will reportedly no longer be supplying AIB partners with the GM200 chip, so companies like ASUS, ZOTAC and MSI will not be making anymore GeForce GTX 980 Ti video cards.

Why? Well, there are plenty of GTX 980 Tis in the wild, and with NVIDIA ramping up towards the Pascal launch, we should expect this to be a reality, as we will soon see GP104- and GP100-powered offerings in the hopefully, very near future.

AMD is readying the more professional and scientific oriented version of the dual Fiji Radeon Pro Duo; the FirePro S9300 X2. This is the first time we're seeing Fiji in a professional capacity such as this.

The dual Fiji-based card mirrors nearly every aspect of the consumer (arguably still professional) focused card except that the TDP has been reworked to have a maximum of 300W, now running each GPU at a modest 850MHz. The FirePro is also trading in the great cooling solution for a passive method of keeping the card cool. It's designed to fit into any number of HPC compute applications, where a large water-cooling solution may not always fit properly. And for the 300W TDP, it still manages to be a startlingly powerful card with up to 13.9 TFLOPs of compute power at its disposal. It can potentially be a very powerful card in the right situations.

Fiji presents an interesting choice for the scientific market. The FP64 performance of Fiji is significantly lower by nature of being limited to 1/16th of the FP32 performance. That performance is only really significant for those applications that make use of double precision math, which is selective at best. The great thing is that AMD seems to have found its niche, regardless of that limitation. The higher memory bandwidth afforded by HBM might also be appreciated in these applications as well, even though it also lacks ECC RAM as well.

NVIDIA will show off its next-generation Pascal architecture at their GPU Technology Conference last week, where we should be introduced to the GeForce GTX 1080 and GTX 1070 video cards - with their purported cooler shrouds teased just a couple hours ago.

We could be surprised, and NVIDIA could unveil their new Pascal-powered Titan X successor - which is something I think we'll see. I think NVIDIA will unveil the GP100 GPU, rocking 16-32GB of HBM2 and a price of $1000-$2000 (I'd say $1499 for the 16GB HBM2 version and $1999 for the 32GB HBM2 card). NVIDIA could then drop the GTX 1080 for something like $599, rolling with GDDR5X - still providing a large increase in performance over the GTX 980 and GTX 980 Ti - with thanks given to the 16nm node and the new Pascal architecture.

AMD has just released their new VR Ready Crimson 16.3.2 drivers, which are recommended for GCN cards - as well as the new dual-Fiji video card, the Radeon Pro Duo.

The new Crimson 16.3.2 drivers also provide support for VR headsets in the form of the just-released Oculus Rift and HTC Vive, with Radeon video card owners enjoying AMD VR features like LiquidVR SDK include Asynchronous Shaders, Affinity Multi-GPU and Quick Response Queue.

AMD is committed to VR, as they've stated they have "83% of all tethered VR experiences" powered by Radeon technology. The new Crimson 16.3.2 drivers support for the Oculus Rift SDK v1.3, the Radeon Pro Duo, and are completely optimized for the Rift.

NVIDIA's new 364.72 drivers are here, and represent a major release. Among the inclusions: support for the Oculus Rift and HTC Vive, optimizations for VR titles (Eve Valkyrie, Elite Dangerous, Chronos, Project Cars, and more), and optimizations for non-VR titles (Dark Souls 3, Killer Instinct, Paragon, and Quantum Break).

As you may know, many reports of stability issues have surfaced with NVIDIA's last couple of driver releases, so hopefully, that's overwith with this release; it's encouraging to see many significant stability improvements in the release notes, and to see mostly positive experience reports across the web.

Grab the drivers here or through GeForce Experience now.

AMD has made a huge deal about their Radeon video cards featuring support for Asynchronous Compute, with one of the standout games with Async Compute support being the recently released Hitman from IO Interactive.

During the Game Developers Conference earlier this month, Lead Render Programmer at IO Interactive Jonas Meyer held a discussion called Advanced Graphics Techniques Tutorial Day: Rendering 'Hitman' with DirectX 12. During the session, it was revealed that NVIDIA GeForce video cards had no benefits from Async Compute, with IO Interactive claiming to be working with NVIDIA to fix this.

But, even with their ACE (Asynchronous Compute Engine) on the GCN architecture, Radeon cards only see a 5-10% performance increase. Async Compute is used for SSAA (Screen Space Anti Aliasing), SSAO (Screen Space Ambient Occlusion) and the calculation of light tiles in Hitman. Ashes of the Singularity also makes use of Async Compute, too.

AMD laid out its GPU architecture roadmap through to 2019 at its huge Capsaicin event during the Game Developers Conference, but now we're hearing rumbles on its exciting new Vega GPU - due out in 2017.

Vega will reportedly rock a huge 4096 stream processors based on the Greenland GPU, with improvements in the way of the GCN 4.0 architecture, which are included in the IP v9.0 generation of graphics chips under development from AMD.

We should expect Vega 10 to be AMD's flagship product from the Greenland GPU era - rocking somewhere between 15-18 billion transistors, and the exciting new HBM2 technology which offers up to 1TB/sec of memory bandwidth. Vega looks like it'll be fighting against NVIDIA's compute-powerful GP100 (the Pascal-based successor to the GTX Titan X and GTX 980 Ti) - as Vega is the only HBM2-powered card on AMD's roadmap for 2017.

As we get closer to NVIDIA's GPU Technology Conference, it should come as no surprise that more leaks are arriving on AMD's next-gen Polaris architecture. This time, it's in the form of the Radeon R9 490 and R9 480 video cards.

According to the new leaks, AMD will slap 8GB of the new GDDR5 (and possibly GDDR5X) memory onto the R9 400 series cards, with a 256-bit memory bus. There'll also be 2304 fourth-generation GCN cores, which should be able to easily compete against the likes of NVIDIA's Pascal-based GeForce GTX 1070 and GTX 1080 cards (placeholder names, I really don't see NVIDIA calling them as the GTX 1000 series).

AMD's Polaris cards will be made on the 14nm FinFET process, but remember - these cards are the R9 490 and R9 480, not the R9 490X that should be a little beefier. The R9 490X could be different, where it will feature the faster GDDR5X, while the R9 490 might retain the GDDR5 standard.

AMD's Capsaicin event at GDC was quite the blast, if not for the reveal of their Radeon Pro Duo dual-GPU, VR-focused monster card, they also took time to show off just how potent Polaris 10 actually was. And someone was lucky enough to get a close-up of its behind. And it looks like any other GPU's derrière.

Looking closely, which is easy here, you can see that the connectors include three DisplayPort (presumably 1.3) one HDMI (likely 2.0) and a DVI-D port. And we get our first peek at the prototype PCB as well. Though it's only a prototype, and this could change. There it is. It is an engineering sample, so things could change in the future. The best part is the small form-factor it happens to be in. AMD is definitely committed to bringing more power to smaller form factors.

This is "Big Polaris", or Polaris 10, that's running the newest Hitman using DX12. It's stuffed in a Cooler Master Elite 110 case, meaning the board is minuscule to be able to fit into that case. The PCB is probably around the size of the R9 Nano and should consume less power than any Fiji based board to date.

The dream of higher-bandwidth it a slightly lower cost compared to HBM is coming closer to reality now that Micron has begun shipping samples of their GDDR5X chips to customers for inclusion in prototypes. This means that AMD and NVIDIA are now able to properly test the increase in bandwidth compared to normal GDDR5 and even HBM(2).

It looks like at the moment they're able to ship two different densities, 8Gb and 16Gb that can allow for VRAM of up to 16GB over a 256-bit wide memory bus. Each chip would be relegated to a single 32-bit channel. Don't fret, however, because even though it's a comparatively small memory bus, the internal changes to the structure still allow for far more bandwidth traveling over that bus. It's akin to increasing the speed limit, despite the lane being the same size. The result is that we could see up to 448Gbps of bandwidth, which is similar to first generation HBM, though without the restrictions on memory die size. Power-consumption, too, has been reduced slightly to offset any increase from higher clock speeds and more memory chips on the board.

As of right now it looks like both AMD and NVIDIA are interested in using GDDR5X in their next generation products. From the Capsaicin event, we learned that HBM2 will not be making an appearance until Vega even though the first generation HBM has been confirmed to be part of Polaris alongside traditional GDDR5 and GDDR5X memory. NVIDIA on the other hand will be making great use of Micron's faster tech by likely including 8GB of it in their upcoming GTX 1080, which should be revealed at GTC in April.

NVIDIA has just doubled down on its latest Quadro M6000 professional video card, with the new Quadro M6000 featuring a huge 24GB of GDDR5. The previous model had 12GB of VRAM.

As for what makes the Quadro M6000 tick, it is technically similar to the GTX Titan X, with a full GM200 core and 6 Graphics Processing Clusters. Each of the Graphics Processing Clusters features four SMM (Streaming Multiprocessing Units) with 12 cores in each SMM block. There's 3072 CUDA cores, 192 TMUs, 96 ROPs and 3MB of L2 cache, with six 64-bit memory controllers - with a 384-bit memory bus, and the GPU clocked at 988MHz.

The only difference between the older Quadro M6000 and the new model is that the new model features 24GB of GDDR5, up from the 12GB on the previous M6000. The card has a 225W TDP, with a single 8-pin PCIe power connector, DVI, HDMI and 3 x DisplayPort outputs.

The stars really aligned for AMD during the Game Developers Conference, with their Capsaicin event being a success. The company not only officially unveiled the Radeon Pro Duo (the dual-GPU based on the Fiji architecture) as well as a double down into VR and a tease of their GPU roadmap through to 2019.

On the roadmap itself, we can see Polaris promising a 2.5x performance-per-watt over the current 28nm-based GPUs, but Vega is sitting above it waiting for a early 2017 release - and that's what we're reporting about today; Vega. AMD should begin playing around with Vega sometime this year, since it is being positioned for an early 2017 release and promising the use of HBM2 - which clocks in at 1TB/sec of memory bandwidth (up from 512GB/sec on HBM1).

Vega has been spotted on the Zauba database, thanks to the predictible nomenclature used by AMD. For their Hawaii XT boards, it was the C67101. Tonga was C76501, while Fiji XT was C88001. Now we have the new C9XXXX series cards flowing through, but we don't think they are Polaris boards. The Baffin XT GPU that was on Zauba was the C98101, which is a Polaris card.

GDC 2016 - AMD was all systems go at its Capsaicin event during the Game Developers Conference, unveiling its new dual-GPU video card, the Radeon Pro Duo. The company also talked about its massive commitment to VR, DirectX 12, its next-gen Polaris architecture, and more.

AMD was super confident during this event, where it had a fair amount of hands to play in its battle with NVIDIA. The laser-focused commitment to VR has me excited, as I believe that being a VR-focused company this early on, will only benefit Radeon Technologies Group, and AMD. The company has made partnerships with both Oculus and HTC, for the Rift and Vive, respectively. The company has gone all-in with VR to the point of having its own APU inside of a headset, partnering with Sulon for the Sulon Q headset.

With HDR-enabled TVs and video cards thanks to its next-gen Polaris card, the company had working 14nm at the show. The Radeon Pro Duo was on stage being used during the demonstration, requiring 3 x 8-pin PCIe power connectors to power the dual-GPU video card, rocking 8GB of HBM (4GB per Fiji GPU).