Yes, that was my point. We've got a 530mm² GPU barely outperforming a 170mm² one by 20%, using twice the power. That's just terrible, and a mobile Cypress would trounce it.
NVIDIA GF100 & Friends speculation
- Thread starter Arty
- Start date
Yes, that was my point. We've got a 530mm² GPU barely outperforming a 170mm² one by 20%, using twice the power. That's just terrible, and a mobile Cypress would trounce it.
I really cant understand why they didnt wait for GF104. On the other hand, it makes me scared of what GF104 really is.
They need to sell out deffective parts...
Not really abt the defective parts, its more difficult to bin the chips for use in laptops. Not all chips will make the grade. Its more abt time to market and the desire to claim that they have the fastest mobile graphics card(ie the halo effect). GF104 still hasnt released on the desktop and mobile parts usually lag desktop parts by 2-3 months. GF104 would have released only by christmas. Coupled with the production problems and the fact that its going to cost an arm and lag, i dont think we'll see too many GTX480M's being sold
In this case, the GTX 480M is clocked so low that it may actually help yields a little. Then again, the top-end desktop-replacement market is so small that it probably doesn't matter much. Plus, GF104 is supposed to be a half-GF100, give or take a few things, so it wouldn't be as fast...
GF104 better be as fast.. Or else this generation is completely bust for nVIDIA. I cant believe GF104 does not beat Juniper...
That's crazy. I guess Nvidia are trying to get every single dollar they can from any salvage parts possible. Surely hardware vendors that will be producing the laptops aren't stupid? Don't they look at issues such as package size, perf/watt and heat which are very important for laptops?
Not really, the low clocks are for TDP reasons, and they have to cherry pick chips with low leakage and voltage requirements that fit the TDP.
Sure it might only be a half GF 100, but the GTX 480M is only 3/4th of a full GF100 and with 2/3rd the clock speed. GF104 is probably gonna end up close to the GTX480M if they can clock it high enough(ie at GTX 280M/285M clocks)
Package size will be standard MXM, thats what most desktop replacements use. Perf/watt really does not matter here, its all about performance(similar case with GTX 480 desktop parts). Heat is an issue but they were already at 75W with previous gen parts and we had dual cards in SLI as well. So a 100W is still doable
Silent_Buddha
Legend
Don't forget also additional disabled SPs. So definitely low clocks + disabled SPs = easier to salvage what would otherwise go into the recycling bin.
This definitely does seem like a salvage part. I have a feeling Mobility 5870 using the same CPU would have scored similar, or at least closer to the Mobility 480 in Vantage.
Heck if AMD felt threatened by this, which they probably don't, they could release a Mobile version of the Cypress and wipe the floor with it and still have lower power use and heat.
Regards,
SB
GF104 should beat Juniper on desktop (by roughly 10%, I would guess) but according to Fudo it's a 180W part, while Juniper has a 108W TDP. Let's hope that 180W figure is more realistic than the GTX 480's 250W "TDP".
So given a power budget of 100W, could it do as well as the GTX 480M?
Sure, but there are always a few who'll build a system with the fastest part, no matter what.
Then again, low-leakage parts tend to be the ones that don't clock very high, don't they?
I seriously doubt they'll reach 1500MHz for the shaders on the mobile GF104 when even the desktop GF100 can't do it, even with 32SPs disabled. Unless they made profound modifications to the design, that is.
We really have no idea how the GF104 will perform but that would be my guess as well. Juniper has a 108W TDP on the desktop with 850/1200 clocks while the mobile version has 700/1000 and reduces TDP to 50W. So similarly if the GF104 is actually 180W, they could easily reduce it to under a 100W with a similar reduction in clock speed.
So could mobile GF104 which is supposed to be half a GF100 X 3/4 desktop clocks(which should be higher than GF100 desktop clocks, similar to how G92 clocked higher compared to GT200b) be as fast as GTX480M which is 3/4 GF100 X 2/3 desktop clocks?
From what i remember reading earlier in this thread, apparently they do. In this case this works fine for them given the TDP limits
Well the desktop GF100 couldnt clock as high due to heat issues(and possibly not enough parts yielding that high a clock speed). Similar to how G92b clocked higher than GT200b, i expect GF104 to clock higher than GF100
We really have no idea how the GF104 will perform but that would be my guess as well. Juniper has a 108W TDP on the desktop with 850/1200 clocks while the mobile version has 700/1000 and reduces TDP to 50W. So similarly if the GF104 is actually 180W, they could easily reduce it to under a 100W with a similar reduction in clock speed.
So could mobile GF104 which is supposed to be half a GF100 X 3/4 desktop clocks(which should be higher than GF100 desktop clocks, similar to how G92 clocked higher compared to GT200b) be as fast as GTX480M which is 3/4 GF100 X 2/3 desktop clocks?
Careful there; again, barring profound changes to the design in GF104, there's a problem: static power is a bitch on Fermi. The GTX 480 idles at 50/100MHz, yet draws 50W. That's an issue that lower clocks won't help you with, and you need to reduce voltage, which was apparently not possible on the GTX 480.
I believe that's why the 480M only has 352SPs: NVIDIA picked the 5 worse SMs (those needing the highest minimum operating voltage) and disabled them so that they could scale voltage down, and keep the TDP at 100W (again, hopefully it's real).
So if GF104 is just a half-GF100, NVIDIA would need to do something similar, albeit on a smaller scale. They could probably get away with disabling only one or two SMs. Let's be optimistic and say only one. Now, let's be optimistic again and say that on desktop, GF104 outperforms the HD 5770 by about 15% (it can't be much more than that, because the GTX 465 is about 25~30% faster).
So, you take one SM out, scale down voltage a bit, reduce clocks by a good 20% and you get (224/256) × 80% = 70%. So the mobile GF104 would be about 70% of the desktop version. Or 70% × 115% = 80.5% of the HD 5770.
The mobile Juniper is 82% of its desktop counterpart.
So give or take a few points here and there, the mobile GF104 should roughly match the Mobility Radeon HD 5870, maybe edge it out, not outperform it by 20%... at 100W.
On desktop, yes, that's likely. I think GF104 should be able to achieve 1450~1500+MHz without too much trouble. On notebooks, that's a different story.
Well my expectations of GF104 are much different then yours, then. I expect GF104 to wipe the floor with Juniper, IF they have more than 256 SP, which i think they will. We have already seen that GF108 mobile has 96 SPs, despite some people thinking its an error. So if GF108 is a 96 SPs part what would be a GF106? Certainly not a 128 SPs part, IMO.
Im more inclined to believe:
GF104 - 384 SPs (max)
GF106 - 256 SPs (max)
GF108 - 128 SPs (max)
GF119 - 64 SPs (max) for ION.
I know this might seem a little bit farfetched (ex: lower-mid end with as much shaders as a G80), but remember that GF100 was short of the mark. If things did went well and would they really hitted their goal with GF100, this speculation might have made sense.
Im more inclined to believe:
GF104 - 384 SPs (max)
GF106 - 256 SPs (max)
GF108 - 128 SPs (max)
GF119 - 64 SPs (max) for ION.
I know this might seem a little bit farfetched (ex: lower-mid end with as much shaders as a G80), but remember that GF100 was short of the mark. If things did went well and would they really hitted their goal with GF100, this speculation might have made sense.
Silent_Buddha
Legend
It's not quite that simple either is it? I could be wrong, but wasn't part of GF100's power problems related to GDDR5? Similar to some of the pains ATI went through when they introduced it?
Then again, I could be wrong. My brain is halfway addled here working 60-72 hour weeks.
Regards,
SB
It's not quite that simple either is it? I could be wrong, but wasn't part of GF100's power problems related to GDDR5? Similar to some of the pains ATI went through when they introduced it?
Then again, I could be wrong. My brain is halfway addled here working 60-72 hour weeks.
Regards,
SB
If memory serves, the GTX 480 operates under 1.05V in full load, and 0.99V while idling. Clearly, there's a voltage problem.
As for GDDR5, I'm not sure, but the memory is clocked at 67MHz when the card is idle, so...
yes....and?
So that 20.6% would be notably less if the CPU's used were equal
I think is not a good test (better: pre-test), but 3d vantage is not much cpu-dipendent
aaronspink
Veteran
That's not necessarily a voltage problem, it just means they are operating just above Vmin.
Similar threads
- Deleted member 2197
- Architecture and Products
