AMD: R9xx Speculation
- Thread starter Lukfi
- Start date
1600 SP?
symmetry eh 64^2
Yeap. With fullrate FP16 and the 4D nerf don't expect the 69s to be 3DMark (Vantage or 11, they're both quite similar in placing) monsters.
I think the nVidia Confidential slide on games speaks wonders for the performance, though. Crysis/Metro etc actually show the 6970 > 580. And this could be a neutered card for all we know.
Yeap. With fullrate FP16 and the 4D nerf don't expect the 69s to be 3DMark (Vantage or 11, they're both quite similar in placing) monsters.
I think the nVidia Confidential slide on games speaks wonders for the performance, though. Crysis/Metro etc actually show the 6970 > 580. And this could be a neutered card for all we know.
If we use the average perf advantage 580 has over 570, 6970 should be faster than 580 in all of the games except for avp & batman
Yeap. With fullrate FP16 and the 4D nerf don't expect the 69s to be 3DMark (Vantage or 11, they're both quite similar in placing) monsters.
I think the nVidia Confidential slide on games speaks wonders for the performance, though. Crysis/Metro etc actually show the 6970 > 580. And this could be a neutered card for all we know.
We still have to know if that is with its "Turbo Mode" on or off
If with it ON it consumes 240W, from 190W, there is probably a nice clock speed bump. Remember it runs at 880Mhz, while GTX580 runs at 772Mhz, with quite some room for OC. All in all, it might not be that far from Fermi 1.1 perf/watt, while having better perf/mm2.
I guess that's a possibility with the launch driver.
It would also only be less than 15% faster than HD5870, which is a bit disappointing.
I think GPU-Z doesn't know yet about VLIW5 and VLIW4, and it possibly misreads number of simds. So it reads 20 simds, calculates tmu count from that and miscalculates sp number from the same bogus number.
PSU-failure
Newcomer
No, it doesn't.
Consider the "1600" GPU-Z gives on that board, it means it has 1280SP if based on NI, or 1600 if based on EG.
Either it's Cypress in disguise, severely crippled Cayman or Cayman dramatically reduced raw throughput in exchange for efficiency (theorically, with 20 VLIW4 SIMDs, it could be up to 42% faster than a GTX580 or as much as 0% faster than an HD6870).
Or in the same ballpark as the 5870 2gb. So i hope all the info so far is fake ...
http://www.hardocp.com/article/2010/11/09/nvidia_geforce_gtx_580_video_card_review/3
http://www.hardocp.com/article/2010/11/09/nvidia_geforce_gtx_580_video_card_review/3
I'm playing Lotro on eyefinity, just 5040x1050 though. What card you using - even my 5770 ran it well with 0xAA (obviously) and mid-high settings.
But yeah I get your point. I signed up for TOR beta testing, giving eyefinity as my reason why I should get a spot
I don't think it will have eyefinity support to start with unfortunately. If AMD had any sense it would, but...well you know what they are like lol.
radeon 5850 but at 5760x1080. I think its more the framebuffer than anything in that game.
But of course I'm just playing that for fun right now since its fun. I can see raiding in TOR getting hairy with eyefinity
Cayman isn't "severely crippled" if its SP count is similar to Cypress and the VLIW4 rumors are true.
It complies with 80 TMUs hence 20 SIMDs like Cypress. That means it is either a lazy fake or if GPU-Z really is (somewhat correctly) reading the amount of enabled units, it simply messes up with the hardcoded 5 SPs per VLIW unit and does not take the new VLIW4 architecture into account (*). But the latter possibility would also mean that the benchmark was done with just 1280 active SPs.
(*):
In AMD stream for instance you get the number of SIMD engines and the wavefront size only. You have to put in knowledge about the VLIW structure to arrive at the correct number of SPs.
PSU-failure
Newcomer
If its performance isn't at least ~40% better than Cypress, it'll have been crippled, whatever the reason.
On the other hand, 20 VLIW4 SIMDs is enough to be even higher than that, so either it's crippled (low specs, low perf) or more efficient (low spec, high perf).
Limiting the SIMD count could allow for a considerably higher transistor/area budget elsewhere, so even if the said "leak" is suspect, it's not totally impossible.
Edit: although what I wrote still stands true in general, we have to remember the "More than HD5870" on the "Texture Filtering Units" line, which implies more than 20 SIMDs.
No. When looking at power consumption of e.g. HD5970, it will show, that one cannot just multiply the power consumption of one complete single card (OK the RV870 on the HD5970 are slightly downclocked, and most likely selected for low power consumption.
Yes. Correct. Power consumption of the HD5970 (2x Cypress XT) is appr. 216W. Cypress XT Crossfire 291W respectively (1x Cypress XT appr. 158W)
http://ht4u.net/reviews/2009/amd_ati_radeon_hd_5970/index11.php
That shows, that you cannot just double the power consumption.
Edit:
That shows, that the second Cypress GPU on Hemlock just added 37% of a complete Cypress XT card. Assuming 50% in case of Cayman Pro/XT at 190W, you would end up at 285W (Depending on chip selection and down-clocking).
Given that Cayman and Cypress are produced at the same process node, it's not feasible for Cayman to be that much faster than Cypress without being significantly larger. AMD (ATi) doesn't create bloated GPUs designed to win the single-GPU performance crown, so how likely is this to occur?
Similar threads
