AMD Teases Its 7 nm Vega Instinct Accelerator - Data-Pushing Silicon Deployed

[Raevenlord]

AMD has announced via its Twitter feed that the Vega die shrink from current 14 nm down to 7 nm has actually coalesced into a hardware product that can be tested and vetted at their labs. Via a teaser image, the company said that "7nm @RadeonInstinct product for machine learning is running in our labs."

Of course, working silicon is only half the battle - considerations such as yields, leakage, and others are all demons that must be worked out for actual production silicon, which may thus be some months off. Only AMD and TSMC themselves themselves know how the actual production run went - and the performance and power efficiency that can be expected from this design (remember that AMD's CEO Lisa SU herself said they'd partner with both TSMC and Globalfoundries for the 7 nm push, though it seems TSMC may be pulling ahead in that field). Considering AMD's timeline for the die-shrunk Vega to 7 nm - with predicted product launch for 2H 2018 - the fact that there is working silicon being sampled right now is definitely good news.



View at TechPowerUp Main Site

 

[Aldain]

This is TSMC based not glo-fo based and given the fact that intel 10 nm is utter garbage and delayed , AMD is looking freaking amazing with the 7 nm process

 

[Valantar]

This is TSMC based not glo-fo based and given the fact that intel 10 nm is utter garbage and delayed , AMD is looking freaking amazing with the 7 nm process

Doubt that. GloFo's 7nm is slated for high volume manufacturing in 2H18.

 

[Aldain]

Doubt that. GloFo's 7nm is slated for high volume manufacturing in 2H18.

No , this is TSMC.. In the AMD conference call Lisa Su was asked does she feel confident that TSMC will deliver on the 7nm front for the upcoming products.. she said yes

p.s TSMC is ahead of glo-fo

 

[Midland Dog]

if amd has tsmcs 7nm it means nvidia has it too

 

[efikkan]

if amd has tsmcs 7nm it means nvidia has it too

Yes, it does.
But like Lisa Su said; "is running in our labs and we remain on track to provide samples to customers later this year",
so they don't expect volume shipments this year.

 

[Midland Dog]

im actually somewhat excited for 7nm i expect intels "10nm" to be a flop like 14nm was with broadwell but every time we have had a node shrink for gpus its been generally a positive experience

 

[ShurikN]

im actually somewhat excited for 7nm i expect intels "10nm" to be a flop like 14nm was with broadwell but every time we have had a node shrink for gpus its been generally a positive experience

You don't need to expect Intel's 10nm to flop, it already did.

 

[Aldain]

if amd has tsmcs 7nm it means nvidia has it too

No it does not..

 

[Midland Dog]

No it does not..

they would make more money from nvidia so of course they have 7nm, just like how they had all the hbm2 even though they were late to the hbm party, im pretty sure they even backed micron yet samsung gave them the hbm supply they needed

 

[Supercrit]

Are they showing that they tamed the 14nm Vega's heat with die shrink? The card doesn't even have a fan.

 

[Aldain]

they would make more money from nvidia so of course they have 7nm, just like how they had all the hbm2 even though they were late to the hbm party, im pretty sure they even backed micron yet samsung gave them the hbm supply they needed

No they would not actually . it seems that EPYC 2 is also tsmc 7 nm , as for this gpu this i only AI . so dont go expecting nvidia to have a 7 nm gaming gpu this year

 

[R0H1T]

they would make more money from nvidia so of course they have 7nm, just like how they had all the hbm2 even though they were late to the hbm party, im pretty sure they even backed micron yet samsung gave them the hbm supply they needed

Say what? AMD makes their console chips from TSMC, Nvidia also has a few products made at Samsung.
The idea that TSMC make more $ from Nvidia, than AMD, is highly suspect. Apple is the only one that has priority access, for many different reasons.
Yeah right, so there were no Intel, Google, Xilinx who gobbled up tonnes of HBM2 either? Not to mention the volume of consumer Vega is anywhere between 2x-5x the Tesla cards with HBM2, also miners!

 

[the54thvoid]

Are they showing that they tamed the 14nm Vega's heat with die shrink? The card doesn't even have a fan.

That's a dual slot card with a long cover. Bearing in mind the die will be smaller and the memory is on die, that's a large space under the hood. It could be water cooled or there could be a fan with an intake at the rear. It will not be passive.

EDIT: it's also a render. The real card in their lab probably looks like Frankenstein.

 

[ppn]

Whell in case Turing is 7nm too I will buy the titan Xv of that, if not 14nm being 3 times bigger it is to be considered ancient. A big no-no.

 

[Shamalamadingdong]

Are they showing that they tamed the 14nm Vega's heat with die shrink? The card doesn't even have a fan.

That's a dual slot card with a long cover. Bearing in mind the die will be smaller and the memory is on die, that's a large space under the hood. It could be water cooled or there could be a fan with an intake at the rear. It will not be passive.

EDIT: it's also a render. The real card in their lab probably looks like Frankenstein.

It's a datacenter/HPC card. They don't have fans. So the render is accurate enough.

 

[iO]

Too bad it wont ever end up in a consumer product.

 

[Casecutter]

For this HPC/AI market all AMD said is they have something running and correct it at best looks like the "Bride of Frankenstein" AMD is on track for engineering samples "a half ready product" by sometime later but supposedly in this year. That kind of says to me the me the throughput/power of the TSCM parts are inline with what they hoped to get. At this point "hearing" anything in regards 7nm consumer parts are in the works is probably best "tempered" till a year from now.

 

[the54thvoid]

It's a datacenter/HPC card. They don't have fans. So the render is accurate enough.

It depends how many cards per rack. It's also machine learning, more nuanced than sheer number crunching power. Good chance when in a rack it will have cooling on board. My brother makes HPC FPGA cards that use cooling (not passive).

 

[Shamalamadingdong]

It depends how many cards per rack. It's also machine learning, more nuanced than sheer number crunching power. Good chance when in a rack it will have cooling on board. My brother makes HPC FPGA cards that use cooling (not passive).

You couldn't fit a fan on the card with more cooling capacity than you'll find in your average rack I'd wager. You can employ various strategies to cool a rack but ultimately you'd most likely end up with worse if you tried to cool individually.

 

[T4C Fantasy]

Whell in case Turing is 7nm too I will buy the titan Xv of that, if not 14nm being 3 times bigger it is to be considered ancient. A big no-no.

turing might be, but if it is the 11 series its Volta.

also im a strong believer and i can be wrong, but i believe 11 series will be volta and all the leaked news is fake.

 

[Fouquin]

It depends how many cards per rack. It's also machine learning, more nuanced than sheer number crunching power. Good chance when in a rack it will have cooling on board. My brother makes HPC FPGA cards that use cooling (not passive).

For large scale compute ASICs it really doesn't depend on count per rack. AMD, nVidia, and Intel all use much larger heatsinks under a plain shroud rather than a smaller heatsink with integrated fan. Fin size and density matter most all day, every day. None of the current generation Radeon Instinct cards have integral fans, that includes cards with Polaris, Fiji, and Vega. Same goes for nVidia's Tesla series, they did away with integrated fans nearly 4 years ago.

 

[Vya Domus]

Finally they can build GPU's on the best nodes available.

 

[eidairaman1]

Finally they can build GPU's on the best nodes available.

Glofo, is that in Dresden?

 

[tarin]

I wonder how machine learning chip differrent from workstation chip or gaming chip.