Yeap. That's what Nvidia is trying to do from the day APUs became a reality. Create higher price points for GPUs. AMD with 290/290X pricing spoiled it in the begging, but the prices in US because of mining, saved Nvidia's plan to increase gpu prices, at least for now.
Just another Nvidia show. ...... Meh Talking about a product 2 years in the future, when they havent delivered the product they talked about in the previous show (Maxwell, Tegra5/K1) . This talk is seriously to woo back the investors.
Chances of that unlikely as Intel will have something to say about killing PCIe. What's the point when we are still stuck with PCIe 2.0 hardware and not even utilized PCIe 3.0
It's a chip to chip solution, not cpu to chip. Thus have different design goals, like not needing tons of pins/traces like PCIe. If they would take it to PCI-SIG nothing would stop any vendor from using it. Especially not those that do add-in boards like AMD.
NVlink looks like it bring PCIe outside the CPU chip back into the Chipset hence allowing HSA design in the Intel infrastructure. Though not as good as AMD, it might be enough to utilise the GPU power for its intended tasks. It would seem to be a replacement for the primary PCIe GPU port on the PC. This is not such a bad idea if the industry license it for free to everyone. We seemed to have come to a time where even PCie 3.0 is being bottlenecked so a replacement would be in order.
Man Maxwell is under some silly wraps right now. We can't seem to get a hint as to when it will be out or what process it will use other than "This year guys".
Very disappointing that there was no news on upcoming big-core Maxwell GPUs. Looks like we might have another 6+ months to wait on that. Are they waiting for 20nm to become available?
I do not think so, they just kept a low profile for Denver becuase there is no rush in the market, nor demand for 64bit Android chips as yet. Besides, they would need the optimizations to ensure competitiveness due to the time duration, so plenty of work effort in the time to do it properly. Maybe waiting for 20nm to get cranking in a big way as well.
So we know Volta still exists, but was pushed to 2016 and Pascal takes it spot in 2014. But what about Tegra Parker? Was that renamed, canned, or pushed back to come after Erista?
Everything appears to be still there, just in a different order. According to their earlier roadmaps:
(Year) GPU [CPU] (2014) Kepler [A15 with Cuda] (2015) maxwell with unified virtual memory [64bit with 16nm] (2016) Volta with stacked DRAM [unknown]
Since 64bit had been moved ahead of schedule, we now have a different roadmap:
(2014) Kepler [32bit A15, or 64bit custom] (2015) maxwell with dx12 [16nm Finfet erista] (2016) Pascal with stacked DRAM and unified virtual memory [unknown]
It looks like unified virtual memory was pushed back a year in order to bring 64bit a year earlier. Hope that helps.
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
Facebook
Twitter
RSS
50 Comments
Back to Article
blanarahul - Tuesday, March 25, 2014 - link
Umm... Is Brian at the HTC Keynote?extide - Tuesday, March 25, 2014 - link
Wow, so they basically renamed Volta to Pascal, and moved Unified memory from Maxwell to Pascal/Volta. Interesting...Ryan Smith - Tuesday, March 25, 2014 - link
Maxwell was unified virtual memory. Big emphasis on virtual.extide - Tuesday, March 25, 2014 - link
Not according to this: http://images.anandtech.com/doci/7894/GTC-2014-021... ..?blanarahul - Tuesday, March 25, 2014 - link
Yeah. Maxwell's dedicated to new DX12 features and power efficiency I guess.grahaman27 - Tuesday, March 25, 2014 - link
According to previous roadmaps, erista should be maxwell based and utilize a 16nm finfet process.iMacmatician - Tuesday, March 25, 2014 - link
I'd like to see 1 Yodaflop of performance.psychobriggsy - Tuesday, March 25, 2014 - link
12:48PM EDT - Wrapping up discussion of machine learningAnd much rejoicing was had.
Ken_g6 - Tuesday, March 25, 2014 - link
Ack! I gagged when I saw the $3000 price tag for a GPU. I guess top-end prices keep climbing. :(RealiBrad - Tuesday, March 25, 2014 - link
its not really a new top end card, its more of 2 markets merged into a single card. Its a workstation/gaming card.No doubt that some will buy them for epeen though.
kyuu - Tuesday, March 25, 2014 - link
Titan Z is for graphics professionals, not for gaming (though you could game on it, if you really wanted).nathanddrews - Tuesday, March 25, 2014 - link
It's Titan SLI for MITX. I'm sure the boutiques already have the configurations figured out.yannigr - Tuesday, March 25, 2014 - link
Yeap. That's what Nvidia is trying to do from the day APUs became a reality. Create higher price points for GPUs. AMD with 290/290X pricing spoiled it in the begging, but the prices in US because of mining, saved Nvidia's plan to increase gpu prices, at least for now.extide - Tuesday, March 25, 2014 - link
It's pretty much aimed at Quadro users, and it is actually a very good value/$ compared to other top end quadro cards.kD7 - Tuesday, March 25, 2014 - link
So when is the Titan Z giveaway? ;)eanazag - Tuesday, March 25, 2014 - link
Save the money on rendering a Honda with 19 Irays and just buy the Honda. You could get several Hondas for the cost.HighTech4US - Tuesday, March 25, 2014 - link
Huh?Honda uses the 19 Irays to help design the cars that they sell in the millions.
mayankleoboy1 - Tuesday, March 25, 2014 - link
Just another Nvidia show. ...... MehTalking about a product 2 years in the future, when they havent delivered the product they talked about in the previous show (Maxwell, Tegra5/K1) .
This talk is seriously to woo back the investors.
magnusmundus - Tuesday, March 25, 2014 - link
What's the benefit of the Titan Z over two Titan Black's in SLI? It costs $1000 more.caladbolg - Tuesday, March 25, 2014 - link
Removing the SLI link heavily improves performance between the two GK110s, plus double the available VRAMextide - Tuesday, March 25, 2014 - link
No, it's the same amount of RAMnathanddrews - Tuesday, March 25, 2014 - link
Sadly SLI/CF graphics don't share VRAM. 6GB and 6GB is not the same as 12GB. All data is mirrored to each GPU's VRAM.extide - Tuesday, March 25, 2014 - link
I assume you aren't replying to me.Wreckage - Tuesday, March 25, 2014 - link
Imagine if NVLINK replaces PCIe, but only on Intel chipsets. That would pretty much kill AMD over night.vision33r - Tuesday, March 25, 2014 - link
Chances of that unlikely as Intel will have something to say about killing PCIe. What's the point when we are still stuck with PCIe 2.0 hardware and not even utilized PCIe 3.0Penti - Tuesday, March 25, 2014 - link
It's a chip to chip solution, not cpu to chip. Thus have different design goals, like not needing tons of pins/traces like PCIe. If they would take it to PCI-SIG nothing would stop any vendor from using it. Especially not those that do add-in boards like AMD.extide - Tuesday, March 25, 2014 - link
They said it can be for CPU to GPU also.Penti - Wednesday, March 26, 2014 - link
Yeah but unlikely in the x86-64 market and not what their slides show. Maybe at IBM.Ryanman - Wednesday, March 26, 2014 - link
As vsion33r said, there's 0 chance of intel abandoning PCIe right now.fteoath64 - Thursday, March 27, 2014 - link
NVlink looks like it bring PCIe outside the CPU chip back into the Chipset hence allowing HSA design in the Intel infrastructure. Though not as good as AMD, it might be enough to utilise the GPU power for its intended tasks. It would seem to be a replacement for the primary PCIe GPU port on the PC. This is not such a bad idea if the industry license it for free to everyone.We seemed to have come to a time where even PCie 3.0 is being bottlenecked so a replacement would be in order.
o3smog - Tuesday, March 25, 2014 - link
Man Maxwell is under some silly wraps right now. We can't seem to get a hint as to when it will be out or what process it will use other than "This year guys".extide - Tuesday, March 25, 2014 - link
It's pretty obvious that Maxwell will use both 28nm and 20nm.grahaman27 - Tuesday, March 25, 2014 - link
You both mean the tegra k1 right?HighTech4US - Tuesday, March 25, 2014 - link
Quote: Tegra 4i (A9 + baseband) was famously pushed back last year to get Tegra 4 to market soonerYou have that backwards. The Tegra 4 was pushed back to bring the Tegra 4i (A9 + baseband) forward.
HighTech4US - Tuesday, March 25, 2014 - link
Quote: NVIDIA normally names mobile products after comicbook heroes. Not sure where "Jetson" fits inYou must have missed the multiple mentions of Robots, Robots, Robots
JDG1980 - Tuesday, March 25, 2014 - link
Very disappointing that there was no news on upcoming big-core Maxwell GPUs. Looks like we might have another 6+ months to wait on that. Are they waiting for 20nm to become available?haukionkannel - Tuesday, March 25, 2014 - link
At the end of this year... hopefully... They want to have something new during the next Christmas season.jasonelmore - Tuesday, March 25, 2014 - link
Keeps AMD Guessing. Why give them a advantage this early?rogueninja - Tuesday, March 25, 2014 - link
nVidia: Here is our upcoming $2999 GPU.nVidia: Also, Maxwell is coming in 2014 that will blow that GPU outta the water. Problem?
jwcalla - Tuesday, March 25, 2014 - link
I guess they gave up on Project Denver.extide - Tuesday, March 25, 2014 - link
No... Project Denver is in the dual core/64-bit version of Tegra K1...grahaman27 - Tuesday, March 25, 2014 - link
Some interesting tidbit of the keynote was that the Denver version will be 4x as efficient... That's huge and they just glossed over it.http://m.youtube.com/watch?v=q25WhVQkQTI
jwcalla - Wednesday, March 26, 2014 - link
So the result of Project Denver is just an ordinary ARM chip? They made it out like they were going to put an ARM CPU in a dGPU.fteoath64 - Thursday, March 27, 2014 - link
I do not think so, they just kept a low profile for Denver becuase there is no rush in the market, nor demand for 64bit Android chips as yet. Besides, they would need the optimizations to ensure competitiveness due to the time duration, so plenty of work effort in the time to do it properly. Maybe waiting for 20nm to get cranking in a big way as well.tviceman - Tuesday, March 25, 2014 - link
Hey Ryan and/or Anand -So we know Volta still exists, but was pushed to 2016 and Pascal takes it spot in 2014. But what about Tegra Parker? Was that renamed, canned, or pushed back to come after Erista?
grahaman27 - Wednesday, March 26, 2014 - link
Everything appears to be still there, just in a different order. According to their earlier roadmaps:(Year) GPU [CPU]
(2014) Kepler [A15 with Cuda]
(2015) maxwell with unified virtual memory [64bit with 16nm]
(2016) Volta with stacked DRAM [unknown]
Since 64bit had been moved ahead of schedule, we now have a different roadmap:
(2014) Kepler [32bit A15, or 64bit custom]
(2015) maxwell with dx12 [16nm Finfet erista]
(2016) Pascal with stacked DRAM and unified virtual memory [unknown]
It looks like unified virtual memory was pushed back a year in order to bring 64bit a year earlier.
Hope that helps.
tviceman - Wednesday, March 26, 2014 - link
Parker was specific to Tegra. It was supposed to merge Denver CPU cores with Maxwell GPU cores. But as I said, it was not on Tegra's updated roadmap.Penti - Wednesday, March 26, 2014 - link
Yeah but unlikely in the x86-64 market and not what their slides show. Maybe at IBM.Ashari - Wednesday, March 26, 2014 - link
"giving away free SHIELDs", I suppose we don't have to ask for sale numbers any more, I wonder how much the write-off will be.boe - Friday, March 28, 2014 - link
Blah Blah Blah - just release the 880 GTX PCIe card.