Making it a multiple of 6 would make sense for GF108 (so not 445M) if my speculation is correct that it's really 3x32 SP rather than 2x48, but then you'd expect 24 TMUs and not 12. None of it makes sense really...
NVIDIA GF100 & Friends speculation
- Thread starter Arty
- Start date
It's definitely wrong. They did the same thing for GF104 based mobile parts and we know that's 8 TMUs per SM.
Those numbers with at the wall measurements aren't particularly reliable, I much prefer direct measurements for that - as far as I can tell the guru3d numbers also include PSU inefficiencies.
So you're suggesting HD5870 draws less nowadays? The site I mentioned tested some newer cards, but they tend to have something extra (like more clock, more memory). Some 2GB variants though indeed seemed to have about the same power draw, so I guess new 1GB variants would indeed have slightly lower power draw but I don't know by how much (10-20W?).
All but TechPowerUp and XBitLabs (and in some cases, XBitLabs too) include the PSU inefficiencies and the whole rig in the power consumption numbers.
And ht4u - from which came the numbers I initially quoted...
But yes most sites certainly just use at the wall numbers. If they'd adjust them for psu inefficiency, they'd not get more reliable...
In any case at guru3d the OC GTX460 delta power was 205W - clearly if you factor in PSU efficiency (a good one might have 90% efficiency) the card power draw was below 200W
Hmm GTX470M vs. GTX480M is quite interesting, and it's not quite obvious GTX480M is always faster. A good thing nvidia doesn't announce gflops ratings too loudly cause the former would actually be very slightly faster using that metric...
These two chips have very similar peak shader and rop throughput. The former has a distinct advantage in texturing throughput, the latter in memory bandwidth and pixel throughput (not so much because of its 3 (?) vs. 2 rasterizers, but because of the 11 vs 6 SMs). I guess that's enough that GTX480M is a bit faster in practice but I wonder how much - in any case I'd suspect in terms of power draw efficiency the GTX470M is in a different class (or rather, the GTX480M defines its own class outside of any other chip).
Something else I'm curious about is the very very low gddr5 memory clock on all mobile chips. It was already that low on GTX480M but that chip wasn't really the right one to draw any conclusions. Is this due to power efficiency? Maybe it's possible to use only 1.35V instead of the nominal 1.5V when operating at such a low frequency? Or would the memory controller need higher core voltage to run at higher frequencies? I'm really wondering what effects that low memory clock has wrt power draw compared to AMD - basically nvidia needs a 192bit bus (hence more chips) to achieve similar bandwidth than AMD with a 128bit bus, using the same gddr5 technology.
Also, possibly related, AMD has pushed gddr5 down to even the slowest chip (cedar). OTOH it seems nvidia is using (well on the mobile front at least) ddr3 exclusively on GF108 (which is really more of a redwood competitor).
These two chips have very similar peak shader and rop throughput. The former has a distinct advantage in texturing throughput, the latter in memory bandwidth and pixel throughput (not so much because of its 3 (?) vs. 2 rasterizers, but because of the 11 vs 6 SMs). I guess that's enough that GTX480M is a bit faster in practice but I wonder how much - in any case I'd suspect in terms of power draw efficiency the GTX470M is in a different class (or rather, the GTX480M defines its own class outside of any other chip).
Something else I'm curious about is the very very low gddr5 memory clock on all mobile chips. It was already that low on GTX480M but that chip wasn't really the right one to draw any conclusions. Is this due to power efficiency? Maybe it's possible to use only 1.35V instead of the nominal 1.5V when operating at such a low frequency? Or would the memory controller need higher core voltage to run at higher frequencies? I'm really wondering what effects that low memory clock has wrt power draw compared to AMD - basically nvidia needs a 192bit bus (hence more chips) to achieve similar bandwidth than AMD with a 128bit bus, using the same gddr5 technology.
Also, possibly related, AMD has pushed gddr5 down to even the slowest chip (cedar). OTOH it seems nvidia is using (well on the mobile front at least) ddr3 exclusively on GF108 (which is really more of a redwood competitor).
Space Giraffe
Newcomer
8 texture units at 700 MHz multiplies to 5.6 GTexel/s. It's right.
Umm no it's not. Where did you see 700Mhz... According to nvidia, GT415M for instance has 48 shader units, 1000Mhz shader clock. That gives us 8 texture units at 500Mhz - which should be 4GT/s, not 3GT/s as stated. Same for all other new mobile chips. I guess whoever updated the website calculated with same ratio as GTX480M/GF100...
5.6 at 700MHz = 8 TMUs
http://www.nvidia.co.uk/object/product-geforce-gt-420-oem-uk.html [specifications]
http://www.nvidia.co.uk/object/product-geforce-gt-420-oem-uk.html [specifications]
Space Giraffe
Newcomer
Oh for the 415M yeah that's wrong. But as Novum stated 6 TA/TF isn't possible. It's 8 TA/TF and for the GT420 which Neliz linked above it's measured correctly.
Ah right, it's correct for the desktop GT420. But it's wrong for all (new) mobile parts.
(btw nice memory on the GT420 - 2GB standard. I'm sure that's gonna help performance. Granted 2gbit ddr3 memory chips are dirt cheap still I'd have thought it would have saved a couple bucks for opting for 8 1gbit chips instead.)
Last edited by a moderator:
I guess he means that as the process matures and yields improve, clocks and/or power can improve over time. In this case due to the troubles of TSMC's 40nm process maybe even more so. Dont forget Cypress and Juniper are almost a year old now.
For example, Intel's processors improve in thermal characteristics and overclockability with stepping changes as the process matures. A famous example was the Q6600 G0 stepping or i7920 D0 stepping
2GB?? Even 1 GB is overkill for this class of cards
Wouldnt they have 4 chips? GF108 is rumoured to have a 128 bit memory interface which implies 4 memory chips. So 4 2Gbit chips would give 1GBAs to the power consumption, here are Anand's numbers for the GTX 460 http://www.anandtech.com/show/3810/nvidias-geforce-gtx-460-part-2-the-vendor-cards/8
Power consumption of reference 1 GB card increases by 9W at 825 Mhz and the 768MB card by 10W at 840 Mhz. He does not mention Shader clocks, and memory clocks stay at default. Interestingly the EVGA 768 MB Superclocked with clocks of 783/1526/3800 only consumes 4W more than the reference card
Last edited by a moderator:
Blazkowicz
Legend
28.8GB/s is a nice bandwith amount, this is the cheap and usable card I expected and wanted. I would enjoy a passive variant or one with a cheap dual slot cooler.
When I said it's going to help performance I didn't mean it for real2GB?? Even 1 GB is overkill for this class of cards
.But ddr3 memory (unlike gddr3/gddr5) is only available in 4, 8 and 16bit versions. Hence you always need 8 chips for 128bit.
It would be completely impossible for any GF108 based card (using the full 128bit interface and ddr3) to have less than 1GB, since ddr3 chips smaller than 1gbit do not even exist
.
Last edited by a moderator:
DavidGraham
Veteran
Gts 450 ..
The news are true , it competes with HD 5750 and replaces GTS 250 (AKA 9800GTX+!!) , it is slightly faster than HD 5750 (8~10%) .
It's half GF104 .
It has 32 Texture Units and 1 GPC .
It consumes 31w more power than HD 5750 at load and 6w at idle.
Temps are 57c at 70% fan , and 72c at default fan (FurMark Burn) ..
Even with 1 GPC , it still beats HD 5750 at heaven by 29% @ 1900x1200.
Oc'ed to 1.0GHz core easily , which pumps up Vantages GPU score by 2000 Points .
The news are true , it competes with HD 5750 and replaces GTS 250 (AKA 9800GTX+!!) , it is slightly faster than HD 5750 (8~10%) .
It's half GF104 .
It has 32 Texture Units and 1 GPC .
It consumes 31w more power than HD 5750 at load and 6w at idle.
Temps are 57c at 70% fan , and 72c at default fan (FurMark Burn) ..
Even with 1 GPC , it still beats HD 5750 at heaven by 29% @ 1900x1200.
Oc'ed to 1.0GHz core easily , which pumps up Vantages GPU score by 2000 Points .
DavidGraham
Veteran
Similar threads
- Deleted member 2197
- Architecture and Products
