NVIDIA GP107 ASIC Pictured, Die Smaller than a Dime

[btarunr]

Here are some of the first pictures of the incredibly tiny "GP107" silicon, on which NVIDIA plans to build the GeForce GTX 1050 Ti and GTX 1050. Pictures of a yet to be released graphics card based on this chip were leaked to the web, revealing a postage-stamp size GPU package, with its die smaller still (about 1 cm² die-area, if we're not mistaken).

It's been rumored that the GP107 could be built on the 14 nm process, by Samsung. The ASIC in the pictures are surrounded by four GDDR5 memory chips, sitting across its 128-bit wide memory interface. The GTX 1050 Ti features 4 GB of standard memory amount. Although the chip's TDP is rated to be 75W, AIC partners could give these cards 6-pin PCIe power connectors, just to be on the safe side.



View at TechPowerUp Main Site

 

[Fluffmeister]

Looking forward to see how these perform, but looks to be a pretty potent little chip in a small package.

 

[TheGuruStud]

Switching to a new fab doesn't make much sense unless TSMC is dangerously swamped, again.

 

[TheLostSwede]

Switching to a new fab doesn't make much sense unless TSMC is dangerously swamped, again.

This is unlikely to be a "switch" as TSMC and Samsung doesn't work together and have quite different manufacturing processes, so this would be a conscious decision on Nvidia's side. Maybe they want a second source for chips, since as you say TSMC can be "swamped" or it could mean that the yields aren't that great from TSMC and this is a first step towards Nvidia moving its chip production to Samsung. This is less of a gamble, as it's a comparably "simple" chip design and it allows Nvidia to tests Samsung's manufacturing abilities while not risking too much. If this had been the 1080 instead, this would've been a huge risk for Nvidia to entrust it with a new manufacturing partner, as it's a much more complex chip and it costs a lot more to tape it out and manufacture duds if things went pear shaped.

 

[Nokiron]

How cute!

 

[64K]

This is unlikely to be a "switch" as TSMC and Samsung doesn't work together and have quite different manufacturing processes, so this would be a conscious decision on Nvidia's side. Maybe they want a second source for chips, since as you say TSMC can be "swamped" or it could mean that the yields aren't that great from TSMC and this is a first step towards Nvidia moving its chip production to Samsung. This is less of a gamble, as it's a comparably "simple" chip design and it allows Nvidia to tests Samsung's manufacturing abilities while not risking too much. If this had been the 1080 instead, this would've been a huge risk for Nvidia to entrust it with a new manufacturing partner, as it's a much more complex chip and it costs a lot more to tape it out and manufacture duds if things went pear shaped.

Yields have been an issue with TSMC when going to a lower process node. We were stuck on 28nm GPUs for 4 years because of TSMC. Nvidia pays per wafer for the chips so when yields aren't good they have to use them for salvage chips if they aren't too defective but Nvidia is the one losing potential profit so they have an incentive to pursue a relationship with Samsung.

 

[Supercrit]

Potent single slot/low profile candidate.

 

[HalfAHertz]

Look at all that empty pcb space...

 

[TheLostSwede]

Yields have been an issue with TSMC when going to a lower process node. We were stuck on 28nm GPUs for 4 years because of TSMC. Nvidia pays per wafer for the chips so when yields aren't good they have to use them for salvage chips if they aren't too defective but Nvidia is the one losing potential profit so they have an incentive to pursue a relationship with Samsung.

Well, technically they did launch 20nm, it just wasn't suitable for GPU's as they managed to mess it up somehow. There's volume production of other things from TSMC at 20nm, but that's a different story.

As I said, I believe this is a move by Nvidia to either make sure they have a second source in the future, even if it means more work on their side, as TSMC and Samsung has quite different ways of doing things, but it worked for Apple so... Or it's a first move to Samsung from TSMC if the rumours prove to be true.

 

[GhostRyder]

Wow, too much blank space on that PCB for that chip. Make us some SFF and low profile variations please.

 

[aldo5]

smallest chip, empty pcb, performance so low as it would make console "gamers" proud.... and it will cost 150$... not 50$ as it should

 

[natr0n]

They should make it low profile this is stupidly over designed/sized.

 

[xorbe]

They should make it low profile this is stupidly over designed/sized.

You just know there's gonna be a 3 fan model!

 

[qubit]

I've gotta have me one of these Little Pascals.

 

[kaspar737]

smallest chip, empty pcb, performance so low as it would make console "gamers" proud.... and it will cost 150$... not 50$ as it should

With an overclock it's close to R9 290, consoles aren't anywhere near that, you know that right?

 

[chaosmassive]

we need half-height card with passive cooling on this cutie
and console graphic processor is not even relevant anymore at this point

 

[Blueberries]

Potent single slot/low profile candidate.

Yeah right! It doesn't matter if it has a 50W TDP, manufacturers will still put 3 fans on it. It's like single-slot cards are illegal now or something.

 

[BirdyNV]

With an overclock it's close to R9 290, consoles aren't anywhere near that, you know that right?

Say again? A 290 provides almost 5 TFLOPS of performance, the 1050 (allegedly) has just over 2 TFLOPS, unless you can overclock the HELL out of a 1050/ti, you won't be getting close to that at all.

 

[ppn]

overclock the HELL out of a 1050/ti,.

50% OC seems doable. Around the performance of GTX 780 @ 1Ghz. R 290 slightly faster.

 

[kaspar737]

Say again? A 290 provides almost 5 TFLOPS of performance, the 1050 (allegedly) has just over 2 TFLOPS, unless you can overclock the HELL out of a 1050/ti, you won't be getting close to that at all.

I'm talking about Firestrike Extreme numbers. TFLOPS are meaningless anyway, a 980Ti with 6TFLOPS is faster than Fury X with 8.6 TFLOPS.

 

[BirdyNV]

I'm talking about Firestrike Extreme numbers. TFLOPS are meaningless anyway, a 980Ti with 6TFLOPS is faster than Fury X with 8.6 TFLOPS.

Even then, its still 700 points off, a fairer comparison is a 280X. (+- 40/50 points).

 

[the54thvoid]

Wow, too much blank space on that PCB for that chip. Make us some SFF and low profile variations please.

That blank space is for the extra VRM, 8GB of memory and other pointless Pascal eye candy that doesn't help you overclock one iota.

 

[$ReaPeR$]

with its size, this should be a single slot card! whats wrong with a small efficient card? not everything has to be massive to get the job done.

 

[marios15]

AIB partners logic:
50Watts???PUT TWO 8-PINS, 4 HEATPIPES AND 3 FANS ON THAT THING!!!!

 

[xorbe]

AIB partners logic:
50Watts???PUT TWO 8-PINS, 4 HEATPIPES AND 3 FANS ON THAT THING!!!!