Support NeoGAF

[nogoodnamesleft]

Any estimate on when the next generation nvidia cards will be out to the public?

 

[Shin]

I'm confused by this: 16GB HBM2, 900GB/s bandwidth.
That's 4GB/s less than 16Gb GDDR5x, doesn't make sense.

 

[dgrdsv]

I'm confused by this: 16GB HBM2, 900GB/s bandwidth.
That's 4GB/s less than 16Gb GDDR5x, doesn't make sense.

That's 350GB/s more actually, if we compare to the Titan Xp's fastest GDDR5X on a 384 bit bus.

 

[RedlineRonin]

Needed a significant increase in transistor count and die size to get there though.

Still, amazing power efficiency improvements.

I see.

Am I right in thinking (obviously just taking shots in the dark) but that 50% improvement would be a rough estimate of improvement for performance of whatever the 1180 would be over the 1080?

I'm curious if Volta has the potential for true 4K60 on a single card, or if that's another gen out past Volta yet.

 

[dgrdsv]

I see.

Am I right in thinking (obviously just taking shots in the dark) but that 50% improvement would be a rough estimate of improvement for performance of whatever the 1180 would be over the 1080?

I'm curious if Volta has the potential for true 4K60 on a single card, or if that's another gen out past Volta yet.

50% improvement is pretty much expected on the same process by going with larger dies. However we still have no idea when and if gaming Volta will even launch.

 

[Sir Abacus]

50% improvement is pretty much expected on the same process by going with larger dies. However we still have no idea when and if gaming Volta will even launch.

We could extrapolate from Maxwell and Pascal? At the current cadence we should expect an 1180 with 56 SMs, 3584 CUDA cores, and 256-bit memory bus in January 2018. April 2018 we should expect an 82 SM, 5248 CUDA core Titan V (TO THE MOON ALICE) with a full 384-bit memory bus, maybe a HBM2 variant? Then in January 2019 an 1180 Ti with 5248 CUDA cores, 384-bit memory bus, maybe a fully enabled 84 SM Titan V variant released with HBM2 to differentiate a couple of months earlier?

They follow a pretty constant pattern: People who can pay squillions get them first when yields are lowest. Then comes the mid variant which beats the old large variant at a lower price point. Then shortly after that the little variant which delivers value. Then once the big chip yields improve they start handing out the ones with a couple of faulty SM or two to rich people who pay twice as much for the next big thing a year early and then, finally, the plebs like us once volume and yields improve.

 

[dgrdsv]

We could extrapolate from Maxwell and Pascal?

We could but you have to remember several things:

a) Volta to Pascal is more like Maxwell to Kepler, not Pascal to Maxwell.

b) The market is completely different now than it was back in Kepler and even Maxwell days, what made sense back then may not do so now.

I still think that if there will be a gaming Volta we'll see it's first chip this Autumn. If we won't then I have some doubts about Volta arch ever being used for gaming GPUs because by spring time NV will already have something new.

 

[Pineapple]

Isn't Volta supposed to be the brand new architecture to replace Maxwell/Pascal? I actually used to think Pascal was a new archiecture, but apparently it's just Maxwell with faster clocks and a few slight improvements. Wasn't Volta supposed to be the next big thing?

Also, how does memory bandwidth translate into gaming performance? For example, HBM(2) offers massive bandwidth increased relative to GDDR5, but what would be the actual result of this? Would you need to be playing at 4K or higher in order to see any tangible performance benefit?

 

[riflen]

Isn't Volta supposed to be the brand new architecture to replace Maxwell/Pascal? I actually used to think Pascal was a new archiecture, but apparently it's just Maxwell with faster clocks and a few slight improvements. Wasn't Volta supposed to be the next big thing?

Also, how does memory bandwidth translate into gaming performance? For example, HBM(2) offers massive bandwidth increased relative to GDDR5, but what would be the actual result of this? Would you need to be playing at 4K or higher in order to see any tangible performance benefit?

Compared to GDDR5, HBM has massive bandwidth improvements. However:
1. It increases complexity massively, which increases cost and impacts yields.
2. GDDR5X was developed rapidly and quickly ate HBM's lunch with comparable performance, greater availability and lower cost.

In games, it's going to be very difficult to quantify in a sentence what the impact of memory bandwidth will be. Essentially it varies from game to game and also varies from one GPU configuration to another. You will only ever know how bandwidth affects one design by increasing the memory clock and testing. A lot.

For some idea, look at the recently released GTX 1060 and 1080 designs that use improved GDDR5(X). The performance increase is a few percent at best. Really, a good design shouldn't be massively improved by increasing memory bandwidth alone. The design should strike a balance in all areas.

 

[AmyS]

Nvidia evaluating multi chip module GPU designs

http://hexus.net/tech/news/graphics/107560-nvidia-evaluating-multi-chip-module-gpu-designs/

Researchers from Arizona State University, Nvidia, the University of Texas, and the Barcelona Supercomputing Centre have published a paper (PDF) that looks at improving GPU performance using Multi-Chip-Module (MCM) GPUs. The team see MCM GPUs as one way to sidestep the deceleration of Moore's Law and the performance plateau predicated for single monolithic GPUs.

Transistor scaling cannot happen at historical rates anymore and chipmakers are staying with certain manufacturing processes longer but optimising performance in other ways. As "the performance curve of single monolithic GPUs will ultimately plateau," researchers are looking at how to make better performing GPUs from package-level integration of multiple GPU modules.

It is proposed that easily manufacturable basic GPU Modules (GPMs) are integrated on a package "using high bandwidth and power efficient signalling technologies," to create multi chip module GPU designs. To see if such a proposal is worthwhile and can bear fruit worth picking, the research team has been evaluating designs using Nvidia's in-house GPU simulator. Theoretical performance comparisons against multi-GPU solutions were also made.

MCM GPUs could do wonders for increasing the SM count and many GPU applications "scale very well with increasing number of SMs," observe the scientists. The research team looked at the possibilities of a 256 SMs MCM-GPU in the paper, and are pleased by its potential. Using the simpler GPM building blocks and advanced interconnects this 256 SM chip "achieves 45.5% speedup over the largest possible monolithic GPU with 128 SMs," assert the researchers

More articles, recently posted:

http://fudzilla.com/news/graphics/44022-nvidia-looking-at-multi-chip-module-gpus
http://techreport.com/news/32189/nvidia-explores-ways-of-cramming-many-gpus-onto-one-package
https://www.neowin.net/news/nvidia-researching-multi-chip-module-gpus-to-keep-moores-law-alive

From the PDF:

http://research.nvidia.com/sites/de...U:-Multi-Chip-Module-GPUs//p320-Arunkumar.pdf

Many of today's important GPU applications scale well with GPU compute capabilities and future progress in many fields such as exascale computing and artificial intelligence will depend on continued GPU performance growth. The greatest challenge towards building more powerful GPUs comes from reaching the end of transistor density scaling, combined with the inability to further grow the area of a single monolithic GPU die. In this paper we propose MCM-GPU, a novel GPU architecture that extends GPU performance scaling at a package level, beyond what is possible today. We do this by partitioning the GPU into easily manufacturable basic building blocks (GPMs), and by taking advantage of the advances in signaling technologies developed by the circuits community to connect GPMs on-package in an energy efficient manner. We discuss the details of the MCM-GPU architecture and show that our MCM-GPU design naturally lends itself to many of the historical observations that have been made in NUMA systems. We explore the interplay of hardware caches, CTA scheduling, and data placement in MCM-GPUs to optimize this architecture. We show that with these optimizations, a 256 SMs MCM-GPU achieves 45.5% speedup over the largest possible monolithic GPU with 128 SMs. Furthermore, it performs 26.8% better than an equally equipped discrete multi-GPU, and its performance is within 10% of that of a hypothetical monolithic GPU that cannot be built based on today's technology roadmap.

Seems this is the way forward for Nvidia in 2019+

 

[The Wart]

Nvidia evaluating multi chip module GPU designs

http://hexus.net/tech/news/graphics/107560-nvidia-evaluating-multi-chip-module-gpu-designs/

Interesting. Would this require them to effectively "solve" the problem of SLI support?

 

[AmyS]

Interesting. Would this require them to effectively "solve" the problem of SLI support?

One can only hope!

 

[TyphoonStruggle]

Seems this is the way forward for Nvidia in 2019+

I'm not sure why you would say such a thing. The one you provided seems to have been lead by a Ph.D student. Are there other journal articles under the same topic that were published by Nvidia?

 

[icecold1983]

Interesting. Would this require them to effectively "solve" the problem of SLI support?

no, not necessarily

 

[CazTGG]

Any news on Xavier?

 

[The Wart]

no, not necessarily

Similar issues would necessarily arise with a multichip setup, no? Though I guess the solution they come up would not necessarily be generalizable to mGPU, or they might not bother doing so.

 

[CariusD]

Similar issues would necessarily arise with a multichip setup, no? Though I guess the solution they come up would not necessarily be generalizable to mGPU, or they might not bother doing so.

No they even talk about dual GPU style cards like AMD x2 and their unsolved problems and how this tech is much better. SLI will always be a completely different beast because of data syncing issues.

 

[icecold1983]

Similar issues would necessarily arise with a multichip setup, no? Though I guess the solution they come up would not necessarily be generalizable to mGPU, or they might not bother doing so.

to completely solve the issues mgpu exhibits requires better hardware and dev support. a fully encapsulated solution by nvidia will likely always have worse scaling than a single gpu with the same amt of resources. increased latency is also unavoidable

 

[dgrdsv]

Any news on Xavier?

Lots of NV related automotive news lately. Not sure why this is relevant though as Xavier is basically an automotive only SoC.

Interesting. Would this require them to effectively "solve" the problem of SLI support?

This is for HPC/supercomputers first and foremost. There is no problem of SLI support in compute.

 

[dgrdsv]

First Tesla V100 PCIE cards are out on the market now: Big Surprise for Top AI Brainiacs: NVIDIA CEO Gives World's Top AI Researchers First NVIDIA Tesla V100s

 

[tkscz]

Hopefully HBM will start catching on after this.

 

[Journey]

Well I was thinking of upgrading after Volta (I bought a 970 and 1070), so in terms of TFs for the 70 series (going by 70% increases each time)

Pascal - 6.5 TF
Volta - 11TF
After Volta - 18TF that's gotta be enough for 4k60 at that point.

Just in time for 8k displays to hit the mainstream.

 

[Dehnus]

Just in time for 8k displays to hit the mainstream.

And just in time for Ethereum miners to figure out how to use Tensor cores to mine even more efficient . AND BAM They're gone.

.

But on a good note, you'll be able to get a lot of 1070's and 1060's by then, used for a few years at maximum ... and needing a bios update to work again.. as they flood the market for the same price as a new one 2 years ago .

 

[icecold1983]

JHH made it pretty clear in the conference call that just took place no consumer volta this year

 

[AmyS]

NVIDIA Gives Xavier Status Update & Announces TensorRT 3 at GTC China 2017 Keynote

Earlier today at a keynote presentation for their GPU Technology Conference (GTC) China 2017, NVIDIA's CEO Jen-Hsun Huang disclosed a few updated details of the upcoming Xavier ARM SoC. Xavier, as you may or may not recall with NVIDIA current codename bingo, is the company's next-generation SoC.

Essentially the successor to the Tegra family, Xavier is planned to serve several markets. Chief among these is of course automotive, where NVIDIA has seen increasing success as of late.

In today's GTC China keynote,Jen-Hsun revealed updated sampling dates for the new SoC, stating that Xavier would begin in Q1 2018 for select early development partners and Q3/Q4 2018 for the second wave of partners. This timeline actually represents a delay from the originally announced Q4 2017 sampling schedule, and in turn suggests that volume shipments are likely to be in 2019 rather than 2018.

https://www.anandtech.com/show/1187...orrt-3-announcement-at-gtc-china-2017-keynote

 

[dgrdsv]

https://www.anandtech.com/show/1187...orrt-3-announcement-at-gtc-china-2017-keynote

It's pretty much their usual schedule for "mobile" SoCs - they are announced at CES in January with some limited availability promised for the end of the same year and a full rollout during the next year. Was like this since Tegra 4 I think. Xavier does seem to be coming a bit later than expected though, and it's hard to tell what could be the reason.

 

[Renekton]

A new mobile Soc at 12nm+ or 10nm sounds very exciting.

 

[Candescence]

I can't wait to see what kind of beast of a chip the Switch 2 is gonna have.

 

[flipswitch]

So Nintendo is officially going with the hybrid approach going forward? I hope they do. Tech is always getting better plus it’s so convenient having a console and handheld in the one package.

 

[Renekton]

So Nintendo is officially going with the hybrid approach going forward? I hope they do. Tech is always getting better plus it’s so convenient having a console and handheld in the one package.

Nintendo doesn't seem to have a fixed design philosophy other than blue ocean. They could go in a wildly different direction next gen.

 

[Timu]

Nintendo doesn't seem to have a fixed design philosophy other than blue ocean. They could go in a wildly different direction next gen.

Knowing Nintendo, that could definitely happen.

 

[dgrdsv]

A new mobile Soc at 12nm+ or 10nm sounds very exciting.

Xavier is not really a mobile SoC, hence why I've put "mobile" in quotes. It's going to be interesting to see if Nintendo will ask NV to actually build something specifically for them (and if NV will agree) as otherwise it doesn't look like NV will have anything for Switch 2 "off the shelf".

 

[Renekton]

Nvidia CEO says Moore’s Law is dead and GPUs will replace CPUs

http://www.pcgamer.com/nvidia-ceo-says-moores-law-is-dead-and-gpus-will-replace-cpus/

 

[Everlya_Apocryha]

Nvidia CEO says Moore’s Law is dead and GPUs will replace CPUs

http://www.pcgamer.com/nvidia-ceo-says-moores-law-is-dead-and-gpus-will-replace-cpus/

it depends on which market...i am not seeing x86 CPU going away anytime soon on the desktop front...

For the Application Servers though....that may very be in the next decade.

 

[dgrdsv]

NVIDIA Announces Drive PX Pegasus at GTC Europe 2017: Level 5 Self-Driving Hardware, Feat. Post-Volta GPUs

I wonder if these "post-Volta GPUs" are what we'll get in GeForces next year or something even beyond that already. Doesn't make much sense to build an AI Drive PX module on a gaming architecture so this may in fact be something like Volta@7nm.

 

[AmyS]

No doubt Nvidia is concurrently working on at least two more GPU architectures beyond Volta given the company's huge size / and massive R&D budget. I mean AMD has Navi for 7nm and beyond that "Next Gen" for 7nm+ so Nvidia is doing at least something similar if not more -- The real question is, what will be the architecture names?

Fermi
Kepler
Maxwell
Pascal
Volta
?????
?????

 

[dgrdsv]

No doubt Nvidia is concurrently working on at least two more GPU architectures beyond Volta given the company's huge size / and massive R&D budget. I mean AMD has Navi for 7nm and beyond that "Next Gen" for 7nm+ so Nvidia is doing at least something similar if not more -- The real question is, what will be the architecture names?

Fermi
Kepler
Maxwell
Pascal
Volta
?????
?????

Well, if Drive PX2 Pegasus is based on some AI/DL oriented architecture that will be different from the one which they'll use for GeForces next year and considering that Pegasus is supposed to become "available" (to selected key partners yada yada) in 1Q18 then I'd say that we're looking at two architectures being developed in parallel for different markets really. And this in turn likely means that they are developing more than two architectures right now - likely closer to four or even six.

I dunno why they don't disclose the names for what comes after Volta though (they are certainly given already). Afraid that no one will buy Pascals after that? Seems like a stupid fear to have.

One thing to note though is that these two post-Volta GPUs on Pegasus board seem to be using 192 or 384 bit GDDRx memory:

This kinda makes it possible that we're actually looking at what will be used for GeForce cards here.

 

[Polygonal_Sprite]

When are these expected to hit retail in 2018?

 

[dgrdsv]

When are these expected to hit retail in 2018?

Not sure what you mean by "these" - Drive PX Pegasus which was unveiled today? It will never hit retail as it's a unit made specifically for automotive / taxi companies and it will be sold via B2B channels only probably.

As for when this will happen - 1H18 is for selected partners and 2H18 for large scale availability.

Next GeForce lineup is expected to launch in Spring 2018 but it's anyone's guess on how these two are related.

 

[Polygonal_Sprite]

Not sure what you mean by "these" - Drive PX Pegasus which was unveiled today? It will never hit retail as it's a unit made specifically for automotive / taxi companies and it will be sold via B2B channels only probably.

As for when this will happen - 1H18 is for selected partners and 2H18 for large scale availability.

Next GeForce lineup is expected to launch in Spring 2018 but it's anyone's guess on how these two are related.

Sorry I should have specified. I meant the next gen GeForce GPU’s.

A mate is building a gaming PC for 4k/60fps and I was wondering if he should hold off for a big leap in power if it’s going to be available in the early part of next year. Monies no object to him so I guess he should just wait until next Spring.

 

[Deleted member 17706]

Given that AMD is not bringing the competition *at all* I could see Nvidia milking the 10 series for as long as possible.

I imagine they will do a similar launch structure for the 11 series, too, with multiple "flagship" card releases (1180 at a jacked up price at first, 11 Series Titan card, 1180 Ti) stretched out over a year or so.

I imagine the sweet spot for 4K/60 FPS will be the 1180 Ti, but it's possible we won't even see it until late 2018 or sometime in 2019.

 

[dgrdsv]

Sorry I should have specified. I meant the next gen GeForce GPU's.

A mate is building a gaming PC for 4k/60fps and I was wondering if he should hold off for a big leap in power if it's going to be available in the early part of next year. Monies no object to him so I guess he should just wait until next Spring.

Current expectation is Spring, around GDC in March or GTC in May.

Edit: Actually, scratch that - GTC 2018 will be held in March as well, back to back with GDC. So March is very likely then.

 

[Spurs]

Time to throw my 1080 in the bin?

 

[Nikokuno]

Time to throw my 1080 in the bin?

*Waiting like It in the trash bin*