RX VEGA - Too shine will need Dev Support

[pmc25]

It is an optical shrink of 16FFC - same design rules but scaled down - also I believe nVidia have access to 9 track libraries over the normal 6? which is allowing them to squeeze a bit more from it.



While the more suited 7nm+ w/ EUV won't be available until about another year TSMC is catering for all market segments with their standard and HPC, etc. implementations of 7nm including GPUs like Volta - people already have products of this kind of requirements in testing with TSMC.

No, they don't, and Volta is categorically a 12nm (16nm) SHP product.

It would require total redesign (i.e. different architecture) to have such a massive optical shrink to either 10nm or 7nm, and a lower power node.

If you weren't trolling and actually believed Volta could or would use a TSMC 7nm or 10nm process, within its design lifetime, explain why it won't, and will face a 2 node deficit to AMD's competing product.

 

[heartburnron]

I'm still planning on getting one - the liquid cooled version too but do we really think it will take a 1000w PSU to run it at the boost clock speed? I thought I'd be safe with my 850w - it was considered safe to run 2x 390Xs in crossfire...

 

[Rroff]

No, they don't, and Volta is categorically a 12nm (16nm) SHP product.

It would require total redesign (i.e. different architecture) to have such a massive optical shrink to either 10nm or 7nm, and a lower power node.

If you weren't trolling and actually believed Volta could or would use a TSMC 7nm or 10nm process, within its design lifetime, explain why it won't, and will face a 2 node deficit to AMD's competing product.

I'm not trolling - TSMC themselves have stated they are launching simultaneously both mobile and high performance variants of their 7nm node additional to the coming plus version:

"TSMC set another industry record by launching two separate 7nm FinFET tracks: one optimized for mobile applications, the other for high performance computing applications."

Unlike the change from 28nm to 20nm planar based technologies like 16FF where a direct shrink wasn't possible due to the differences in lithography TSMC's first round of 7nm is much more compatible to direct shrink of designs for nodes based off 20nm planar.

 

[pmc25]

I'm not trolling - TSMC themselves have stated they are launching simultaneously both mobile and high performance variants of their 7nm node additional to the coming plus version:

"TSMC set another industry record by launching two separate 7nm FinFET tracks: one optimized for mobile applications, the other for high performance computing applications."

Unlike the change from 28nm to 20nm planar based technologies like 16FF where a direct shrink wasn't possible due to the differences in lithography TSMC's first round of 7nm is much more compatible to direct shrink of designs for nodes based off 20nm planar.

No specific products have been launched, nor will they be ready. They've stated similar things on other nodes. They're always later, usually much; it's the same at all foundries. Except the originally IBM designed GF 7nm is by definition a high performance process - it's not intended for mobile stuff.

Do you actually believe that NVIDIA would have paid TSMC to produce an expensive new SHP derivative of 16nmFF purely for V100, which given the titanic die size will have incredibly low yields anyway (some wafers might have no good chips at all)? They're the sole customer. Even at the price they're selling them for, this would almost certainly ensure that they turned their biggest margin product segment into a huge loss maker.

Also, if you believe that Q2 '18 V102/104/106 will be using a high power variant of TSMC 7nm, then I have a bridge to sell you.

Volta is '12nm' through the range.

 

[Molesy79]

I'm still planning on getting one - the liquid cooled version too but do we really think it will take a 1000w PSU to run it at the boost clock speed? I thought I'd be safe with my 850w - it was considered safe to run 2x 390Xs in crossfire...

IMO you'll be fine with a good 650w PSU tbh

 

[Rroff]

Do you actually believe that NVIDIA would have paid TSMC to produce an expensive new SHP derivative of 16nmFF purely for V100

You do realise why they did this right? they had legally binding commercial obligations with huge penalties to deliver GPUs with capabilities designed for Volta within a strict thermal and power spec, that was impossible to hit on normal 16FF+, due to the type of installation they were going in. This was purely down to 10nm not being ready in time for their needs and nothing to indicate what they might or might not do going forward.

What they might or might not do in Q2 18 is one thing but what I am saying is that don't assume they are going to be stuck on a 2 node difference and be unable to compete with Navi.

 

[pmc25]

You do realise why they did this right? they had legally binding commercial obligations with huge penalties to deliver GPUs with capabilities designed for Volta within a strict thermal and power spec, that was impossible to hit on normal 16FF+, due to the type of installation they were going in. This was purely down to 10nm not being ready in time for their needs and nothing to indicate what they might or might not do going forward.

What they might or might not do in Q2 18 is one thing but what I am saying is that don't assume they are going to be stuck on a 2 node difference and be unable to compete with Navi.

Of course they're doing the weird custom node purely for them because 16nmFF wouldn't suffice, and smaller nodes aren't ready.

The economics of doing just V100 on it would not work, and TSMC 10nm let alone 7nm processes suitable for smaller Volta chips will not be ready in H1 '18.

Volta is a '12nm' architecure, whether you like it or not.

It certainly won't be a long lived architecture, because they'll need a smaller node. Were it 2 years, like Pascal or Maxwell, that would probably see them totally wiped out. They will be hoping that NAVI gets pushed into at least Q2 '19. If Volta's successor is on the market in Q3 or Q4 '19 on 10nm or 7nm, then that will be no great loss to them. If NAVI is Q3 or Q4 '18, or even Q1 '19, then it begins to look pretty bad.

 

[shankly1985]

Why wouldn't you care what Nvidia is doing or bringing to the table? To act like they don't exist for a pro AMD approach is pure fanboyism.

I have no interest in buying Nvidia. I'm very well aware of what they're offering however which is why I find vega to be such a failure. And would rather not purchase any card.

Because I have no intention of buying why would I care? Let people care that are interested in there products.

 

[Rroff]

Of course they're doing the weird custom node purely for them because 16nmFF wouldn't suffice, and smaller nodes aren't ready.

The economics of doing just V100 on it would not work, and TSMC 10nm let alone 7nm processes suitable for smaller Volta chips will not be ready in H1 '18.

Volta is a '12nm' architecure, whether you like it or not.

It certainly won't be a long lived architecture, because they'll need a smaller node. Were it 2 years, like Pascal or Maxwell, that would probably see them totally wiped out. They will be hoping that NAVI gets pushed into at least Q2 '19. If Volta's successor is on the market in Q3 or Q4 '19 on 10nm or 7nm, then that will be no great loss to them. If NAVI is Q3 or Q4 '18, or even Q1 '19, then it begins to look pretty bad.

Given they are struggling to get Vega actually out to customers by 2018 at this rate it would be hugely optimistic that Navi was coming H2 2018

Though as I've said before 7nm from both GF and TSMC is ramping up at pace so I wouldn't write off next generation GPUs turning up sooner rather than later.

I think you are seriously under-estimating where TSMC are actually at never mind where they will be in 2018, 10nm saw some hiccups and 12FF is serving a gap in their product line up beyond just nVidia but 7nm seems to be on track. We are back, for now, in the days where more direct shrinks are more possible - Volta certainly isn't "stuck" on 12FF regardless of what nVidia actually do with it - I suspect its quite likely we will see a 12FF run in the mean time though.

EDIT: PS I know where you are getting a lot of your info from - and I've shown before in response to your posts they tend to have a heavy bias compared to the truth with stuff often turning up much sooner or in better shape than they'd allow for.

 

[D.P.]

http://www.anandtech.com/show/11337/samsung-and-tsmc-roadmaps-12-nm-8-nm-and-6-nm-added/2

TSMC will have 7nm high volume shipping in Q2 2018, that is well ahead of Global Foundaries schedule.

 

[Rroff]

http://www.anandtech.com/show/11337/samsung-and-tsmc-roadmaps-12-nm-8-nm-and-6-nm-added/2

TSMC will have 7nm high volume shipping in Q2 2018, that is well ahead of Global Foundaries schedule.

Was talking to someone recently who works in relation to ARM stuff - apparently GF 7LP is currently struggling to hit 30% frequency uplift over their already not particularly headline speeds while TSMC 7nm is enjoying better than 35% performance uplift in the same kind of design scenario.

I think any hopes for GF 7nm being a dramatically different story in terms of implementation success compared to their rather woeful attempts with 14nm is being a bit optimistic unless things change dramatically.

EDIT: Not that this is particularly relevant to potential high performance implementations but still I think it tends to indicate its probably prudent not to get too hopeful 7nm at GF will be a dramatically different story to 14.

 

[CAT-THE-FIFTH]

What is a thread meant to be about RX Vega got to do with who wins some battle to XYZ node??

Ultimately it will be people like Apple who win it anyway.

 

[pmc25]

Given they are struggling to get Vega actually out to customers by 2018 at this rate it would be hugely optimistic that Navi was coming H2 2018

Though as I've said before 7nm from both GF and TSMC is ramping up at pace so I wouldn't write off next generation GPUs turning up sooner rather than later.

I think you are seriously under-estimating where TSMC are actually at never mind where they will be in 2018, 10nm saw some hiccups and 12FF is serving a gap in their product line up beyond just nVidia but 7nm seems to be on track. We are back, for now, in the days where more direct shrinks are more possible - Volta certainly isn't "stuck" on 12FF regardless of what nVidia actually do with it - I suspect its quite likely we will see a 12FF run in the mean time though.

EDIT: PS I know where you are getting a lot of your info from - and I've shown before in response to your posts they tend to have a heavy bias compared to the truth with stuff often turning up much sooner or in better shape than they'd allow for.

Yet again, for the Nth time:

Volta is a 12nm product. You are the only person to try to claim that non-V100 Volta is likely to be 7nm ... I haven't even seen a single person claim it will be 10nm, let alone 7nm.

Your theories will be a huge surprise to NVIDIA, because they seem to think it's 12nm top to bottom.

 

[Rroff]

Yet again, for the Nth time:

Volta is a 12nm product. You are the only person to try to claim that non-V100 Volta is likely to be 7nm ... I haven't even seen a single person claim it will be 10nm, let alone 7nm.

Your theories will be a huge surprise to NVIDIA, because they seem to think it's 12nm top to bottom.

I haven't claimed that Volta GeForce products will be on 7nm - I said it would be a mistake to assume that nVidia is stuck on 12nm and hence 2 nodes behind should Navi release in a timely fashion on a successful 7nm. Backing that up with the fact that 7nm is currently well on schedule and relatively easy, compared to for instance 28nm to 16nm, to shrink to from 12FF.

 

[pmc25]

Was talking to someone recently who works in relation to ARM stuff - apparently GF 7LP is currently struggling to hit 30% frequency uplift over their already not particularly headline speeds while TSMC 7nm is enjoying better than 35% performance uplift in the same kind of design scenario.

I think any hopes for GF 7nm being a dramatically different story in terms of implementation success compared to their rather woeful attempts with 14nm is being a bit optimistic unless things change dramatically.

EDIT: Not that this is particularly relevant to potential high performance implementations but still I think it tends to indicate its probably prudent not to get too hopeful 7nm at GF will be a dramatically different story to 14.

That sounds rather unlikely. Seeing as GF 7nm is mainly aimed at IBM Power, Zen2 and NAVI initially. There's no way any ARM embedded chips or microcontrollers will be on it, and it's highly unlikely anyone orders ARM SoCs on it over Samsung or TSMC's 7nm, when their processes are designed from the ground up to mainly be for LP ARM chips ...

 

[Rroff]

That sounds rather unlikely. Seeing as GF 7nm is mainly aimed at IBM Power, Zen2 and NAVI initially. There's no way any ARM embedded chips or microcontrollers will be on it, and it's highly unlikely anyone orders ARM SoCs on it over Samsung or TSMC's 7nm, when their processes are designed from the ground up to mainly be for LP ARM chips ...

I'm talking about someone who works in that industry, albeit primarily in relation to development of products based off ARM, not that anyone is making ARM devices on GF 7LP.

 

[pmc25]

I haven't claimed that Volta GeForce products will be on 7nm - I said it would be a mistake to assume that nVidia is stuck on 12nm and hence 2 nodes behind should Navi release in a timely fashion on a successful 7nm. Backing that up with the fact that 7nm is currently well on schedule and relatively easy, compared to for instance 28nm to 16nm, to shrink to from 12FF.

You're implying exactly the same thing.

Also, are you seriously trying to claim that a 2 node shrink of the same architecture, to a much lower power process would be "relatively easy" ? Furthermore, what would they do with these newly tiny chips that would probably clock well under 12nm SHP Volta? Their version of Infinity Fabric is highly unlikely to be ready any time soon, so I don't think they'll be going chiplets with an imagined shrunk Volta, like NAVI ... and that would require a totally different memory controller anyway.

Volta is 'stuck' on 12nm. Period.

NVIDIA will hope to not be 'stuck' on Volta for long, or that NAVI is late.

 

[pmc25]

I'm talking about someone who works in that industry, albeit primarily in relation to development of products based off ARM, not that anyone is making ARM devices on GF 7LP.

I simply don't believe you. The processes while directly comparable in size and density are not at all comparable in what will be their initial forms for any design scenarios, as you previously put it ... so why would an ARM engineer be directly comparing them?

When they've both been around a while, and both are offering HP and LP derivatives, and EUV, it would make sense to compare them. Now? Utterly pointless. Especially since the prototype chips that will have been produced in the labs on both nodes will have been completely different. One being ARM SoCs aimed at phones and tablets, one lot being monster IBMPower chips ... I highly doubt there are any entire Zen2s yet.

 

[Rroff]

You're implying exactly the same thing.

Also, are you seriously trying to claim that a 2 node shrink of the same architecture, to a much lower power process would be "relatively easy" ? Furthermore, what would they do with these newly tiny chips that would probably clock well under 12nm SHP Volta? Their version of Infinity Fabric is highly unlikely to be ready any time soon, so I don't think they'll be going chiplets with an imagined shrunk Volta, like NAVI ... and that would require a totally different memory controller anyway.

Volta is 'stuck' on 12nm. Period.

NVIDIA will hope to not be 'stuck' on Volta for long, or that NAVI is late.

Not implying the same thing at all. Relative being the key word there a 2 node shrink of Volta from 12FF to 7nm is quite feasible don't forget that it was for a long time designed for use on 10nm but 10nm didn't come up with the goods in time for nVidia to meet their contractual obligations - they didn't go with 12FF just for fun when 10nm was only a relatively short delay away - its ramping upto production now - under any other circumstances it would have been insane for nVidia not to delay for 10nm.

Oh come on Navi with a proper "chiplet" type design is a long way off yet - same as nVidia's own take on MCM while its likely first generation Navi will lay the ground for it. You can't just slap current monolithic GPU architectures together with an interconnect like IF and get any significantly better results than current mGPU as you are still stuck with the same rendering limitations. nvlink is more nVidia's answer to IF - for their eventual MCM design they will be using highly specialised interconnects and not something that is a competing product to IF which will remain nvlink.

 

[Rroff]

so why would an ARM engineer be directly comparing them

I'm not talking about an ARM engineer. I'm talking about someone who works for a company making custom ASICs and FPGAs, etc. albeit their job has more involvement with regard to using ARM based products.