Geforce GTX1180/2080 Speculation thread

bigmike20vt · 17 Jul 2018 at 10:30

Paladine said:
Nvidia have increased the limit on their existing range in their web shop from 2 to 10 per person. Seems they are trying to shift as much stock as possible before the end of the month.

if they knocked £200 off the current RRP they would sell a lot more

btw, this 300k cards which were allegedly sent back to NV... how will NV sell them? Are these literally just the chips or something, because surely aftermarket boards tend to have their own design so it is not like nvidia can put their own blower cooler on it and sell as an nvidia card? Even if their cooler fits on it, that could cheese off a buyer if, for instance, they thought they were getting an nvidia built card, but it turned out to be a palit PCB with a different cooler on it?

(sorry if i am talking crap)

D.P. · 17 Jul 2018 at 10:39

Gerard said:
Somehow, i think the team of engineers would have realised that in the decade plus it took to get it to production if it were the case.

They knew it but not much they can do about it because they can't make interposer much bigger.

HBM is still likely more efficient overall but I don't think it gains quite as much as one might expect. The biggest failing of HBM is you don;t really get the bandwidth advantages until you get 4 or more stacks, by which point the cost and complexity have soared well over GDDR

4K8KW10 · 17 Jul 2018 at 10:40

Kaapstad said:
I happen to believe that HBM actually increases power consumption a bit due to its proximity to the GPU core resulting in poorer cooling and higher running temps. As we all know the hotter something runs the more power is needed to make it happen.

Do we know how warm do these HBM2 chips running at mere 950MHz get?
Usually, memory chips don't require dedicated cooling, even less so at low clocks.

Navi with next-generation memory, read HBM3..

D.P. · 17 Jul 2018 at 10:45

4K8KW10 said:
Vega has 50% the memory bus width of Fiji - 2048-bit vs 4096-bit. Also, historically, AMD aren't able to optimise their chips with the first iteration of a new memory.
With GDDR5, they had RV790 sticking to full-memory speed at idle.

So, 50% less memory bandwidth, and ultimately, slower clocks on their own.

What rubbish.

Vega does't have 50%of the bandwidth of Fiji.
The Furyx had 512GB/s
Vega64 has 483.8 GB/s

You are getting confused with bus width. FuryX needed 4 stacks of HBm1 to get only 4GBM of VRAM. This added much more expensive and lowered yields. AMD were lucky in that HBM2 standard increased density allowing 2 stacks to be used, but this meant they couldn't increase bandwidth at all.

4K8KW10 · 17 Jul 2018 at 10:48

Vega with 4096-bit MI and the same 950MHz HBM2 would have ~970GB/s bandwidth.

Gerard · 17 Jul 2018 at 10:58

Frontier edition has 16 gigs of hbm, wonder if that's 4 stacks or 2.

Gerard · 17 Jul 2018 at 11:01

D.P. said:
They knew it but not much they can do about it because they can't make interposer much bigger.

This all would have been taken into consideration when it was being designed, one of the key features of hbm is less power draw. I'm pretty sure those claims have been scrutinised since its inception in fury-x and if lower power wasn't the case an article or 2 would have been done about it by now.

KillBoY_UK · 17 Jul 2018 at 11:46

4K8KW10 said:
Vega with 4096-bit MI and the same 950MHz HBM2 would have ~970GB/s bandwidth.

While Vega is bandwith staved I dont think doubling the bandwidth is going to make the performance difference they need. Also the memory controllers on AMD cards suck up loads of power by them self, doubling the size of it is nto going to help.

I do miss what the R520 did back in the day a 512bit ring bus which was closer to the size of 256bit bus's of the time.

D.P. · 17 Jul 2018 at 12:41

Gerard said:
This all would have been taken into consideration when it was being designed, one of the key features of hbm is less power draw. I'm pretty sure those claims have been scrutinised since its inception in fury-x and if lower power wasn't the case an article or 2 would have been done about it by now.

I'm not doubting the lower power figures, I'm just given a slightly nod to Kapp that the heat of the HBM will increase the heat of the GPU die and power consumption rapidly increases with heat, offsetting a small amount of the power gains..

HBm2 will save maybe 30w but for a lot of extra expensive. For the GV100 this is OK with the bandwidth offered from 4 stacks, but i don;t think it has helped Vega much and was a bit of a disaster for Fiji.

D.P. · 17 Jul 2018 at 12:42

Gerard said:
Frontier edition has 16 gigs of hbm, wonder if that's 4 stacks or 2.

2 stacks, but see the cost difference.

Gerard · 17 Jul 2018 at 13:19

D.P. said:
I'm not doubting the lower power figures, I'm just given a slightly nod to Kapp that the heat of the HBM will increase the heat of the GPU die and power consumption rapidly increases with heat, offsetting a small amount of the power gains..

HBm2 will save maybe 30w but for a lot of extra expensive. For the GV100 this is OK with the bandwidth offered from 4 stacks, but i don;t think it has helped Vega much and was a bit of a disaster for Fiji.

Fiji was designed for hbm, they couldn't just change it at the last minute, the tech was in its infancy so they had to go with what was available.

Kaapstad · 17 Jul 2018 at 13:29

Gerard said:
Somehow, i think the team of engineers would have realised that in the decade plus it took to get it to production if it were the case.

Those same engineers also said that HBM would do other things that has turned out to be a myth.

Kaapstad · 17 Jul 2018 at 13:32

4K8KW10 said:
Do we know how warm do these HBM2 chips running at mere 950MHz get?
Usually, memory chips don't require dedicated cooling, even less so at low clocks.

Navi with next-generation memory, read HBM3..

Even if they generated no heat whatsoever (not going to happen) their physical presence packed around the GPU would still act as insolation making cooling more difficult.

Gerard · 17 Jul 2018 at 14:16

Kaapstad said:
Those same engineers also said that HBM would do other things that has turned out to be a myth.

Such as? It's draws less power than gddr5 , enables shorter boards to be made and the hbcc is essentially what amd were trying to do manually via drivers with Fiji. So what has it not done that it was meant to?

Also I think you're mixing up amd marketing with the people actually making the ram.

Kaapstad · 17 Jul 2018 at 16:07

Gerard said:
Such as? It's draws less power than gddr5 , enables shorter boards to be made and the hbcc is essentially what amd were trying to do manually via drivers with Fiji. So what has it not done that it was meant to?

Also I think you're mixing up amd marketing with the people actually making the ram.

Glad you brought up the fact that shorter boards can be used.

I don't want a shorter board on a gaming card as it is next to useless.

For a gaming card (or even a professional card) a full size cooler is far more important than a worrying about a shorter PCB.

The only use I see for shorter boards is if people want to build mini systems where the compromises are acceptable.

Gerard · 17 Jul 2018 at 16:14

Kaapstad said:
Glad you brought up the fact that shorter boards can be used.

I don't want a shorter board on a gaming card as it is next to useless.

For a gaming card (or even a professional card) a full size cooler is far more important than a worrying about a shorter PCB.

The only use I see for shorter boards is if people want to build mini systems where the compromises are acceptable.

Might be useless for you but other people can have uses for it. The vega pcb could indeed have been shorter than it currently is as there's a lot of empty pcb space, again down to hbm. But they done the same thing they done with fiji and only produced one pcb variant that could be used across cards for 56,64 and FE.

Kaapstad · 17 Jul 2018 at 16:29

Gerard said:
Might be useless for you but other people can have uses for it. The vega pcb could indeed have been shorter than it currently is as there's a lot of empty pcb space, again down to hbm. But they done the same thing they done with fiji and only produced one pcb variant that could be used across cards for 56,64 and FE.

Tell that to AMDs AIB partners who use very large coolers on the latest Vega cards.

Making something small for the sake of it is pointless.

Gerard · 17 Jul 2018 at 16:44

Kaapstad said:
Tell that to AMDs AIB partners who use very large coolers on the latest Vega cards.

Making something small for the sake of it is pointless.

How is it pointless exactly? It gave people with small cases the option of an amd card with good performance that would fit their form factor. You might find it pointless, others don't.

Zotac even jumped on the bandwagon and made a mini version of the 1080ti so obviously theres a market for things like this.

Meaker · 17 Jul 2018 at 17:15

Kaapstad said:
Even if they generated no heat whatsoever (not going to happen) their physical presence packed around the GPU would still act as insolation making cooling more difficult.

Heat only moves a little sideways in a chip, having the HBM there even if the heat did move to those chips they are making contact with the cooler so would help transmit it.

They have little impact in that sense.

nomadd · 17 Jul 2018 at 19:33

Nearly 60 pages fellas..

..and not a single mention of anything credible as regards the GTX 1180.

Keep up the good work.