Read more.Quote:
Features 768 CUDA cores that can boost to 1380MHz. Said to arrive mid-October.
Printable View
Read more.Quote:
Features 768 CUDA cores that can boost to 1380MHz. Said to arrive mid-October.
So basically a GTX 950 with higher clocks? I suppose the big question will be the pricing. £130 (~$149?) would make it 30% more expensive than the cheapest RX 460, for around 30% more performance?
Interesting that it's clocked a lot lower than the other Pascal parts though ... weight to the rumour that GP107 will be fabbed on Samsung 14nm?
Or perhaps this is actually a load of rubbish (although the specs wouldn't be at all surprising given the market position)...?
If the clockspeeds are true,it shows you how much performance AMD has potentially lost using GF 14NM for desktop parts!! :(Quote:
Originally Posted by scaryjim
Its also a shame we are not seeing a consumer desktop part with a full Polaris 11 chip either.
ftfy ;)Quote:
Originally Posted by CAT-THE-FIFTH
I don't think there's much doubt that AMD has deliberately gone for using one set of silicon for both mainstream laptop and mainstream desktop GPUs this time around. I look forward to seeing the laptop implementation of Polaris 10 (actually, I'd love to get hold of an RX 470 and mess with the clocks and voltages to see just what it could: I can't help thinking that it might be able to creep down towards a 75W TDP at ~ 950MHz/0.9v (which is probably the silicon sweet spot)).
TBH, in the mainstream space I don't think they've lost all that much - the absolute performance isn't so critical if you're deliberately not going for the high end, and they've been able to engineer good cards despite being on the ragged edge of the capabilities of the process. And - if the rumours are true about nvidia going to Samsung 14nm for GP107 - I've got to wonder whether it's significantly cheaper per working die than TSMC 16nm...
75W would be really odd. Wouldn't you either clock a little slower for 70W and stay bus powered, or stick a 6 pin socket on it and crank clocks to the moon? If 75W *is* clocks cranked to the moon, then that implies it need a 6 pin power connector anyway so perhaps the power is a bit bursty.
75W *is* bus powered. It's the hard edge of bus powered, but it's there. It's another indicator - if this is all true - that it might not be on TSMC 16nm - 1280 shaders @ 1700MHz for 120W TDP, or 768 @ 1380MHz for 75W? A naive calculation* (based on linear scaling for both clock speed and shaders) would suggest it should have around half the TDP of a GTX 1060 - around 60W. A higher than expected TDP could easily be down to it being pushed to the limits of a less performance-focussed process...Quote:
Originally Posted by DanceswithUnix
*I rather suspect thatmy naive calculation is actually on the high side - whilst it doesn't account for the memory subsystem (which is 2/3 of the 1060, rather than 1/2) lower clocks usually allow for a lower voltage to be employed, and power usage scales with the square of voltage so this has a larger effect on bringing the power down.
Ta, of course it is, must be time for a coffee :DQuote:
Originally Posted by scaryjim
The RX480 memory interface clocks happily at 2GHz so I'm not convinced we can blame transistor speed on the process, I think Nvidia just have a clock monster design (which makes the power usage more impressive). Perhaps there are Arm A57 designs out there on the two processes that would give a better comparison for what they are capable of in terms of clock speed. Dunno, I'm off to that coffee machine :)Quote:
Originally Posted by CAT-THE-FIFTH
There's a rather significant margin over the 75w limit where the PCIe slot can still power the card, Nvidia's been using it for a while on their GTX x50s and AMD probably would've gotten away with it too with the RX480 if it wasn't a higher power part. Also few people buying this kind of part are going to overclock so there's not much reason to add the headroom to a reference card as it's not a massive selling point.Quote:
Originally Posted by DanceswithUnix
When is it not? ;)Quote:
Originally Posted by DanceswithUnix
The problem with that is that there's no physical spec requiring the motherboard to handle that much power. Don't forget that, whilst we talk about "the slot" powering the card, what we mean is "the 24pin power connector and motherboard traces". Traces are pretty high resistance, and the ones from the 24pin connector to the PCIe slot are going to be fairly long. That's not a system you want to be pushing beyond its rated spec.Quote:
Originally Posted by HW90
PCIe cables, on the other hand, have physical specifications that can handle currents way outside of the electrical spec for PCIe - iirc it's up to around 190W for a 6 pin and 290W for an 8 pin, and those have a reasonable tolerance built in too. So if you're going to risk pulling more than 75W, you might as well slap a 6 pin connector on there and use it for some, or even all, of the 12v power you need...
Well have you noticed the clockspeeds on the GTX1050?? 1.3GHZ to 1.4GHZ,instead of 1.5GHZ to 1.6GHZ upwards for all other Pascal parts?? That is a heck of a loss in clockspeed if the process node not to blame.Quote:
Originally Posted by DanceswithUnix
Maxwell was a high clockspeed design too and Pascal is based on it.
Even mobile Pascal cards seem to be above 1.5GHZ too,and apparently Nvidia might have engineered slightly slower shaders,so they could run the cores faster(people have made some comparisons IIRC).
I expect the GTX1050 will be under 75W too,so I think that would rule out a TDP limitation.
If it is 75W,then again that looks like a process node issue.
Plus,its GF,they have a history of issues.
However,AMD signed the WSA,so they have to use GF.
It does make me worried whether Zen will hit sufficiently high clockspeeds.
Kaveri was massively delayed and did not hit the clockspeeds AMD wanted as GF 28NM had issues. GF follow-up process was cancelled.
The Samsung process they are using was developed more with mobile chips in mind too,and they had issues implementing it too,apparently.
TSMC,has far more experience with bigger chips too.
Or it could be simply the case,they realised that 14NM was only really optimised for that kind of voltages and clockspeeds??Quote:
Originally Posted by scaryjim
Remember - AMD themselves have said Polaris 11 is the laptop optimised part,especially since it is meant to have low Z height too.
Plus,its worrying that the RX460 is only a 896 shader part - it makes me wonder how bad yields must be if the desktop parts are primarily all salvaged parts??
Remember,until recently the RX460 was a 1024 shader part in AMD bumpf.
IIRC,I read somewhere there are certain power optimisation options which only Polaris 11 has(Polaris 10 does not have them).
14NM is meant to be denser than 16NM,which should help with costs,but it makes me wonder whether it comes with the disadvantage of clockspeeds.
I wonder what company AMD will use for Vega??
I think TSMC,especially since they have experience with interposers too with AMD Fiji.
Only allow myself one per day, but as it is lunchtime I am now on Pepsi Max so caffeine level should be maintained :)Quote:
Originally Posted by scaryjim
Sort of true, but:Quote:
Originally Posted by scaryjim
1/ Some traces are wider than others, and wide traces are surprisingly capable.
2/ I would expect a major voltage level to go to a plane rather than a trace. That gives you an entire PCB layer, a rectangular slab of copper with holes in it to route signal vias through, and a *lot* of current capability.
I would have thought the biggest problem was on most PSUs you don't want all the power coming out of one or two rails in a multi rail design. So if you are going to have to plug those rails in somewhere, straight into the video card seems like a sensible move.
Could be a tweak to the shaders to make them more power efficient, or a denser layout library on the same node to reduce costs. Or that a smaller die is more prone to thermal issues increasing leakage.Quote:
Originally Posted by CAT-THE-FIFTH
It could be that they've shifted to a different process/node/fab.
It could be that this rumour is wrong, or is based on an engineering sample that's not representative of shipping silicon.
At the minute, it's all conjecture...
Quote:
Originally Posted by scaryjim
Well considering unlike some of you I have owned a Maxwell card too(and have mates who have them too) - the Maxwell V2 cards across the whole range boost quite similarly!!Quote:
Originally Posted by scaryjim
Also,considering even the mobile Pascal parts are clocking above 1.5GHZ it seems more like a process node issue,if the GTX1050 is not hitting similar base clockspeeds on GF 14NM.
Plus its funny,some of you are now defending GF,just so you can have an opposite viewpoint to me(just for the sake of having it,like in the last few threads),which contradicts what you implied before.
OFC,as you very well know,GF has had loads of issues in the last few years,so know apparently that should be ignored. All the missed targets,etc - its all well know.
I expect the only reason Nvidia would want to use Samsung ,is since TSMC 16NM might be over subsribed or Samsung is desparate to get more business since they lost those massive Apple contracts.
You known this very well - the fab in NY was built to service Apple contracts.
So know what you are saying is conjecture,since you choose to ignore history.
The only way what you are saying is not conjecture is if they are using TSMC 16NM still.
Not defending GloFo CAT, just pointing out that we don't know if those are the shipping clocks, or if it's being fabbed by Samsung, or ANYTHING.Quote:
Originally Posted by CAT-THE-FIFTH
GLoFo/Samsung's 14nm process being less clock oriented than TSMC's 16nm is certainly one possibility, as I've already pointed out. But there are other possibilities, and I think it makes sense to keep some perspective about rumours.
TSMC has experience producing graphics chips - the GF/Samsung process was developed mostly for mobile products,ie,mostly for Apple. Samsung would not have predicted Apple would abandon them for TSMC. Even the XBox One refresh is TSMC.Quote:
Originally Posted by scaryjim
If the GP107 is 16NM,then you are probably right. If it is 14NM,then there is something not quite right with the process. Its a massive drop in clockspeeds over Pascal mobile chips(above 1.5GHZ base clockspeeds) - 1.3GHZ to 1.4GHZ is not that much higher than the 1.2GHZ for the RX460 especially when you realise on the same node,Maxwell had much higher clockspeeds than what AMD could achieve.
Its worrying for Zen. Remember what happened with Kaveri?? Prototypes were much higher clocked but when it came to actually producing them in quantity,it massively missed clockspeed targets and was massively delayed. The same happened with Llano.
If AMD has issues with Zen clockspeeds,it would be a disaster.