take it with a teaspoon of salt: Pascal up to 17 Billion Transistors, 32GB HBM2

[Notional]

Pascal taped out in April. Greenland taped out late June/early July given GloFo's own risk production timelines.

 

Well the story about Pascal broke early june and Lisa Su's official statement broke july, but we don't know when exactly these things happened. Either way that is the actual GPU die, not HBM2 or interposers or anything. We still don't know how Pascal will even connect to HBM.

[jos]

so are chips still 2d? they seem very flat. I dont know why manufactures havent gone vertical like a skyscraper. 

need to cool each transistor.. Localised heating is worse..youhave 3D one but with internal heat sink... proof of concept by IBM.. not practical.

[GoodBytes]

I can believe it. But that will be, with Nv Link, for the Tesla cards.

Companies interested in Tesla farm / Nvidia GRID for their R&D needs, have plenty of money. 1 million $ plus or minus on these card for extra RAM, is pretty much pocket change.

It is a bit harder to swallow for many of us.

If we look at 4K gaming, depending on the game, as the latest GPU with HBM1 shows, some games are memory bandwidth limited, others are memory quantity limited.

Assuming it does get HBM2 (meaning the 4GB limit won't be there), it would solve both of these problems. Nvidia can put more memory as it now costs less, and have high bandwidth. I expect AMD to soon release a HBM2 version of their card around the same time... and maybe OCed a bit or a tweak architecture model that can bring additional performance.

[Curufinwe_wins]

This is actually pretty expected. You have a single dimensional shrink of around 40% which means the area taken up for the same number of transistors at 16nm should be around 30-40% of that at 28nm.

Now obviously not everything gets that much smaller, but this mirriors pretty well what we see from Intel as well with die shrinks.

[patrickjp93]

Well the story about Pascal broke early june and Lisa Su's official statement broke july, but we don't know when exactly these things happened. Either way that is the actual GPU die, not HBM2 or interposers or anything. We still don't know how Pascal will even connect to HBM.

Via interposer. Nvidia's already produced renderings of the end product.

[patrickjp93]

I can believe it. But that will be, with Nv Link, for the Tesla cards.

Companies interested in Tesla farm / Nvidia GRID for their R&D needs, have plenty of money. 1 million $ plus or minus on these card for extra RAM, is pretty much pocket change.

It is a bit harder to swallow for many of us.

If we look at 4K gaming, depending on the game, as the latest GPU with HBM1 shows, some games are memory bandwidth limited, others are memory quantity limited.

Assuming it does get HBM2 (meaning the 4GB limit won't be there), it would solve both of these problems. Nvidia can put more memory as it now costs less, and have high bandwidth. I expect AMD to soon release a HBM2 version of their card around the same time... and maybe OCed a bit or a tweak architecture model that can bring additional performance.

GPU-based NVLink may actually be replacing the SLI fingers we've come to know with another cable system even on Quadro and GTX solutions.

[Notional]

Via interposer. Nvidia's already produced renderings of the end product.

 

Cool, do you have an image? I've only seen the weird looking image with no connectors anywhere. That does not shot a silicon interposer at least, but just a pcb. Have no idea how they managed that.

[ChrisxIxCross]

Nvidia Pascal GPU to feature 17 Billion Transistors, Almost twice the transistors of Fiji

 

 

In an exclusive report published by Fudzilla, the site reveals that NVIDIA’s next generation Pascal GPU will feature 17 billion transistors crammed inside its core. Currently, the flagship GM200 core found on the GeForce GTX Titan X comes with 8.0 Billion transistors while the competitor, the Radeon R9 Fury X has a total of 8.9 Billion transistors inside its Fiji GPU. The 17 Billion transistors on the Pascal GPU are twice the transistors found on the GM200 Maxwell and the Fiji XT GPU core which is literally insane. Pascal is meant to be NVIDIA’s next high performance, compute focused graphics architecture which will be found on all market segments that will include GeForce, Quadro and even Tesla. Based on TSMC’s 16nm process node, NVIDIA’s Pascal GPU will not only feature the best performance in graphics but also the most power efficient architecture ever made by a GPU manufacturer.

 

It was revealed a few days ago that NVIDIA’s Pascal GP100 chip has already been taped out on TSMC’s 16nm FinFET process, last month. This means that we can see a launch of these chips as early as Q2 2016. Given that the transistor count is correct, we can expect a incremental performance increase from Pascal across the range of graphics cards that will be introduced.

 

TSMC’s 16FF+ (FinFET Plus) technology can provide above 65 percent higher speed, around 2 times the density, or 70 percent less power than its 28HPM technology. Comparing with 20SoC technology, 16FF+ provides extra 40% higher speed and 60% power saving. By leveraging the experience of 20SoC technology, TSMC 16FF+ shares the same metal backend process in order to quickly improve yield and demonstrate process maturity for time-to-market value.

 

The 17 Billion transistors are an insane amount but what’s more insane is the amount of VRAM that is going to be featured on the new cards. With HBM2, NVIDIA gets the leverage to feature far more memory than what’s currently allocated on HBM1 cards (4 GB HBM on Fury X, Fury, Nano, Fury X2). With HBM2, NVIDIA gets access to more denser chips that will result in cards with 16 GB and up to 32 GB of HBM memory across a massive 4096bit memory interface which will dominate the next high-resolution 4K and 8K gaming panels.Although they may have to wait a little bit longer thanks to AMD’s priority access to HBM2 with SK Hynix, the makers of HBM. With 8Gb per DRAM die and 2 Gbps speed per pin, we get approximately 256 GB/s bandwidth per HBM2 stack. With four stacks in total, we will get 1 TB/s bandwidth on NVIDIA’s GP100 flagship Pascal which is twice compared to the 512 GB/s on AMD’s Fiji cards and three times that of the 980 Ti’s 334GB/s.

The Pascal GPU would also introduce NVLINK which is the next generation Unified Virtual Memory link with Gen 2.0 Cache coherency features and 5 – 12 times the bandwidth of a regular PCIe connection. This will solve many of the bandwidth issues that high performance GPUs currently face. One of the latest things we learned about NVLINK is that it will allow several GPUs to be connected in parallel, whether in SLI for gaming or for professional usage. Jen-Hsun specifically mentioned that instead of 4 cards, users will be able to use 8 GPUs in their PCs for gaming and professional purposes.

With Pascal GPU, NVIDIA will return to the HPC market with new Tesla products. Maxwell, although great in all regards was deprived of necessary FP64 hardware and focused only on FP32 performance. This meant that the chip was going to stay away from HPC markets while NVIDIA offered their year old Kepler based cards as the only Tesla based options. Pascal will not only improve FP64 performance but also feature mixed precision that allows NVIDIA cards to compute at 16-bit at double the accuracy of FP32. This means that the cards will enable three tiers of compute at FP16, FP32 and FP64. NVIDIA’s far future Volta GPU will further leverage the compute architecture as it is already planned to be part of the SUMMIT and Sierra super computers that feature over 150 PetaFlops of compute performance and launch in 2017 which indicates the launch of Volta just a year after Pascal for the HPC market.

   

 If these alleged rumors hold true Pascal is looking like its going to be a massive leap in performance compared to leap in performance. If the full fat GP100 chip is going to have 17 billion transistors then that is over double the amount compared to GM200 and over 3 times the amount compared to GM204. Also 32 GB VRAM!? I can't wait to see what kind of setup you're going to need to even come close using up that much frame buffer. Also Pascal will be utilizing TSMC's all new 16nm FinFet + technology allowing for 65% higher speed, 2 times the density and 70% less power than that of 28HPM which is Maxwells production process. So expect to see even lower power usage, ultra low temperatures and cards finally breaking the 2GHZ barrier on air. Personally I was on the fence with upgrading to two 980 Tis from my 980s but I think Im just going to sit and wait a bit longer!

What do you guys think? 

 

Source(s):

 

http://www.fudzilla.com/news/graphics/38304-nvidia-pascal-gpu-has-17-billion-transistors

http://wccftech.com/nvidia-pascal-gpu-17-billion-transistors-32-gb-hbm2-vram-arrives-in-2016/

[AlwaysFSX]

Sure it is.. I'll believe it when I see it because this is a VERY bold claim.

[Tim Drake]

*sigh* Just adding vram won't sell cards guys..

[Bob Jim]

RIP AMD.

[werto165]

These really don't seem like consumer cards... 

[AlwaysFSX]

These really don't seem like consumer cards... 

Yeah I would expect the GeForce variants to not have 32 gigs of VRAM, maybe 16.

[SurvivorNVL]

I'll just get one 980ti for now, and then when Pascal is out I'll get two 1080's or 1080ti's.  Whichever.

[werto165]

Yeah I would expect the GeForce variants to not have 32 gigs of VRAM, maybe 16.

Even 16GB is insane, I really don't see why you would need that much for gaming. Unless games start to be programmed different, have literally everything in VRAM... (I'm no programmer so I have no idea if that's actually possible and how it would benefit you). 

[Jerakl]

I was planning on waiting for Pascal anyways :^)

[DarkBlade2117]

Price is the biggest thing. Hopefully the 16gb variants aren't to much more expensive then the 980Ti.

[patrickjp93]

Cool, do you have an image? I've only seen the weird looking image with no connectors anywhere. That does not shot a silicon interposer at least, but just a pcb. Have no idea how they managed that.

http://wccftech.com/nvidia-pascal-gpu-17-billion-transistors-32-gb-hbm2-vram-arrives-in-2016/

 

It's a very different design from AMD's and is designed to provide additional smoothing of electrical current (much like Intel's FIVR, even similar in form if you look at the integrated .capacitors sitting on the interposer itself on the left and right sides).

[Notional]

http://wccftech.com/nvidia-pascal-gpu-17-billion-transistors-32-gb-hbm2-vram-arrives-in-2016/

 

It's a very different design from AMD's and is designed to provide additional smoothing of electrical current (much like Intel's FIVR, even similar in form if you look at the integrated .capacitors sitting on the interposer itself on the left and right sides).

 

But that's just the already know prototype pic of Pascal:

I mean that cannot be a production unit, as it doesn't seem to have any connectors to anything. Just seems like a proof of concept thing to test HBM. But again the HBM looks to be connected to a PCB, which sounds to be near impossible with all the connecting points (ifbga rolls). I wonder if that testing unit is actually functional or if it's just another "wood screw" card by NVidia. Either way the picture's older than Fiji pictures, so I don't think too much should be put into it really.

[The Benjamins]

you may want to add more options to the poll, as in already own a 980 Ti, and to go amd or fury. plan to get a lower end card like 980/970.

[patrickjp93]

But that's just the already know prototype pic of Pascal:

 

I mean that cannot be a production unit, as it doesn't seem to have any connectors to anything. Just seems like a proof of concept thing to test HBM. But again the HBM looks to be connected to a PCB, which sounds to be near impossible with all the connecting points (ifbga rolls). I wonder if that testing unit is actually functional or if it's just another "wood screw" card by NVidia. Either way the picture's older than Fiji pictures, so I don't think too much should be put into it really.

Correct, it's not a production unit in the flesh, but the connectors are secondary to all of that. That's about exactly what the interposer will look like. And are you blind? The 4 HBM stacks are on an independent square interposer which sits on the PCB.

[flibberdipper]

Hot damn.

[KirbyTech]

Even 16GB is insane, I really don't see why you would need that much for gaming. Unless games start to be programmed different, have literally everything in VRAM... (I'm no programmer so I have no idea if that's actually possible and how it would benefit you). 

Everything in memory is a bad idea in general but games would see a performance increase if the gpu itself was ale to handle rendering the frames fast enough to make a difference. 

[Sauron]

Nice, let's hope AMD can keep up.

[Notional]

Correct, it's not a production unit in the flesh, but the connectors are secondary to all of that. That's about exactly what the interposer will look like. And are you blind? The 4 HBM stacks are on an independent square interposer which sits on the PCB.

 

All I see is the NVidia die surrounded by the 4 HBM modules on a green PCB package. Sure, a PCB can be an interposer, but the reason AMD made one in silicon was because of the massive amounts of tiny BGA connectors would be extremely difficult to connect on a PCB with copper tracing. The black board only contains VRM and the GPU package.