Samsung to Optical-Shrink NVIDIA "Pascal" to 14 nm

[FR@NK]

Krhm, if you mean that picture from this article, that's the tesla P100.

Ah thanks.

 

[mcraygsx]

Based on that picture it looks like a mobile chip.

This is sold news for enthusiast who have been waiting for VOLTA hopefully next year.

 

[HisDivineOrder]

Sounds like the 1160, 1170, 1180, 1180 Ti, and Titan X Pascal X got their upgrade.

 

[ensabrenoir]

Sounds like the 1160, 1170, 1180, 1180 Ti, and Titan X Pascal X got their upgrade.

....probably to directly coincide when Amd releases their "big guns." Have a feeling the entire 10 series will be short lived.

 

[Prima.Vera]

Now that die shrinks are getting ever harder, I'm not surprised that they are taking this step.

If I remember correctly, at around 3-7nm you reach the size of single atoms, so we're close to the limit now.

No.
An atom size is between 0.2 - 0.3nm, which is looooong way from 3nm even.

 

[ArdWar]

No.
An atom size is between 0.2 - 0.3nm, which is looooong way from 3nm even.

Yeah, silicon atom size is aporox 0.1nm, but the distance between silicon atoms in silicon crystal is approx 0.2nm

Controlling where an electron should or shouldn't go when you only have 10 atoms between them is difficult nonetheless.

 

[TheGuruStud]

Yeah, silicon atom size is aporox 0.1nm, but the distance between silicon atoms in silicon crystal is approx 0.2nm

Controlling where an electron should or shouldn't go when you only have 10 atoms between them is difficult nonetheless.

You can pretty much forget (actual) single digit nm nodes. It'll leak like a sieve.

 

[FordGT90Concept]

sounds more like a trial&error run

I concur. TSMC's 16nm is probably better than Samsung's 14nm like it has proven to be better than Global Foundries' 14nm.

You can pretty much forget (actual) single digit nm nodes. It'll leak like a sieve.

It'll probably be on silicon-germanium.

 

[Siman00]

Makes me wonder if TSMC is having yield problems on 16nm like they did when they went from 40nm to 28nm. That could in part explain the shortages but I think most of the shortage of 1070 and 1080 is due to their extremely good performance and a lot of people want them. Maybe Nvidia is finally fed up with TSMC.

Nvidia has yeild problems in general their die sizes are massive...

 

[efikkan]

There have been talks of producing Tegra on Samsung "14nm" FinFET for a long time.

But Samsung's "14nm" FinFET isn't really significantly smaller than TSMS' "16nm" FinFET, and the 16nm HP node is working quite well, so if any chips are moved at all it's because of prize, not because of "shrinking".

 

[efikkan]

Nvidia has yeild problems in general their die sizes are massive...

No, contrary to initial 28nm and 40nm batches, 16nm have no yield issues. The yields of GP100 is excellent, but all the production capacity is already reserved.

 

[qubit]

No.
An atom size is between 0.2 - 0.3nm, which is looooong way from 3nm even.

That's great, we've got some way to go before that hard limit then.

 

[DeathtoGnomes]

So whats next after the hard cap?

tiny crystal of beryllium ions??

 

[FordGT90Concept]

Quantum computing

~500 qubits translates to about 500 GFLOPS in terms of normal processor loads but upwards of 5 PFLOPS when dealing with problems where superposition is useful (true and false simultaneously):
http://newatlas.com/d-wave-quantum-computer-supercomputer-ranking/27476/

Every two years, they've been doubling the number of qubits and the growth in performance is damn near exponential. In solving problems conducive to its design, it will be passing up traditional supercomputers if it hasn't already. NASA, Google, etc. are buying them up as fast as they can make them.

 

[Keullo-e]

GTX 1080+ anyone?

Spoiler

throwback to 9800 GTX and 9800GTX+

And first there was 8800GTS 512 and last GTS250.

 

[dont whant to set it"']

with driver and driver specific uhm specific software improvements at driveer level might yeld a pitential new lineup relegating curent one to mainstreme-high tendencie depending what review/s I believe.

 

[cdawall]

And first there was 8800GTS 512 and last GTS250.

8800gt, 9800gt and gts 240

 

[Alexander Kaiser]

So, is Pascal the new G92? Since AMD isn't providing strong enough competition, they'll just refresh and rebrand the crap out of it like the 8800gt -> 8800gts 512MB -> 9800gt -> 9800gtx -> 9800gtx+ -> GTS 250?

 

[cdawall]

So, is Pascal the new G92? Since AMD isn't providing strong enough competition, they'll just refresh and rebrand the crap out of it like the 8800gt -> 8800gts 512MB -> 9800gt -> 9800gtx -> 9800gtx+ -> GTS 250?

8800GTS 512 was not a rebranded 8800GT...few more shaders in there.

8800GT->9800GT->GTS240

8800GTS 512->9800GTX->9800GTX+->GTS 250

 

[KevinCobley]

It's pretty clear what will happen with 1080, 1070 and 1060, an optical shrink to 14nm with some tweaks like widening the memory bus to 512 and raising base clock to 1800mhz and boost clock to 2000mhz, probably with 8Gb and 16GB HBM memory versions, they will release as 1180, 1170 and 1160's probably early February 2017 with around 15-20% improvement. The Volta release will be scheduled to Christmas 2017.

 

[jabbadap]

It's pretty clear what will happen with 1080, 1070 and 1060, an optical shrink to 14nm with some tweaks like widening the memory bus to 512 and raising base clock to 1800mhz and boost clock to 2000mhz, probably with 8Gb and 16GB HBM memory versions, they will release as 1180, 1170 and 1160's probably early February 2017 with around 15-20% improvement. The Volta release will be scheduled to Christmas 2017.

Uhm how about no.
1. Are you really suggesting samsungs 14nm manufacturing node is up-to task to deliver such high freqs. Or even higher than tsmc.
2. Widening memory bus to 512bit and then use hbm, what?
3. Reaching Feb2017 chips should have been taped out already. And it would have leaked somewhere.

If they are forced to update something on that time(feb2017), it will be based on current gp104 chips with faster gddr5x and higher core clocks and gp102 parts with different configurations other than anything else.