Subject: Graphics Cards | June 20, 2016 - 01:57 PM | Scott Michaud
Tagged: tesla, pascal, nvidia, GP100
GP100, the “Big Pascal” chip that was announced at GTC, will be coming to PCIe for enterprise and supercomputer customers in Q4 2016. Previously, it was only announced using NVIDIA's proprietary connection. In fact, they also gave themselves some lead time with their first-party DGX-1 system, which retails for $129,000 USD, although we expect that was more for yield reasons. Josh calculated that each GPU in that system is worth more than the full wafer that its die was manufactured on.
This brings us to the PCIe versions. Interestingly, they have been down-binned from the NVLink version. The boost clock has been dropped to 1300 MHz, from 1480 MHz, although that is matched with a slightly lower TDP (250W versus the NVLink's 300W). This lowers the FP16 performance to 18.7 TFLOPs, down from 21.2, FP32 performance to 9.3 TFLOPs, down from 10.6, and FP64 performance to 4.7 TFLOPs, down from 5.3. This is where we get to the question: did NVIDIA reduce the clocks to hit a 250W TDP and be compatible with the passive cooling technology that previous Tesla cards utilize, or were the clocks dropped to increase yield?
They are also providing a 12GB version of the PCIe Tesla P100. I didn't realize that GPU vendors could selectively disable HBM2 stacks, but NVIDIA disabled 4GB of memory, which also dropped the bus width to 3072-bit. You would think that the simplicity of the circuit would want to divide work in a power-of-two fashion, but, knowing that they can, it makes me wonder why they did. Again, my first reaction is to question GP100 yield, but you wouldn't think that HBM, being such a small part of the die, is something that they can reclaim a lot of chips by disabling a chunk, right? That is, unless the HBM2 stacks themselves have yield issues -- which would be interesting.
There is also still no word on a 32GB version. Samsung claimed the memory technology, 8GB stacks of HBM2, would be ready for products in Q4 2016 or early 2017. We'll need to wait and see where, when, and why it will appear.
Subject: Graphics Cards | June 18, 2016 - 10:37 PM | Scott Michaud
Tagged: nvidia, graphics drivers
GeForce Hotfix 368.51 drivers have been released by NVIDIA through their support website. This version only officially addresses flickering at high refresh rates, although its number has been incremented quite a bit since the last official release (368.39) so it's possible that it rolls in other changes, too. That said, I haven't heard too many specific issues with 368.39, so I'm not quite sure what that would be.
As always with a hotfix driver, NVIDIA pushed it out with minimal testing. It should pretty much only be installed if you have a specific issue (particularly the listed one(s)) and you don't want to wait until one is released that both NVIDIA and Microsoft looked over (although Microsoft's WHQL certification has been pretty lax since Windows 10).
Oddly enough, they only seem to list 64-bit links for Windows 8.1 and Windows 10. I'm not sure whether this issue doesn't affect Windows 7 and 32-bit versions of 8.1 and 10, or if they just didn't want to push the hotfix out to them for some reason.
Subject: Graphics Cards | June 14, 2016 - 01:46 PM | Jeremy Hellstrom
Tagged: GTX1070, nvidia, overclocking
Overclocking the new Pascal GPUs can be accomplished with the EVGA Precision X tool as it allows you to bump up the power, temperature target and fan speed as well as the frequencies for the GPU and memory easily and effectively. [H]ard|OCP set out to push the 1070 as far as it would go with this software in a recent review. The power target can only be increased to 112%, which they implemented along with setting the fan to 100% as this is about the maximum performance, not about peace and quiet. After quite a bit of testing they settled on 2062MHz GPU and 4252MHz RAM clocks as the highest stable frequency this particular card could manage. The results show a card which leaves the TITAN X in the dirt and this card does not even have a custom cooler; we anxiously await the non-Founders Edition releases to see what they can accomplish.
"In our overclocking review of the NVIDIA GeForce GTX 1070 Founders Edition we will see how far we can overclock the GPU and memory and then compare performance with GeForce GTX TITAN X and GeForce GTX 980 Ti. How high will she go? Can the $449 GTX 1070 outperform a $1000 GTX TITAN X? The answer is exciting."
Here are some more Graphics Card articles from around the web:
- NVIDIA GeForce GTX 1070 On Linux: Testing With OpenGL, OpenCL, CUDA & Vulkan @ Phoronix
- MSI GeForce GTX 1070 Gaming X Review - It's RGB! @ HiTech Legion
- NVIDIA GeForce GTX 1070 8 GB @ techPowerUp
- MSI GTX 1080 & GTX 1070 Gaming X 8G Review @ OCC
- Deep Learning & CUDA Benchmarks On The GeForce GTX 1080 Under Linux @ Phoronix
- Gigabyte GTX 1080 G1 Gaming 8 GB @ techPowerUp
- Gigabyte GeForce GTX 1080 G1 GAMING @ Guru of 3D
- MSI GTX 1080 Gaming X 8 GB @ techPowerUp
Subject: Graphics Cards | June 8, 2016 - 08:44 PM | Ryan Shrout
Tagged: sli, pascal, nvidia, GTX 1080, GP104, geforce, 4-way sli, 3-way sli
IMPORTANT UPDATE: After writing this story, but before publication, we went to NVIDIA for comment. As we were getting ready to publish, the company updated me with a shift in its stance on multi-GPU configurations. NVIDIA will no longer require an "enthusiast key" to enable SLI on more than two GPUs. However, NVIDIA will also only be enabling 3-Way and 4-Way SLI for a select few applications. More details are at the bottom of the story!
You'll likely recall that during our initial review of the GeForce GTX 1080 Founders Edition graphics card, we mentioned that NVIDIA was going to be moving people towards the idea that "only 2-Way SLI will be supported" and promoted. There would still be a path for users that wanted 3 and 4 GPU configurations anyway, and it would be called the Enthusiast Key.
As it turns out, after returning from an AMD event focused on its upcoming Polaris GPUs, I happen to have amassed a total of four GeForce GTX 1080 cards.
Courtesy of some friends at EVGA and two readers that were awesome enough to let me open up their brand new hardware for a day or so, I was able to go through the 3-Way and 4-Way SLI configuration process. Once all four were installed, and I must point out how great it is that each card only required a single 8-pin power connector, I installed the latest NVIDIA driver I had on hand, 368.19.
Knowing about the need for the Enthusiast Key, and also knowing that I did not yet have one and that the website that was supposed to be live to enable me to get one is still not live, I thought I might have stumbled upon some magic. The driver appeared to let me enable SLI anyway.
Enthusiasts will note however that the green marker under the four GPUs with the "SLI" text is clearly only pointing at two of the GTX 1080s, leaving the remaining two...unused. Crap.
At this point, if you have purchased more than two GeForce GTX 1080 cards are simply out of luck and are waiting on NVIDIA to make good on it's promise to allow for 3-Way and 4-Way configurations via the Enthusiast Key. Or some other way. It's way too late now to simply say "we aren't supporting it at all."
While I wait...what is there for a gamer with four GeForce GTX 1080 cards to do? Well, you could run Ashes of the Singularity. It's multi-GPU mode uses MDA mode, which means the game engine itself accesses each GPU on its own, without the need for the driver to handle anything regarding GPU load balancing. Unfortunately, Ashes only supports two GPUs today.
Well...you could run an OpenCL based benchmark like LuxMark that access all the GPUs independently as well.
I did so, and the result is an impressive score of 17,127!!
How does that compare to some other products?
The four GTX 1080 cards produce a score that is 2.57x the result provided by the AMD Radeon Pro Duo and 2.29x the score of SLI GeForce GTX 980 Ti cards. Nice!
So there you go! We are just as eager to get our hands on the ability to test 3-Way and 4-Way SLI with new Pascal GPUs as some of the most extreme and dedicated enthusiasts out there are. With any luck, NVIDIA will finally figure out a way to allow it - no matter how it finally takes place.
IMPORTANT UPDATE: Before going to press with this story I asked NVIDIA for comment directly: when was the community finally going to get the Enthusiast Key website to unlock 3-Way and 4-Way SLI for those people crazy enough to have purchased that many GTX 1080s? The answer was quite surprising: NVIDIA is backing away from the idea of an "Enthusiast Key" and will no longer require it for enabling 3-Way and 4-Way SLI.
Here is the official NVIDIA statement given to PC Perspective on the subject:
With the GeForce 10-series we’re investing heavily in 2-way SLI with our new High Bandwidth bridge (which doubles the SLI bandwidth for faster, smoother gaming at ultra-high resolutions and refresh rates) and NVIDIA Game Ready Driver SLI profiles. To ensure the best possible gaming experience on our GeForce 10-series GPUs, we’re focusing our efforts on 2-way SLI only and will continue to include 2-way SLI profiles in our Game Ready Drivers.
DX12 and NVIDIA VR Works SLI technology also allows developers to directly implement and control multi-GPU support within their games. If a developer chooses to use these technologies then their game will not need SLI profiles. Some developers may also decide to support more than 2 GPUs in their games. We continue to work with all developers creating games and VR applications that take advantage of 2 or more GPUs to make sure they’ll work great on GeForce 10-series GPUs.
For our overclocking community, our Game Ready Drivers will also include SLI profiles for 3- and 4-way configurations for specific OC applications only, including Fire Strike, Unigine and Catzilla.
NVIDIA clearly wants to reiterate that only 2-Way SLI will get the attention that we have come to expect from the GeForce driver dev team. As DX12 and Vulkan next-generation APIs become more prolific, the game developers will still have the ability to directly access more than two GeForce GTX 10-series GPUs, though I expect that be a very narrow window of games simply due to development costs and time.
NVIDIA will enable support for three and four card configurations in future drivers (without a key) for specific overclocking/benchmarking tools only, as a way to make sure the GeForce brand doesn't fall off the 3DMark charts. Only those specific applications will be able operate in the 3-Way and 4-Way SLI configurations that you have come to know. There are no profiles to change manually and even the rare games that might have "just worked" with three or four GPUs will not take advantage of more than two GTX 10-series cards. It's fair to say at this point that except for the benchmarking crowd, NVIDIA 3-Way and 4-Way SLI is over.
We expect the "benchmark only" mode of 3-Way and 4-Way SLI to be ready for consumers with the next "Game Ready" driver release. If you happened to get your hands on more than two GTX 1080s but aren't into benchmarking, then find those receipts and send a couple back.
So there you have it. Honestly, this is what I was expecting from NVIDIA with the initial launch of Pascal and the GeForce GTX 1080/1070 and I was surprised when I first heard about the idea of the "enthusiast key." It took a bit longer than expected, and NVIDIA will get more flak for the iterated dismissal of this very niche, but still pretty cool, technology. In the end, this won't have much impact on the company's bottom line as the quantity of users that were buying 3+ GTX GPUs for a single system was understandably small.
Subject: Graphics Cards | June 8, 2016 - 06:26 PM | Scott Michaud
Tagged: nvidia, giveaway, e3 2016, E3
Update, June 8th @ 8:15pm: Just to clarify, this giveaway is not affiliated with PC Perspective. We just found it on Twitter and thought that our readers might like to have a chance at free hardware.
Fairly simple bit of news for this one. NVIDIA has announced that they will be giving away $100,000 of prizes to people who message @NVIDIA and use the #GameReady hashtag, on either Twitter or Instagram, during one of five E3 keynotes.
Sunday (June 12th, 2016):
- EA at 1PM PDT / 4PM EDT / 8PM GMT
- Bethesda at 7PM PDT / 10PM EDT / 2AM GMT (Monday)
Monday (June 13th, 2016):
- Microsoft at 9:30AM PDT / 12:30PM EDT / 4:30PM GMT
- PC Gaming Show at 12PM PDT / 3PM EDT / 7PM GMT
- Ubisoft at 1PM PDT / 4PM EDT / 8PM GMT
Interestingly, Sony was not listed on their rundown. Sure, they rarely have anything relevant to PC gamers, but it's still an amusing omission none-the-less.
According to their Terms and Conditions, the sweepstakes is open to a large portion of the world. They will be giving away fifty GTX 1080s, “up to” thirty $500 Steam Gift Cards, and “an ultimate PC battlestation”??? I'm not sure what that is, but it sounds like Mark Hamill will be trying to destroy it a few times.
E3 starts this weekend! Stay tuned for coverage. (You can also sleep, eat, and do laundry, though.)
Subject: Graphics Cards | June 8, 2016 - 02:11 AM | Scott Michaud
Tagged: nvidia, graphics drivers
NVIDIA has released a new graphics driver, in line with EA's new title, Mirror's Edge: Catalyst. Version 368.39 is another of their WHQL-certified, Game Ready-branded drivers that integrates all of their tweaks to improve the game's performance, including an updated SLI profile. It also includes performance tweaks for Insomniac's Oculus-exclusive VR title, Edge of Nowhere, which released on June 6th.
Beyond performance enhancements for specific titles, the driver also includes new features and fixes to known bugs. On the feature side of things, a handful of OpenGL extensions were added to support new features in Pascal. Extensions allow hardware vendors to add features without the Khronos Group needing to officially support it in the standard (although many turn multi-vendor and eventually end up in a later core specification). In this case, NVIDIA has added Single Pass Stereo to increase VR performance, Lens Matched Shading to also increase VR performance, Improved Conservative Rasterization to reduce the chance that a pixel fragment will be missed during rasterization of degenerate or otherwise odd geometry, and Double Precision Atomic Operations to increase reliability when doing GPU-compute on 64-bit, double-precision values in OpenGL.
On Windows 10, seven bugs were fixed in 368.39, and two of those were fairly high profile. First, the GTX 1080 Founders Edition fan speed revving issue has been fixed, as NVIDIA mentioned a few days ago. Second, performance issues (stuttering) in Total War: WARHAMMER were fixed. They also fixed an issue where Metal Gear Solid V would fail to launch (white screen).
The new drivers are available on GeForce Experience or their website.
Subject: Graphics Cards | June 7, 2016 - 08:07 PM | Scott Michaud
Tagged: zotac, pascal, nvidia, GTX 1080, GP104, asus
Update @ 10:30pm, June 7th: Annnnnnnnd it's gone.
Update @ 9:45pm, June 7th: ASUS is now out-of-stock, so I crossed out the relevant links. ZOTAC is still around for now.
Update @ 8:45pm, June 7th: Turns out that it's also available on Newegg US. In fact, it's possible that both sites share from the same stock pool, at least for the
US ASUS and US ZOTAC cards, given that Newegg Canada says it ships them from the US.
A couple of GeForce GTX 1080s are available at Newegg Canada at the moment. Both models, one from
ASUS and one from ZOTAC, are listed at $909. This seems high, but it's actually the current US-to-Canada exchange rate from the $699 MSRP. If you were interested in the Founders Edition cards, then you have a brief moment to pick one up.
That said, it's looking like the custom-cooled versions might be a better bet. The EVGA dual-fan GAMING SC ACX 3.0 version is listed at $824.99 CDN (~$635 USD) and, from what we've seen so far, seems to be quite a bit cooler than the Founders Edition (albeit we haven't tested sound levels yet). Those should be coming out fairly soon, and will apparently lean on the cheaper side of the Founders Edition fence.
But, if you don't care, go go go go go.
Subject: Graphics Cards | June 7, 2016 - 03:50 PM | Jeremy Hellstrom
Tagged: geforce, GP104, gtx 1070, nvidia, pascal
With Computex behind us it is time to catch up on all the reviews which were launched during the show, including [H]ard|OCP's review of the GTX 1070 Founders Edition. Their testing was done using an NVIDIA provided driver, GeForce 368.19 the same one which Ryan used in his review. They did not have a chance to delve into overclocking or utilizing the new power settings. From their testing they concluded the GTX 1070 is a great upgrade for those using a vanilla GTX 980 or R9 390X, while the card performs faster than a R9 Fury X or GTX 980 Ti the jump is not quite enough to recommend dumping it for anything less than a GTX1080.
"In our review of the NVIDIA GeForce GTX 1070 Founders Edition video card we will explore the price competitive performance and find out what kind of gameplay advantage the new GeForce GTX 1070 Founders Edition offers over the previous generation cards. We compare both the GTX 980 and Radeon R9 Fury GPUs to the new GTX 1070"
Here are some more Graphics Card articles from around the web:
- Asus Republic Of Gamers Strix GTX 1080 Aura RGB OC @ Kitguru
- Nvidia GTX 1070 Founders Edition @ Kitguru
- NVIDIA GeForce GTX 1080 Overclocking & Best Playable Settings At 4K & Ultrawide @ Techgage
- NVIDIA GeForce GTX 1070 Overclocking Review @ OCC
- NVIDIA GeForce GTX 1080 On Linux: OpenGL, OpenCL, Vulkan Performance @ Phoronix
- Performance & Perf-Per-Watt From NVIDIA's GeForce 9800GTX To GTX 1080 @ Phoronix
- OpenGL Performance & Perf-Per-Watt From The Radeon HD 3850 Through R9 Fury @ Phoronix
Subject: Graphics Cards, Motherboards | June 6, 2016 - 05:14 PM | Scott Michaud
Tagged: pascal, nvidia, motherboard, gtx 1070, GP104, colorful
So here's an interesting bit of news from Colorful, via Videocardz and LG Nilsson. Remember when on-board graphics was a pejorative? Since the GPUs that are attached to many CPUs tend to sufficiently cover everything below a discrete graphics add-in board, there is not a whole lot of mind-share for discrete, on-board GPUs. You get the occasional desktop-style device with a mobile add-in module, but that's about it.
Image Credit: LG Nilsson (via Videocardz)
In this case, it looks like Colorful, the Chinese PC hardware manufacturer, integrated the required components from a GTX 1070 directly onto a motherboard's PCB. We've heard rumors that GP104 would be available in mobile form-factors in a few months, so it's possible that this draws from some laptop initiatives, but it's interesting to see others consider it too. As Videocardz pointed out, this is not an ATX-standard board, so it's possible that Colorful is planning on getting into (or supporting someone getting into) small form factor desktops or is building hardware for all-in-one PCs.
So what's next? A vendor like ASUS making a VRWorks Audio sound card with integrated Pascal?
Subject: Graphics Cards, Mobile | June 4, 2016 - 04:28 PM | Scott Michaud
Tagged: nvidia, GTX 1080, gtx 1070, pascal
Normally, when a GPU developer creates a laptop SKU, they re-use the desktop branding, add an M at the end, but release a very different, significantly slower part. This changed with the GTX 980, as NVIDIA cherry-picked the heck out of their production to find chips that could operate full-speed at a lower-than-usual TDP. With less power (and cooling) to consider, they were sent to laptop manufacturers and integrated into high-end designs.
They still had the lower-performance 980M, though, which was confusing for potential customers. You needed to know to avoid the M, and trust the product page to correctly add the M as applicable. This is where PCGamer's scoop comes into play. Apparently, NVIDIA will stop “producing separate M versions of its desktop GPUs”. Also, they are expected to release their 10-series desktop GPUs to their laptop partners by late-summer.
Last time, NVIDIA took almost a year to bin enough GPUs for laptops. While we don't know how long they've been stockpiling GP104 GPUs, this, if the rumors are true, would just be about three months of lead-time for the desktop SKUs. Granted, Pascal is significantly more efficient than Maxwell. Maxwell tried to squeeze extra performance out of an existing fabrication node, while Pascal is a relatively smaller chip, benefiting from the industry's double-shrink in process technology. It's possible that they didn't need to drop the TDP threshold that far below what they accept for desktop.
For us desktop users, this also suggests that NVIDIA is not having too many issues with yield in general. I mean, if they were expecting GPU shortages to persist for months, you wouldn't expect that they would cut their supply further with a new product segment, particularly one that should require both decent volume and well-binned chips. This, again, might mean that we'll see desktop GPUs restock soon. Either that, or NVIDIA significantly miscalculated demand for new GPUs, and they needed to fulfill partner obligations that they made before reality struck.
Call it wishful thinking, but I don't think it's the latter.