NVIDIA GF100 & Friends speculation
- Thread starter Arty
- Start date
The ati one is 1680*1050 the nvidia 1920*1200. The textures and details are quite different.
"Despite the fact that Nvidia will have a massive volume launch with thousands of cards"
It would be moronic to say something like that without being sarcastic ... he never once really made any critical note about the credibility of the release dates while pushing them onwards week by week on his site though ... so why should I believe now he is not just being a moron? That said, I think it's more about plausible deniability ... this way he can say to he is pushing the message and still pretend to himself he is placing a critical note (which goes right over the top of the head of his remaining readership I'd say). So basically he's chicken and this is as far as he is willing to go.
Last edited by a moderator:
Super-sampling is not available in DX10+ mode. And I used the default AF setting to run the test, which is lower than the 16xAF, seen in the GF100 shot.
On a different note, if you want to see how bad is the AF under-sampling on HD5000 hardware, check out TrackMania Nations game -- all the high-frequency and uniform textures, used for most of the surfaces, are being "sliced" along the filtering stages.
On a different note, if you want to see how bad is the AF under-sampling on HD5000 hardware, check out TrackMania Nations game -- all the high-frequency and uniform textures, used for most of the surfaces, are being "sliced" along the filtering stages.
Last edited by a moderator:
thanks, just what I wanted to know.
Mintmaster
Veteran
Why would GTX470 have considerably less bandwidth? Did I miss something? Is the memory bus severly crippled and clocked low?
Secondly, why on earth would a comparison with HD5850 seem better? This isn't a bandwidth test. ATI could easily get more FPS with the same bandwidth if it made a bigger chip. A 4870 is never 80% faster than a 4850.
I believe the GDDR5 clocked at 800Mhz supposedly, compared to 1200Mhz on the 5870?
I can't go into details unfortunately but one building block that works very well is atomics, which are becoming increasingly important in compute workloads (including graphics). Local atomics (shared memory) in particular are *really* fast on Cypress, attaining rates like ~280 uncontended atomics/clock across the chip even on midrange hardware. This opens up a lot of algorithmic options and makes some fairly complex stuff viable.Care to elaborate on what exactly it is you're doing, and what the respective performance numbers are...? *curious*
I expect Fermi will improve their atomics a lot too (they said they have) so it will be interesting to see how they compare.
Back to the original point though, I have yet to run into significant weaknesses of Cypress for forward-looking workloads and in a number of places it has really impressed me. Obviously a lot of it is relative so maybe Fermi will make some stuff on Cypress look weak in comparison, but we'll see.
Last edited by a moderator:
Mintmaster
Veteran
If it's one game, there's a good chance that a negative LOD bias was used. I've seen it from racing games before, especially ones destined for the console. I don't know if NVidia and ATI handle AF with -ve bias the same way.
Jawed
Legend
Rumours suggest it's crippled with 800MHz clock and 320-bit width.
HD5850 (128GB/s) and GTX470 (~128GB/s) both have about the same bandwidth as HD4890 (124.8GB/s).
DS is likely a bandwidth test, for what it's worth, when close-up to the dragon in Heaven (and other parts of the benchmark, it seems). Jury's still out, though.
Feel free to pitch in with ideas why this benchmark is not significantly setup limited on ATI:
http://forum.beyond3d.com/showthread.php?p=1397787#post1397787
It may not be a simple "available bandwidth" test, but more like an "efficient use of available bandwidth" test.
Jawed
Mintmaster
Veteran
Is it 320-bit for the GTX470?
Mintmaster
Veteran
So what's the reason for using 800 MHz GDDR5? They couldn't clock the memory controller high enough? You'd think that with experience from GT215 they could at least get this right.
This is all really shocking to me. I couldn't believe Charlie's contention that the GTX470 would be so slow, but it looks like he may be right. I still have my doubt, though. It's just unfathomable to me that NVidia could screw up so much.
This is all really shocking to me. I couldn't believe Charlie's contention that the GTX470 would be so slow, but it looks like he may be right. I still have my doubt, though. It's just unfathomable to me that NVidia could screw up so much.
Jawed
Legend
Power, I suspect. GDDR5 I/O on a GPU uses a fair bit of power. And the rumours suggest low clock and commensurately low-voltage.
Jawed
Mintmaster
Veteran
That's crazyness. ATI is getting more bandwidth out of a 256-bit bus than NVidia does out of a 320-bit bus due to power constraints? Wow.
Does that mean the GTX480 is even more boned?
The low voltage was in relation to the memory device voltages, not the GPU core memory memory. The device voltage reduction would save little power overall.
The GTX480 is 6pin+8pin while the 470gtx is just 2*6pin for some reasons. But it seems the gtx480 will be much higher clocked with the masive cooler. Thats 75W difference for quite similar gpu.
Jawed
Legend
I presume the bulk of the memory controllers lie in the Fermi die shot's central stripe. It seems to me that there are 3 "similar" regions at the ends of the stripe. i.e. one half of the stripe has a pair of these regions and the other half has just one.
From the stripe it's a fair way to the majority of the I/O pads.
RV770 appears to have its memory controllers at the periphery of the chip, distributed into locations that are relatively close to sections of the GDDR5 I/O pads.
Does the difference in distances have much effect on power consumption?
Jawed
From the stripe it's a fair way to the majority of the I/O pads.
RV770 appears to have its memory controllers at the periphery of the chip, distributed into locations that are relatively close to sections of the GDDR5 I/O pads.
Does the difference in distances have much effect on power consumption?
Jawed
Silent_Buddha
Legend
So you really think GTX 470 is going to launch at half the price of GTX 480? After all pretty much all rumors for GTX 480 point to 600-679 USD launch price.
Regards,
SB
Similar threads
- Deleted member 2197
- Architecture and Products
