Parhelia projected clockspeed

[Jon P. Inghram]

Where were all these people with all the inside connections when the G550 was still a mystery? Amazing how all the super secret inside leaks are coming from people who joined up here the week the Parhelia became "real".

[ben2207]

Errr I've been around over 2 years. Just haven't posted much. And I was briefed on G550 last year (Parhelia was expected last year btw (meeting with Matrox E3 2001)

[Wombat]

Well, at least now I know how good his sources are.

[DuRaNgO]

Keep in mind, mhz is not the end all, especially when you have a new architecture

[Joel]

Even if his information is correct, which I don't think it is, there still should not be a problem.

Matrox hadn't finalized clock rates at press time, citing the need to determine what clock rates they'll be able to hit in quantity using UMC's 0.15-micron process, but we estimate that clock rates will be somewhere between 200-250MHz. Giving the benefit of the doubt, Parhelia's speeds n' feeds (as seen in the above table) based on a 250MHz clock are certainly competitive: 1Gpixel/sec pixel fill rate and 4Gtexels/sec of bilinear-filtered texture fill rate. A 200MHz clock would yield 800Mpixels/sec of pixel fill-rate and 3.2Gtexels/sec of bilinear-filtered texture fill rate. Parhelia's 256-bit wide DDR memory interface will help keep all of those texturing units churning, which will be important not only for straight texture reads, but also for enabling features like anisotropic filtering and fragment antialiasing, which we'll get to in a bit.

Parhelia's Memory Controller
Since we're on the subject, and because it's so crucial to hitting optimal fill-rates, let's take a quick look at Parhelia's memory controller. Its 256-bit-wide interface (256 bit wide read bus and separate 256-bit wide write bus), coupled with 325MHz (650MHz data rate) DDR that would yield a staggering peak bandwidth of 20.8GB/sec (650MHz * 32 byte wide interface), and even if you factor in DDR memory is about 75% clock-efficient, you're still looking at over 15GB/sec of usable bandwidth. Ignoring cache efficiencies, in order to feed Parhelia's 16 texturing units without a stall, and assuming a 250MHz core clock, you'd need 64GB/sec of texture memory read bandwidth to maintain 4Gtexel/sec of bilinearly filtered texels, which Parhelia can supposedly achieve. So obviously, Parhelia relies heavily upon high texel cache hit rates in order to minimize memory touches. (See more details below).

Obviously, there are many other chores to tend to that require memory bandwidth including: writing to the color and depth buffers, vertex and pixel shader processing, feeding the RAMDAC, etc. Having this nearly 2X increase in usable frame buffer bandwidth will be crucial if Parhelia is going to hit its marks. It's also worth noticing that snaking its way throughout Parhelia's core is a 512-bit-wide data path, which, assuming a 250MHz core clock frequency, would yield internal data rates of 16GB/sec. Finally, note that the chip supports up to 256MB of frame buffer memory.

http://www.extremetech.com/article/0,3396,s=1017&a=26865,00.asp

Joel

[GT98]

A little late in the game for this but had to do it ......

[CaineTanathos]

the 220 mhz would proberly be the speed of the alpha board

[ben2207]

Joel , notice Extremetech's article mentions clockspeeds in the range I mentioned ....

[Kruzin]

There are likely to be different clock speeds on different versions of the card. It's possible that 220 is going to be one of the available speeds...

[Joel]

Joel , notice Extremetech's article mentions clockspeeds in the range I mentioned ....

All I was trying to point out is that even if it is clocked at 220 it will still be a good performing card and nothing to get too upset about.

Joel

[ben2207]

I never said anything to the contrary.

[blacklotus]

Let's not play the stupid Intel games, MURCers

What's with all these obsessions with Clockspeeds aka Intel and Nvidiots? I thought MURCers should be better than them.

I don't care if the Parhelia is running at 10 Hz, as long as it gives me the Image quality exceeding those of G400 Max and frame rates of Quake 3 on par with R8500 at a affordable price point.

And if Parhelia can outdo a GF4 at a lower clockspeed, all the better, I say. It goes to show the Parhelia architecture is better than the GF4, and there will be more headroom to grow when the clockspeeds starting going upwards.