Pascal: What do we know? Discussion, Latest News & Updates: 1000M Series GPU's

No they don't, lol. The rumors used to...not anymore.

There's a possibility for the 390x being different, as in, like how the Titan X differs from the rest (GM200).

The entire 300 series being 14nm by June? Unlikely. New processes are expected in 2016. Nvidia also has plans for a Hybrid Titan X, and at least one more GM200 card before then.

 

No I don't mean 14nm, but I thought the rumours all claim that it's going to be a new version of GCN?

 

Well, yeah, that's possible and likely, as I've said a dozen times: to respond to NVIDIA Maxwell. I thought everyone was expecting new architecture and whatnot in June, lol.

Still, I wouldn't expect magic from AMD. They most often disappoint, and very rarely surprise.

 

Most likely its just a refresh of the current GCN 1.x architecture with some new stuff. Not an entirely new architecture.
AMD`s TDP gain for 300 series seems to be mostly from a different manufacture process 28nm SHP from GloFo or 20nm since architecture looks to be the same. Kyle Bennet from HardOCP said yesterday that "400 series will be the real new architecture from AMD"

If they havent already started mass production of 16nm from TSMC they will very soon. We will absolutely see it in 2016 in products from Nvidia and co.

 

All I care is if Pascal will work with my current MXM. AMD can burn and die.

 

Yes, I'm right there with you!

Man, I hope I'm right in believing Pascal will be the last official MXM 3.0b GPU. This means at least two years of potential upgrades for those of us with upgradeable units... Why? Because Volta is 2H 2018. That will most definitely have a new socket type considering 3.0b will have reached its limitations by then.

By the time Volta hits, I'll definitely need a new CPU. If Pascal is expected to be 200% faster than Maxwell, it's safe to say Volta may very well be at least 100% faster than Pascal. But by that time, I'd expect 1080p to be the new 720p, 1440p/1600p to be the new standard, and 4k to be the common enthusiast grade resolution.

 

I don't know. MXM3.0b should be able to accommodate even beyond Pascal unless bus width increases beyond 256-bit or speeds triple. The current spec is basically 16x PCIe 3.0 as I understand it, so it has to saturate a 16GB/sec bus before it becomes a bottleneck, and this is sufficient for desktop CPU's with much higher throughput than laptops.

 

Could be. No way of telling, though. I was just assuming Volta would be a lot better. It could just be 20% better, or it could be 200% better.

A good point was made in another thread about DX12. That will be very nice to have for everyone with 900 series or later graphics.

 

By 2018, market wise I'm willing to bet that 4K will be the new standard.

 

If the trend towards thin and light doesn't change, by 2018 there will be nothing but BGA left.

 

Mobile GPU's aren't that powerful. In desktop gaming, sure, I can see that.

I think there will be a lot of 4k available, but it won't be a standard for gaming laptops. 1440p/1600p will replace 1080p most likely. Pascal will make 4k just playable at "acceptable" frame rates. Volta may be the beginning of a 4k gaming at 60+ FPS. Of course, I am referring to the highest possible settings (inc. MSAA).

 

Sorry, I didn't mean to say that by 2018 the GPU would be powerful enough to handle 4K gaming at highest possible settings, mainly that 4K panels cost wise would start would make sense to make the jump with Intel & AMD integrated GPU's will be capable of handling windows tasks. Hopefully Win10 has addressed the scaling issue with going from 4k to 1080P if you need to crank everything up to 11, but handle 4K at 8... maybe 9. Right now we're at the point where manufacturers are bordering on the 1080P/4K where the 1080P screen is cheaper and makes the most sense to use, but the 4K screen is optional and you pay for that premium (it's like $300-$400-ish now?). A lot can happen in 3 years.

 

A lot can happen in 3 years, including a lot of rebrands. I expect a delay as well. We all know that's likely.

Pascal should actually be a bigger jump than Volta - the whole "tic-tock" process and everything.

 

Nvidia doesn't rigidly follow a "tick-tock" process like Intel does. Kepler and Maxwell were on the same 28nm process node and Maxwell isn't a reworking of Kepler but an entirely new architecture. Pascal is expected to be an entirely new architecture as well.

nVidia rebrands their architectures to release "new" graphics cards by enabling cores that were already there and/or increasing clock speeds. It's like tick-tock-tock-tock-tock-tock-tock and that still isn't right because there is no die shrink on the rebrands.

 

Not suggesting NVIDIA does. The market follows Intel, though. If Intel doesn't release anything new, then it's likely other processor companies, including AMD and NVIDIA, will follow. Just how the cookie crumbles.

 

I have always been Intel and Nvidia guy, but it makes me sad that AMD and ATI are struggling. The lack of competition speeds things down I feel. And I wish there where more than 2 CPU and GPU makers.

 

I agree. AMD has fallen behind. They're only just now responding to NVIDIA Maxwell this June at Computex.

 

NVIDIA is planning to have Samsung manufacture their Pascal GPU's, beginning production sometime this summer.

http://wccftech.com/samsung-rumored-manufacture-14nm-gpus-nvidia-finfet-3d-transistors-technology/

Looks like they're sticking to the "early 2016" release plan. I'm pretty damn excited to see what this generation performs like. The desktop cards should be on a 16nm process with mobile taking the 14nm process.

 

If by mobile you mean laptop, then no, it will be on the same process as desktop. The LP process will be for Tegra SoC. Maxwell 2 was the same deal--desktop/notebook on 28nm and Tegra X1 on 20nm.

 

The "early 2016" 16 nm Pascal GPU will likely be based on a small(ish) die, and probably not particularly interesting to anyone running a desktop 680 or above.

 

Thread cleaned. Come on, let us all be mature adults here.

 

When you say cleaned, do you mean you hid the bodies? JK

Can't wait to see what next gen AMD and NVIDIA chips can do.

 

FTFY

 

Added the conference video to the main post.

 

Now that Clevo is introducing their refreshed models next month, it looks like the confirmed specs include MXM modules. Therefore, it's fairly safe to assume Pascal will be MXM. We shouldn't have much to worry about in that regard. It looks like we have a good while of upgradeability ahead!

2018 is Volta, so we've got at least 2 years of upgrades, probably more if there are delays.

 

Why would you say that? Chances are Pascal will use MXM, but just because this refresh contains it doesn't mean Pascal will. We're still over a year away from Pascal.

 

It's highly unlikely Clevo will do two refreshes in 12 months time - that's not profitable. That's pretty much why.

Manufacturers have inside information that consumers do not in order to prep for future product releases.

 

Of course, but it's happened before.

 

Soooo... would you get a MXM3.0b notebook if you were in the market or would you not recommend it, wink-wink, nudge-nudge.

 

I have no idea where it's going at this point in time to be honest. I'd go with MXM 3.0b and I'm looking to do just that myself with this fall refresh more than likely of the Clevo P770ZM refresh. I was just saying that there's no certain correlation between refresh getting MXM and Pascal being MXM 3.0b. But in any case he's probably right that MXM 3.0b is here to stay for a while.

 

MXM will need to go through a heavy revision to support Pascal due to HBM.
Current MXM cards have a graphic chip with GDDR5 laid out around it.
HBM however you have the HBM on the same silicon as the GPU package. Basically one big die with the graphic chip and HBM under a big heatsink.
Power wise cards with stacked memory will go hand in hand due to less power from HBM vs GDDR5, meaning more power envelope for the GPU itself.

MXM boards can do 4096/8096bit just fine actually. Current MXM cards are restricted to 256bit due to the size requirements for the GDDR5 chips and the routing to the memory controller. You have 32/64bit memory controllers all adding up to 256bit and each memory controller connected to one or several GDDR5.
With HBM you have a vertical stack, 1GB giving 1024bit and combining 4 of these stacks you get 4096bit. HBM2 will increase that to 2/4GB stacks and up to 8096bit. The reason why I say MXM would support HBM is because if they can get the GPU package in a MXM board, the routing from the graphic chip to the HBM stacks are all done within the silicon. The size requirement have gone down incredibly from current MXM design. If they can do current cards, they can do HBM/Pascal too in MXM.

The potential problem I see is that with Pascal we will see the trend with slimmer notebooks continue/escalate. We have all soldered CPUs starting with Skylake.
A) HBM will decrease power requirements = slimmer notebooks due to TDP go down
B) Graphic cards will go down in size due to stacked DRAM instead of GDDR5 laying all around the MXM card = easier to make room for and manufacture on a motherboard.
C) MXM will need to do a brand new specification for HBM due stuff mentioned above. There may not be enough movement in the industry to make that happen.

 

Well it sounds like all the reasons you gave were because it's getting smaller and the MXM PCB would be large for the HBM design. If that is the case then maybe they'll be able to do SLI on a single MXM card. One can dream.

 

Those are the exact same deductions you were peddling with the AW 14/17/18... and look what happened there

While I think Pascal will be MXM on a hunch. No one here knows anything yet.

 

No it's not. With that I was saying they'd be idiots to make a 15" notebook to replace the 18". It's different.

Technically, I wasn't wrong. They're idiots. We all know that now. Clevo isn't as stupid as Dell.

 

No, I meant you were touting a AW14/17/18 refresh (r2) was a sure thing since they were only 12 month old designs at that time.
Its fine though, many thought that.
Case in point, no one here can say want's going to happen yet with any conviction

 

That would have been the logical lineup. They took the illogical route. Nope, can't foresee irrationality.

I bet on odds, and odds are Clevo is smarter - the Taiwanese* are smart.

 

Uh Clevo is Taiwanese...

 

Really? Well, that's not the point.

 

Sure it isn't

 

If the Japanese are smart.. I wonder what that makes the Taiwanese then..

 

Are you that guy? Someone will be telling you a story and you ruin it because a detail is incorrect? I can see you doing that. "The sky isn't blue. It's light blue."

And it isn't the point. They don't just ditch profitable products and invest in some long shot product, hoping it'll "all be okay." There's no way they do that. The culture of business in Asia is completely different than it is here in America. Totally different.

@TBoneSan,

Doesn't matter when they're both smart.

 

http://wccftech.com/nvidia-ceo-talks-pascal-pk100-pk104-gpus-produced-finfets-hbm/

"Pascal GPUs are slated to be pretty damn amazing. They will feature 4X the mixed precision performance, 2X the performance per watt, 2.7X memory capacity & 3X the bandwidth of Maxwell."

"The PK104 will be the GM204 successor while the PK100 will be the flagship Pascal GPU from Nvidia. Both will be manufactured on either the 14nm or 16nm node using FinFETs."

"Pascal could very well be poised to offer a lineup of one of the first 4K capable single-gpu cards at the 4k @ 60+ fps standard."

 

New updates last month: Forget HBM... HBM2 coming.

Spoiler: NEWS AND UPDATES! LAUNCH EXPECTED Q1 2016.

"Next-generation graphics processing units from Nvidia will support second-generation stacked high-bandwidth memory (HBM2). The HBM2 will let Nvidia and its partners build graphics boards with 16GB – 32GB of onboard memory and 820GB/s – 1.2TB/s of bandwidth." - Source

"The 3D memory that Nvidia refers to above is believed to be the latest HBM2 type, it boasts of a maximum memory bandwidth of around 1.2TB/sec. We have heard that AMD plans to move onto using HBM2 in its graphics cards next year sometime but perhaps Nvidia will get there first as the Pascal GPU could be launched as early as Q1 2016." - Source

According to a couple articles online, production has already begun for Pascal and they may have the first GPU samples launched as early as August of this year. However, official release for the 'gaming' segment is not expected until early 2016.

Confirmed: Up to 32GB HBM2 memory supported. Gamer variants probably only 16 GB.

http://wccftech.com/nvidia-pascal-gpu-gtc-2015/

 

Q1 2016? Earlier than I had predicted. Maybe that's why NVIDIA aren't giving us a 980M replacement but rather a Titan-class mobile GPU (according to the latest rumours).

If it turns out that NVIDIA only have the power-guzzling 990M (or whatever it'll be called) and nothing to replace the 980M in existing 15.6" gaming notebooks then I'll just wait for Pascal instead.

 

Definitely going to be able to handle 4k next year.

I think I'm most excited about NVLink. Scaling is going to be amazing with Pascal.

 

GP100 taping out is old news, and I can almost guarantee the first Pascal chips will be reserved exclusively for the compute market and become Tesla cards, with consumer cards coming much later. Why?

Because Maxwell has terribad DP (1/32, even worse than non-Titan Kepler's 1/24), and the Tesla line is still using Kepler. The most powerful Tesla K80 uses two GK210 chips, and has a theoretical 2.91 teraflops of peak performance assuming perfect scaling and full boost.

Here's the kicker: Intel will be launching Knights Landing in 2H 2015. This is a monstrous chip built on Intel's most advanced 14nm process, and will offer ~6 teraflops of SP performance and >3 teraflops of DP performance. So almost K80-like performance but with only a single die so no potential scaling issues to deal with, and will likely offer (much) better thermals. This thing is a compute monster and will present strong and heavy competition for nVidia. In fact I'd even go far as to say if Intel really wanted to corner nVidia, they could sell these at cost, and nVidia would be in for some real pain.

Given Maxwell can't DP to save itself (Titan X only has ~13% the DP performance of the original Titan), and Kepler based Tesla cards are going to get long in the tooth, it's not too hard to see why nVidia would want to push out big Pascal with proper DP capabilities ASAP.

Btw completely neutered DP, as well as the ridiculous dynamic throttling is why Maxwell is so "efficient".

P.S. HBM2 isn't even supposed to enter volume production until Q2 2016, so no idea where they're pulling that Q1 2016 number from.



Also if you thought AMD was bad with release timing, you should read up on TSMC's yield issues on both the 40nm and 28nm nodes. Their "promises" and "timelines" should be taken with a huge grain of salt.

 

Why so much memory, though? Is there one day to be 128GB VRAM cards? Doesn't seem too far off.

Seems completely unnecessary to me. Wouldn't making memory "faster" (or more efficient) be more effective? As it becomes more efficient, you shouldn't need as much of it, correct? Or is that not the case. I'm not sure I understand this.

 

8GB for 4K resolution should be more than sufficient. But then again, I guess people will want triple monitor 4K gaming next, and the cat/mouse game continues.

 

Was this in response to my post? If yes then for compute you need the memory to contain ginormous amounts of data sets.

 

Nah, not in response to anything in particular. Just came to mind after reading the articles and your post. You'd think by now memory would be much faster or efficient, thus requiring less. In other words, less is more (as technology advances). But I suppose it doesn't work that way, at least not yet.

What setup needs 32GB's of HBM2 memory? Holy crap, lol. I've never used more than (roughly) 3.5GB's. Greater amounts of memory usually brings with it greater power consumption and greater heat. Seems redundant to me, especially as components are becoming smaller, thus generating less heat. They're essentially countering the "savings" by adding more?

One step forward, two steps back.