Meet Pascal, Nvidia's next-generation GPU that could render PCIe obsolete

Well this is very interesting no doubt, but let's see if these advantages actually translate into a massive real world performance jump rather than another incremental boost.


1000 times better bandwidth and 12x faster than PCIe with direct CUP-GPU memory access.... uh, how would this NOT translate in a huge increase in performance while becoming much more energy efficient as well. I know I would love it if my video card didn't have to be so loud when cranking out game graphics... the cooling apparatus is so large now.
 
Looks similar to current embedded solutions available from AMD, a socketed embed solution gpu + ram.
http://www.amd.com/us/products/embedded/graphics-processors/Pages/embedded-display-graphics.aspx
http://www.amd.com/us/Documents/AMD-Radeon-E6760-Discrete-GPU-product-brief.pdf

The only difference and most innovative thing in my opinion is the ram.
The rest is old news, quite hyped by the title of this article by the way. Except this is an NVidia paid article and I didn't noticed it.
So this is the future... gpu on proprietary sockets? quite lame....
 
I'm sure that it's just pure coincidence (cough, cough) but as per their settlement with Nvidia Intel will no longer have to support PCI-Express in 2016. Obviously PCIe isn't going anywhere but Intel will be able to make life very difficult for Nvidia and with the advent of the APU they'll have a good reason to do so. NVLink may be a way out for Nvidia, providing they're ready and able to cough up enough cash to OEMs to support their proprietary technology.
 
If nVidia wants to make a difference or be creative, how about finally introducing a combo of a video card + SSD on the same board?

Most of the time the PCI Express 3.0 is doing nothing anyway, and it would make a perfect balance as a boot-able solution for any system.

Will we ever live to see the day when this happens?
 
I have to say as impressive as this is it is also irrelevant at the moment due to the implications of implementing a new standard.

Right now, we still do not saturate the bandwidth of PCIE-2.0 Standard (Let alone 3.0 standard) which I doubt is going to change to much by the time this comes around (At least for the 3.0 standard). The other part of that is the fact that boards would have to implement it which means a new standard of boards, CPU manufacturers (AMD and Intel) would have to also designate support for the standard, and they would have to make this new design open sourced. I see this being beneficial in the future which is where this is listed right now, but nothing that should or probably will be out in 2 years.

@JC713
Stacked DRAM is essentially the idea that through the silicon you can stack up multiple DRAM die on top of eachother which will in turn provide a much higher density for the DRAM package. They will also put a hole through the center which will allow a center connection for all the stacked ram modules which will allow for a connection to all stacked chips in the lowest latency with the highest speeds possible.

In short, it would allow for higher bandwidth speeds of (Well NVidia has said) 1tb and beyond. You could say the technique (This is a poor example, take it more as just a generalizing with the way it works/reacts and not the actual methodology. In other words just how the speed is achieved and such) is similar to running raid on a SSD/HDD except the DRAM is basically latched onto eachother and connected together to make it even faster.
 
@JC713
Stacked DRAM is essentially the idea that through the silicon you can stack up multiple DRAM die on top of eachother which will in turn provide a much higher density for the DRAM package. They will also put a hole through the center which will allow a center connection for all the stacked ram modules which will allow for a connection to all stacked chips in the lowest latency with the highest speeds possible.

In short, it would allow for higher bandwidth speeds of (Well NVidia has said) 1tb and beyond. You could say the technique (This is a poor example, take it more as just a generalizing with the way it works/reacts and not the actual methodology. In other words just how the speed is achieved and such) is similar to running raid on a SSD/HDD except the DRAM is basically latched onto each other and connected together to make it even faster.
Ah ok. Why did it take so long to implement if it is that simple (in concept)? Was the manufacturing tech not there yet? Was it too expensive?

RAM modules stacked on top of one another. Is that simple enough? (Just yanking your chain). :)
Hahaha xD

If nVidia wants to make a difference or be creative, how about finally introducing a combo of a video card + SSD on the same board?
Wishful thinking my friend :D.
 
Ah ok. Why did it take so long to implement if it is that simple (in concept)? Was the manufacturing tech not there yet? Was it too expensive?

Well that has many reasons of course in part it is of course due to the rise of such technologies and the abilities brought on by changes to silicon and enhancements along the way.

The amount of ram has not been increasing in high levels mostly do to the fact that requirements for ram have not been skipping to such high levels so fast before. We have essentially gone from 480p being the norm to 1080p in a matter of 5 years essentially and already 1440p(Or 1600p) and the new 4K resolutions have already been pushed out the door bringing in a new level of requirements to the gaming world. The easiest solution has been to just add more higher capacity chips (Or just more DRAM chips) in general because the levels have not been moving up at any extreme rates. But now, we have effectively brought out needs for high end gaming cards to have a minimum of 3gb per GPU on a 1080p screen in many games and with requirements skyrocketing throughout the years. But we are also running into an area where continuously cramming more chips is just going to make power consumption higher and make card cooling and power delivery a much higher chore.

This improvement is going to allow more ram to be crammed in a smaller space while upping the bandwidth and keeping power consumption better in the long run which these changes won't happen for quite some time. It just was unnecessary for such a thing to be developed right now and was unneeded until they have realized they are getting close to a barrier with the current DRAM system that is going to need improvements to continue into the future.
 
Well that has many reasons of course in part it is of course due to the rise of such technologies and the abilities brought on by changes to silicon and enhancements along the way.

The amount of ram has not been increasing in high levels mostly do to the fact that requirements for ram have not been skipping to such high levels so fast before. We have essentially gone from 480p being the norm to 1080p in a matter of 5 years essentially and already 1440p(Or 1600p) and the new 4K resolutions have already been pushed out the door bringing in a new level of requirements to the gaming world. The easiest solution has been to just add more higher capacity chips (Or just more DRAM chips) in general because the levels have not been moving up at any extreme rates. But now, we have effectively brought out needs for high end gaming cards to have a minimum of 3gb per GPU on a 1080p screen in many games and with requirements skyrocketing throughout the years. But we are also running into an area where continuously cramming more chips is just going to make power consumption higher and make card cooling and power delivery a much higher chore.

This improvement is going to allow more ram to be crammed in a smaller space while upping the bandwidth and keeping power consumption better in the long run which these changes won't happen for quite some time. It just was unnecessary for such a thing to be developed right now and was unneeded until they have realized they are getting close to a barrier with the current DRAM system that is going to need improvements to continue into the future.
I bet we will have >4GB GPUs by default within the next few years because of 4K.
 
If NVidia licensed 8086 architecture from Intel like AMD does, then Intel and AMD would be in serious trouble.
Wouldn't happen. Intel doesn't do business that way.
AMD only has an x86 licence because IBM stipulated that a second source for 8088 processors was required for the IBM PC, both to ensure supply and quality. Intel were strong armed into accepting AMD as a second source supplier.
VIA only has an x86 licence because they bought Cyrix (minus MediaGX) from National Semiconductor who included their x86 license, who were originally obtained the x86 licence through a cross-license of IP with Intel .
Before NVidia gets any bigger Intel should buy it
But they likely wont. Intel are wedded to x86 - even their math co-processor (Xeon Phi) leverages it.
Intel could consider a buy out for Nvidia's in-place markets and IP (which is still found in Intel products and will likely be subject of a continued cross-licence agreement), but Nvidia seem to have been guarding against that eventuality by instituting a rigorous stock buy back program. Buying a cash-rich Nvidia won't be cheap, and it would be a certainty that Jen Hsun would demand a high profile position in the new hierarchy- two obstacles Intel would probably balk at.
Intel has tripled the on chip graphics of the 4th gen i3/5/7 chips and for most applications eliminates the need for a GPU from AMD or NVidia, which is a major threat to them. Can't see discrete graphics existing in 2016.
A few points to consider:
Intel has zero presence in workstation graphics
Intel is still a minor player in GPU math co-processors, and even with a process lead over Nvidia, can't get anywhere near them in efficiency.
Intel's iGP is gaining steady ground on discrete graphics, but they are still a long way from destroying the discrete market. Think an Intel HD graphics is ready for 4K gaming? multi-monitor gaming? post process menu options in games (HBAO/HDAO/SSAO, depth of field etc.)? tessellation ? driver stability and user options? and of course, how many people associate Intel with PC gaming ?
As iGP's increase in performance, Nvidia and AMD will keep pushing gaming hardware requirements to safeguard their own future products.
Ah ok. Why did it take so long to implement if it is that simple (in concept)? Was the manufacturing tech not there yet? Was it too expensive?
Yes and yes.
Stacking silicon means finding a way to either effectively heatsink the whole stack ( volume increases versus surface area), or reducing input power. Smaller process nodes means more effective transistor tech which is required as heat increases with transistor state switching speed. A brief overview >>here<<. I would also note that HBM isn't slated for Nvidia only.
Stacked DRAM (or any stacked silicon) requires connections between layers. TSV's (Through Silicon Via -Via standing for Vertical Interconnect Access), and they are only really feasible on small process nodes, as their mass has a direct relationship with input power demand (and waste heat output).
I bet we will have >4GB GPUs by default within the next few years because of 4K.
With double density memory IC's now hitting the market, 4GB will be the new 2GB...although I suspect that 4K won't drive the change but marketing.
More is better, right? Advertising says this is so, so it must be right.
 
Last edited:
Stacking silicon means finding a way to either effectively heatsink the whole stack ( volume increases versus surface area), or reducing input power. Smaller process nodes means more effective transistor tech which is required as heat increases with transistor state switching speed. A brief overview >>here<<. I would also note that HBM isn't slated for Nvidia only.
Stacked DRAM (or any stacked silicon) requires connections between layers. TSV's (Through Silicon Via -Via standing for Vertical Interconnect Access), and they are only really feasible on small process nodes, as their mass has a direct relationship with input power demand (and waste heat output).
The TSV diagrams are fascinating on that site. Amazing how we can stack little DRAM modules on top of each other just like blocks. I am very excited for this new tech. Now this is engineering at its best.

With double density memory IC's now hitting the market, 4GB will be the new 2GB...although I suspect that 4K won't drive the change but marketing.
More is better, right? Advertising says this is so, so it must be right.
Yeah advertising is a big part of it. In order to advertise 4K they must also market big
DRAM amounts.
 
The TSV diagrams are fascinating on that site. Amazing how we can stack little DRAM modules on top of each other just like blocks. I am very excited for this new tech. Now this is engineering at its best.


Yeah advertising is a big part of it. In order to advertise 4K they must also market big
DRAM amounts.

Well remember, in this world the person with more is the best right? in reality, 4k is really going to push the envelope of GPU's and ram so this was inevitable (Or something similar).
 
I'm just a simple guy....is anything in any home PC bandwidth restricted? PCI-E 3.0, Sata III or 6? Doesn't seem to be.
 
I'm just a simple guy....is anything in any home PC bandwidth restricted? PCI-E 3.0, Sata III or 6? Doesn't seem to be.
Not really- but that won't stop the hardware industry telling you that everything you own is obsolete.
The only real tangible benefit I can see for the average (?) PC user might be those who use SLI or CrossfireX. The PCI Express bus has limitations in latency, so something along the lines of NVLink has the promise of a unified memory pool between GPUs, where all GPUs contribute to the render of a frame rather than one GPU per frame as is most common with both SLI and Crossfire. The limitations of Alternate Frame Rendering (AFR) which usually manifest as stuttering, screen tearing, and the requirement of dedicated coding have plagued the tech since day one.
 
Back