No such thing as a free lunch red....
Higher clockspeed or higher memory bandwidth -or a juggled combination to keep with boards specification. Pretty much SOP I would think for any AIB's custom boards. Not sure how that affects those cards that have
proven to be poor/average overclockers. Is pouring more voltage into the GPU the answer ? And if it is, then surely you're either raising board wattage or throtting kicks in earlier to cap the power use ? Again, I'd ask if there is a third option.
It might be interesting to see if there are performance differences in the Jan 9th launch. i have heard from a normally reliable source that the first round of review cards may have been throttling back a bit prematurely.
Seems somewhat mutually exclusive to this thread from my PoV, if throtting limits are the defining statement. AFAIA, a less aggressive throtting should allow better performance on games/apps where the GPU/VRM is being throttled yet is still within TDP specification (i.e. when either the GPU or vRAM is being under utilised while the other is being constrained by throttling), I'm not entirely sure how the throttling limit is supposed to affect yield variances, unless the throttling can be completely relaxed in the face of increased core voltage*.
It might be interesting to see if there are performance differences in the Jan 9th launch. i have heard from a normally reliable source that the first round of review cards may have been throttling back a bit prematurely.
You'll still need to compare apples-to-apples. Stock clocks-for-stock clocks at paper launch and product launch. As for...
or if they clock them up another 100Mhz or so.
...as I said before, I would agree that it is very possible...but it must come at the expense of either lower yield (to eliminate those poor samples which accounted for ~50% of review cards), or increasing voltage -assuming this would gain the required speed, and that is not the case in all instances from what I'm seeing.
What I'm seeing here is soft modded voltage, and obviously, cherry picked GPU's (Sapphire) as was the case with previous Atomic/Toxic HD 38xx and 48xx and OC'ed cards in general. Good to see, but the parameters that guide other OC'ed components still apply here.
1. Fan speed at 100% on a reference blower fan. Not a 24/7 solution IMO.
2. Heavily OC'ed cards are still likely going to require a waterblock or an (expensive) third-party cooler such as Arctic's xtreme plus - you could conceivably get HD 6990 performance at HD 6990 pricing -there is no way on earth that a 1335M core uses any reference component aside from top-binned GPU*
3. Performance never scales linearly to clock rate. The 32 ROP count is likely to cap performance gains, even if you profile the OC for both high core/low mem and average core/high mem. If this wasn't the case you would have the GTX 560Ti leaving the 580 in the dust....
which doesn't seem to be the case
*Since the boards in their present guise are pulling ~277w (average) @ ~1.15v (VR Zone's example) that gives you 229 Amps through 5+1 phases of
Coiltronics 1007R3-R15 @ 61A per phase (366A max).
Assuming that the boards TDP isn't to rise (i.e. throtting profile makes efficiency gains without blowing the power budget) then VR Zones example would be 277 x 1.25v = 346.25A (lumping GPU and vRAM together at the same numbers-so not an exact science) ...which is
5% off the boards design threshhold, which kind of leaves raising the power or increasening the voltage regulation which would require a complete deviation from the reference model's power deleivery/trace layout/PCB layers....and pricing. In effect what happens with every limited edition (MSI Lightning, Gigabyte SOC, Asus DCII, PowerColor PCS+ et al)
thoughts? anything I've missed ?
EDIT:
Preview of "Sea Island " capabilities perhaps?
That would be "Canary Islands" I think, and yes, probably. AMD are locked into GCN for the foreseeable future, so it isn't unreasonable that the next iteration of cards will be a more refined design on a more mature process node.
Good god! ...look at the bandwidth!
Cool....the GK110 with 320 GB/sec should kick a$s then (512-bit x ~1250MHz memory clock). Haven't really ever put much stock in the bandwidth number generally:
GTX 285.........512 bit....2484 memory..........158.98 GB/sec
HD 5870.........256 bit....4800 memory..........153.6 GB/sec