Nah :biggrin:, it became faster over time.Fermi was a new architecture for NV so it took them some time to harness it's full potential.
When you say it aged much better, what do you mean? Is it like how the silicon resists flexing or something due to heat cycling over time?
When you say it aged much better, what do you mean? Is it like how the silicon resists flexing or something due to heat cycling over time?
Are you still touting bandwidth-limited benchmarks under extreme OC? You are funny.
They aren't all that different. Still GCN.
Pretty sure he has that post copied to his desktop in a .txt file.
read the hardwarecanucks review
http://www.hardwarecanucks.com/foru...s/63051-evga-gtx-780-classified-review-8.html
By moving to a core speed of 1.32GHz, we were able to all but eliminate the ultra high temperatures and ensure a continual clock speed, regardless of the situation. Temperatures normalized as well. After 20 minutes of constant load (see the image above), the core frequency didn’t budge one iota, though we were still hitting the Classified’s Voltage Limit.
"Did we forget to mention memory? Well, that topped out at 7426MHz before the GDDR5’s error correction stepped in, which is nothing short of incredible."
http://www.hardwarecanucks.com/foru...ews/62594-galaxy-gtx-770-gc-4gb-review-8.html
"Unfortunately, Galaxy doesn’t seem to offer their own home-grown overclocking software anymore so MSI’s AfterBurner was used for these tests. With it, we were able to hit a constant Boost clock of 1313MHz which is within spitting distance of the results achieved with samples from MSI and Gigabyte. Memory speeds weren’t quite as flexible with frequencies topping out at 7404MHz before error correction kicked in. "
so you see GTX 780 was running 1.32 Ghz core and 7.4 Ghz memory on a 384 bit memory bus with 48 ROPS while GTX 770 was running 1.31 Ghz core and 7.4 Ghz on a 256 bit memory with 32 ROPs.
same core clocks and memory clocks on both chips. but 50% more cores, 50% more bandwidth, 50% more ROPs with GTX 780. but only 25 - 26% more performance in Crysis 3. so stop giving excuses.
the reason is GTX 770 has 4 GPC, 4 raster engines for 1536 cc while GTX 780 has the same 4 GPC, 4 raster engines for 2304 cc. the perf/cc on GTX 780 is lower than GTX 770 due to the same amount of front end resources feeding 50% more cores. The GK104 is a extremely well balanced chip. :biggrin:
GK110 being "nearly" 30% bigger (actually less) means that Hawaii will be over 423.8 mm^2, no less.
This DOES make it the biggest GPU silicon AMD (or ATI) has ever made, beating HD 2900XT at 420mm^2.
This also means that it will be at least 20% bigger than Tahiti's 352mm^2.
It seems that:
15-25% more shaders/TMUs
and
48 ROPs
(or even more shaders/TMUs while still having only 32 ROPs which might hurt @ higher resolutions with AA)
is just about what AMD can do with this increase in silicon die size, without some other magical optimizations or expansions (remember, HD 4890 was actually LARGER than HD 4870 despite having practically identical GPU layout).
It certainly does seem like 30% increase in performance is not out of reach, as long as AMD is willing to allow a 300W TDP headroom on this single-GPU card.
However, I'd expect a modest TDP ceiling, and a 20% increase out of the box rather than 30%... because going anywhere close to 300W is just pushing it too hard. There are too many risks (longevity, warranty, noise, notoriety of heat output, etc..) for this to become a practically viable option.
25% increase at 265W might, just might be possible. AMD did it really nicely with Bonaire, after all, so I think they can do it (hint, that is, beat GTX 780).
.. I'm just hoping that AMD will no longer stubbornly stick to 32 ROPs, and just move onto a symmetrical ratio of 48 ROPs to 384-bit bus (so that there's full access to the RBEs rather than limited crossbar access). That's why Tahiti's 384-bit bus seems rather lackluster, like as if the GPU had less than 320-bit bus access in the real world. Just guessing - there seems to be about 5% overall performance overhead that could be regained due to having a straight 48 ROP design alone. Yet it would probably also translate to 5% increase in silicon die size as ROPs are notoriously difficult to shrink with each new fab process nodes, from what I have been told. After all, it was a logical move by AMD to do that with Tahiti early on 28nm process in order to get it out ASAP while keeping the size and power consumption down. I have no idea on exactly how much the ROPs affect power consumption though (up to 5 percent for this 50% increase in ROPs??).. but if AMD manages all of this well (48 ROPs, +25% performance, say 15+% reduction in power consumption-to-performance ratio, etc..) then I'd be amazed! Or more!!!
Still not enough for perfect scaling at these clock speeds. Just keep ignoring the other scaling review I had posted a while ago:
http://www.pcgameshardware.de/Grafi...force-GTX-780-Taktskalierung-im-Test-1082208/
And their normal GTX 780 review disagrees with you:
http://www.pcgameshardware.de/Gefor.../Tests/Geforce-GTX-780-Test-Review-1070700/4/
GTX 770 review:
http://www.pcgameshardware.de/Gefor.../Tests/Geforce-GTX-770-Test-Review-1071496/3/
Clock for clock the GTX 780 is 49% faster than the GTX 680 in Crysis 3. This continues with OC if you increase memory bandwidth accordingly as shown above. There is nothing you can say against those reviews. PCGH is the very best when it comes to documented reviews. They document absolutely everything in minute detail, step by step and further more they eliminate the uncertainty of Boost by disabling it and setting clocks to a fixed value.
To sum up:
GTX 770@1.085/3.505 = 34.9 fps
GTX 780@1.046/3.506 = 51.5 fps (47.6% more despite slightly lower core clocks -> perfect scaling)
So who is right? I trust PCGH test methodology way more than HWC and their result also is way more plausible given the specs of the cards involved. GK110 scales perfectly, deal with it.
I thought gk110 actually was somewhat bad when it came to gaming perf/mm2 which kind of left the door open for AMD
raghu, do you doubt that the gk110 performance per mm^2 would go up significantly if with everything unlocked with a higher target tdp? See performance per mm^2 really is a poor metric when you do not take in performance per watt.
show me a comparison of GTX 780 and GTX 770 in a game using the same benchmark settings at 1.3 Ghz and lets see the diff. pull out a pcgh game benchmark with both cards at 1.3 Ghz and lets see the gap.
btw your first chart shows GTX 780 at 42.0 fps and the second chart shows GTX 780 at 44.9. both at stock. so perf has gone up due to driver version changes. the first review is May 23 with 320.18 drivers and the second one is Aug 10 with 326.41 drivers. you see right there is a mistake.
you are so wrong with your 50% perf increase at same clocks. the actual number is closer to 30%. anyway keep believing what you want to :biggrin:
1. Couple of points RS, though 480 was ~15% faster than 5870 at launch it aged much better compared to 5870.Look at any recent review and you will see most of the time 480 is faring much better than 5870.It also had 512MB of extra vram lets not forget that.
2.Same way 580 aged much better than 6970 though this time the vram advantage went to AMD.Another important thing is overclocking, Fermi scaled way better than 69xx series.
Why do you insist on this clock speed, that makes no sense at all. At 1+ GHz GK110 will always be bandwidth-bottlenecked to some degree and it gets worse the higher the core clock is. Thus a comparison between GK104 and GK110 in regard to scaling with available cores simply makes no sense at such high clock speeds. I'm talking architecture here, not specific SKUs.
Even if the GTX 770 performance has gone up accordingly, that still puts the gain at close to 50%.
you are argumentative. I repeat you are yet to show a single benchmark of GTX 770 and GTX 780 at 1.3 Ghz clocks in a game at same settings and prove the 50% perf diff. :whiste:No, I'm right - but only up to a certain frequency because of bandwidth limitations. Above that frequency, GK110 gains are smaller, but that is definitely not a problem of the frontend or the architecture, but it's a matter of bandwidth. Clock for clock comparisons only make sense if you are aware of the boundary conditions. Do you honestly believe that a badly balanced OC will net perfect scaling? I'm sorry to say this so bluntly, but that is just idiotic.
For the gk110, i do not think nvidia is near that point. They have a little room left, there obviously is more to squeeze out of the chip. They havent even released a full gk110 geforce variant.
Ok but now watch this. By Feb 2011, you could get 6950 unlocked for only $230 + $250 savings from not buying 580 provided 85% of the performance of the 580 for ~2 years before 7970 launched. Sell 6950 for $175 + $250 savings from not buying the 580 = $549 - $375 resale + savings ~ $125 to upgrade to 7970.
Once 7970 was overclocked it gave 50%+ more performance over 580 oc. Therefore, for the majority of competition between 6950 @ 6970 vs. 580, 6970 provided about 85% of the performance and then when it came to newer games, the amount of $ saved from not getting 580 + reinvesting it into a future 28nm GPU blew 580 away.
Therefore, the point for 480/580's longevity is not really that important imo since upgrading more often is the more preferred route. The same thing will happen with 7970GE vs. 780/R9 280. By the time next gen PC games come out, all 3 of these cards will be too slow but right now 7970GE has little trouble with most games since they are still PS360 console ports. The difference now is even more in favour of the 1Ghz 7970 since they are dropping to $280 which represents a $370-400 savings over a 780! You know how much $370-400 savings + 7970's resale value will get you in GPU power in the next 2 years? I bet 40-50% faster than 780 by the time next gen PC games hit and we really need the extra GPU power.
And now people are talking about perf/mm^2? :awe:
I showed the hwc benchmark where the GTX 780 was running at same core clock and memory clock as GTX 770. so it had 50% higher bandwidth with 50% more cores and 50% more ROPs. the diff was still below 30%.if you say the GTX 780 was bandwidth bottlenecked then so was GTX 770.
you took a stock GTX 770 with stock power target against a GTX 780 OC with power target maxed out. the stock GTX 770 could be throttling for all you know. thats such a bad comparison. max out power target and voltage. push to max clocks. if clocks are similar draw comparisons. btw hwc even verified that the card is not throttling.
you are argumentative. I repeat you are yet to show a single benchmark of GTX 770 and GTX 780 at 1.3 Ghz clocks in a game at same settings and prove the 50% perf diff. :whiste:
Don't you think that will depend on each Individual game? here is one
780 is more than 50% faster here(both are at stock). This is also a TWIMTBP title so I believe both the devs and NV knew what they were doing.
the question is how many games are showing such a huge difference. the answer is very few. In the majority of games the GTX 780 is 25 - 30% faster than GTX 770 on a clock for clock basis. so even if you account for extreme cases the avg is nowhere close to 50%.
Nah :biggrin:, it became faster over time.Fermi was a new architecture for NV so it took them some time to harness it's full potential.
Thank god non of AMD cards are compute focused, that would make them worthless for games!I believe Raghu's assertion is valid as Titan has to pay it's DP tax.GK110 is too much compute focused.
If NV need more die space for the same performance then it's a point to talk about.If AMD needs more power for the same performance then it's a point to talk about.
But perf/mm^2? Lol. That is used by people who try to defend AMD if they are not able to beat the GTX780.
Aha.. That is a good thing. But when AMD did this with HD7000 they failed to release new arch and their driver team sux... logic!
Thank god non of AMD cards are compute focused, that would make them worthless for games!
If NV need more die space for the same performance then it's a point to talk about.
But perf/W Lol. That is used by people who try do defend NV if they are not able to beat 7970 in compute.