Speculations On R400
- Thread starter David G.
- Start date
Steve Dave Part Deux
Newcomer
KFC's chickens have 4 wings - R200 has 4 rendering pipes.
They have eyes everywhere....
They have eyes everywhere....
There's lots of other interesting things as well. Transistor count is in no way inversely proportional to MHz - just because something is ram-packed full of transistors doesn't necessarily mean it's going to run slower.
In fact, the reverse is often true, or so it transpired during my very brief stint as a 3D hardware engineer. Some types of cell may use more transistors and/or silicon area to operate at higher frequencies, for reasons of stability, noise reduction, capacitance, inductance, phase of the moon, etc. etc... It's all a bit of a black art.
The most important thing is to target the design from day one at the right clock speed for the process you're using, the chip cost, and the target market. If you simulate at 300MHz, pay the extra silicon area for more complex noise insulation and routing, and you have good design people, you will get 300+MHz out and with further process optimisation you may push it up evem further. If you simulate at 200MHz, then you're not likely to achieve 300MHz.
I would suspect that is what happened with Parhelia - at the time when they started the project they made a decision to simulate at one particular speed givem the target market - then the market overtook it somewhat.
In fact, the reverse is often true, or so it transpired during my very brief stint as a 3D hardware engineer. Some types of cell may use more transistors and/or silicon area to operate at higher frequencies, for reasons of stability, noise reduction, capacitance, inductance, phase of the moon, etc. etc... It's all a bit of a black art.
The most important thing is to target the design from day one at the right clock speed for the process you're using, the chip cost, and the target market. If you simulate at 300MHz, pay the extra silicon area for more complex noise insulation and routing, and you have good design people, you will get 300+MHz out and with further process optimisation you may push it up evem further. If you simulate at 200MHz, then you're not likely to achieve 300MHz.
I would suspect that is what happened with Parhelia - at the time when they started the project they made a decision to simulate at one particular speed givem the target market - then the market overtook it somewhat.
NTD said:Geeforcer said:
Yeah, me too, I find it strange that a 55M P4 runs at 2,6GHz while these measly 40-60M GPU's run at 300MHz.
Yeah ... but you probably forget that that's 0.13 micron technology in a perfect implementation with BIGGER cooling devices on it than what video cards usualy use along with the fact that a whole mobo is filled up with capacitors to make the circuit stable .
Geeforcer said:
Well, from what I know, the RV250 actually yields at ~325MHz, so that is in fact higher. It's of little significance though... R300 is obviously massively improved from a design point of view, compared to the RV250 who's architecture is based on one that was originally designed to run at 250MHz. We all know that the original at-speed target for R300 was 350MHz, although no doubt the part has grown in complexity since then.
Don't forget that RV250 can operate at very high clockspeeds, passively cooled, while R300 will probably sport a pretty hefty HS/fan. This is of course intimately linked with power consumption; the R300 consumes MUCH more power at-speed than the RV250 (it could be as much as 5 times more I believe).
MuFu.
That information was correct; yes, A12 is capable of over 50MHz more than A11 for reasons I have already explained. So, err... nothing to argue about, I hope.
Are you suggesting that R300 does not use a 256-bit memory interface? That's not true. It's a 256-bit bus with a 64-bit x 4 channel organisation.
MuFu.
Randell said:
That's almost correct; the first revision was only capable of 250MHz as the 00A PCB on which it was tested did not have a external power connector to source more juice (the ASIC consumed a full 30W, putting the total consumption at about 54W). The AGP spec violation was simply not anticipated. The problem lay in the A11's memory interface which was drawing a lot of power and causing instability at high clockspeeds. In the A12 stepping, this bug is not present, so subsequently the ASICs can generally be clocked up to about 300MHz, at which point they need an external power connection to go beyond the AGP power spec. This shouldn't really be necessary on high quality mobos (Intel mobos can source about 55W across the AGP, for example), but is understandably a good move to ensure reliable operation on cheaper boards. The AIB/OEM R300 cards might not have the external connector to save production costs and be clocked at <300MHz, although it may be possible to flash the BIOS, add an old FDD power connector and "upgrade" to a BBA card.
MuFu.
Diminishing returns? Requires faster, more expensive RAM (thus obfuscating the low end element)? Requires active cooling (again, pushing prices up)?
There could be all manner of reasons. Why does NVIDIA cripple the MX in its 420 specification? Could it not really do more than the 460 specification? Low end parts have different needs & targets most likely.
Similar threads
