iXBT's claims of the G450

[vanguardian1]

"We've got the confirmation of the earlier posted info about Matrox's upcoming G450 graphics chip. It will be G400-like core based chip produced with .18 micron process. It will use 64-bit memory bus and G450 based cards will be equipped with 166MHz (333MHz effective) DDR SGRAM."

Ok, why the **ll would they down the 256-bus to 64-bit (I know, it's a dual 128 bus) just for DDR sgram? Maybe someone else here can make more sense of this...either complete BS or they're are some details that I don't understand somewhere....

btw, here's the url:
http://www.ixbt-labs.com/news.shtml#955997200

[AJ]

Well, I guess they are trying to reduce costs. They will have far fewer traces on the board with this setup (if it's true) and get more chips per wafer with the .18um process (assuming good yeilds). The 166MHz DDR DRAM will cost a pretty penny though (but not so much 6 months from now). So when the G800 comes out, this becomes Matrox's "value" card.

Based on the data from NV10 benchmarks, they should be able to clock the g450 up to 200MHz and still have plenty of bandwidth (assuming it still is a single dual textured pipe like the g400). If the process will allow it, they could even pump the clock up to 250Mhz. Then the G450 would have geforce level multitexture fill rates (1/2 the number of pipelines, twice the clock rate) - not bad if priced right!

-AJ

[Nuno]

The G400 256 bit bus is internal, not external. I think it has a regular 128 bit memory bus.
But that doesn´t make sense, a 64 bit 333 Mhz memory bus should be slower than a 128 bit 200 Mhz.

[AJ]

DOH! Good point. Damn, well from the info I've seen, the G400 had plenty of memory bandwidth @ 3.2GB/s (the core was the limiting factor), but this move will drop the bandwidth down to 2.6GB/s

And you are right on the DIB: http://www.matrox.com/mga/press_room...df/dualbus.pdf

Hopefully the meaning of this is that there will be a "cheaper" g450 with 16MB of mem at 64bits and a "Max" version with 32MBx128 bits. Some OEMs (Nvidia) have shown that having a lower cost OEM part is really important (The bus width crippled TNT m64 being the example here) - maybe that's what Matrox is after.

If the bus width really is 64bits for all versions, then It seems to me that clocking the core over 166Mhz wouldn't do much good.

Of course, the efficiency of the core may be much higher, so the effective framerates for various games may go up significantly anyway. Plus they may have optimized the core for better OpenGL performance or reduced the framerate drop off for EMBM.

This is a bummer to me. The g800 rumors say twice the fillrate of a g450. Well, twice 200/400 (MP/MT per sec) isn't a good as twice (166/333). Rats, it sure will be nice when the real specs are released!

-AJ

[skinner]

Maybe because of this:Fujitsu has begun sampling it's first 64 Megabit Fast Cycle RAM chips. This new FCRAM is targeted specifically towards graphics applications. FCRAM is expected to double the performance currently seen by DDR SDRAM. One nice feature is that FCRAM can interface directly to a SDRAM interface which should make the RAM that more attractive http://www.fujitsu.co.jp/en/news/2000/01/18-2.html

[dneal]

A little detail has been left out of these conversations.

The rumor bin said it was FCRAM (Fast Cycle), not plain DDR DRAM.

Take 266 MHz DDR for example. On a 64-bit bus, it has an effective bandwidth of 897MB/s. FCRAM has an effective bandwidth of 1165MB/s under the same conditions.

What this would mean for 200MHz 128-bit SDRAM versus 300MHz 64-bit FCRAM, I don't know.

But FCRAM is more advanced, it's most notable feature is lessened latency over SDR DRAM and DDR DRAM. And we all know, lower latency is good :P

It does seem odd that Matrox would cripple the G450 after putting effort into improving it at all. Plop a G450 chip on a streamlined G400 board, with its 128bit bus and some cheapo SDRAM, that would seem a more economical way to go.............

Question is, how much would G450 benefit from a 128-bit FCRAM interface at 300MHz?


FCRAM Info:

http://www.toshiba.com/taec/componen...P_memory.shtml


------------------
ABIT BF6, Pentium III SL35D 450MHz -> 630MHz, 192MB PC100 SDRAM @ 140MHz, Toshiba 6X DVD,
Matrox Millenium G400 DH @ 160/200, Creative SBLive Value, 3Com Fast Etherlink XL PCI
Supermicro SC701A ATX 300watt TurboCool PC Power & Cooling PS, Panasonic Panasync S17

[Maggi]

'lo dudes ...

maybe I completely miss the boat, but:

The G400 uses a 256bit dualbus for data transfer, resulting in 8 x 32bit transfer per clockcycle.

Isn't its onboard gfx RAM also 32bit deep ?

Wouldn't be a step to 64bit RAM enable twice the witdh of our current RAM and hence give us more bandwidth, performance etc. ???

...

Just a thought though.

Cheerio,
Maggi

<FONT SIZE=1>PS: I'm really not too deep into this technological stuff so if there's any nonsense posted in the above, feel free to correct me ... <FONT SIZE=2>

[Proton]

I think the solution to this puzzle can be found at the MURC newssection ("Flight Of The Condor"):"G450 will use 166MHz memory clock (in case of FC DDR 200MHz is feasible) and uses 64bit dual memory bus." If this is true, it would make perfect sense to me, because the G400 has an internal 256bit dual bus architecture, which consists of two 128bit busses working parallel to each other as explained in the link posted above by AJ. AFAI understand it, the great thing about this dual bus is, that the chip can read from and write to the frame buffer at the same time, but the memory interface of the G400 is still a "one way" 128bit bus. If I'm right, the G450 will have an "internal unidirectional input bus"@ 128bit + "internal unidirectional output bus"@ 128bit(=G400 256bit dual bus), but in contrast to the G400 the G450 will feature some sort of "external unidirectional input memory bus"@128bit(64bitDDR) + "external unidirectional output memory bus"@128bit(64bitDDR). Maybe this way Matrox dual bus architecture will unveil its true potential.
This could have another significant advantage over the "one way" bus that everybody else is using at the time. Somehow, it seems to be very difficult to develop a "classical" 256bit bus and thus overcoming the bandwidth limitations of current "single-chip" graphics boards. NVidia used DDR memory for the NV10, but this wont be enough bandwidth for the astronomical (texel)fillrate the NV15 is supposed to achieve. IMHO the NV15 will be severely bandwidth limited at highres/32bit color/32 bit texture, if they stick with a DDR 128bit bus. Now look at the Matrox dual memory bus I mentioned above: 64bitDDR dual bus(64*2*2) is the same as the NV10 DDR, basically an effective 256bit memory bus. The great advantage of the Matrox solution is its dual bus nature, which allows them to read from and write to the memory simultanously and all of this with a physically tiny and easy to handle 64bit bus. Since the G800 is supposed to have a massive fillrate, they could match it with an 128bitDDR dual bus (128*2*2=512!!). This would solve the bandwidth limitations handily, while they could still operate with a physically 128bits wide memory bus.
I don't know if this is going to be true or if anything I stated above is technically correct or even possible, because I'm not an expert at all. Please tell me what you all think about my "dual memory bus theory" and feel free to correct anything of the technical stuff, like I said, I'm no expert at this, but if it's true, it's really exiting news for the G450/800.

Ole

------------------
My Sytem: P3 450@558, Asus P3B-F(Bios 1005), 128 MB PC-100(2,3,3,4, fast), G400 Max@174/218 (MGA-Tweak:2,5/2/2,5,435) PD 5.52wTGL1.30, 13GB Quantum Fireball CR, Pioneer 40*CD-Rom, SBLive!Value (Liveware 3.0)

[AJ]

Hmmm, alot of interesting comments. If the G450 is using FCRAM, and gets about a 30% boost in effective bandwith from it, then the 166(333)Mhz x 64 bit interface would drive about the same data rate as the 200MHz x 128 bit interface. Very interesting. So this could mean that the G450 may sport a 200MHz core clock rate and make decent use of it. With the 200MHz FCRAM, they could support an even higher core clock rate (I think). Seem to me I had seen some benchmarks indicating that the G400 was limited by core fill rate, not memory bandwidth, so a 1:1 ratio core to mem clock would be ok.

As far as the dual bus business, I'm not sure whether they are refering to their (internal)DIB or a dual channel external intreface (which would give insane bandwidth and an insane cost). But it does sound to me like the FCRAM temorarily solves the bus width problem with high clocks and lower latency, expecially if the 250(500)MHz variant is out this fall.

This is definitely good news. I really hope that the G800 will hit 1GigaPixel per second in its single chip form, that would be worth the wait!

dneal: the effective bandwidth of 133(266)MHz DDR DRAM is that bad! I knew DDR had a pretty nasty latency hit, but that much is nuts, seems to me that the NV15 is going to choke unless they are using FCRAM as well.

I just saw a preview of the V5 5500 (with benchmarks) at beyond3D.com, it's nice, but I'm glad that I'm waiting for the G800. Downside is that the 6 months or so is going to seem like forever!

-AJ

Hey, this is fun! I think I like little bit of data coming out piece by piece, I get to learn much more from others rather than just reading the specs and not thinking about them.

[CannyOne]

AJ I couldn't agree with you more! On multiple counts!! I'm sure the "huwy" will get quite deep at some sites in the coming weeks. It already seems to be happening.

I will admit to breifly considering the Voodoo 5 5500. Then I discovered that this board will only have 32 MB onboard. (It was supposed to have 64!) This memory is not shared. Simply put, this is the equivalent of two 16 MB cards in an SLI configuration. And when people try to use 4X FSAA, at higher resolutions with 32-bit color, the framerates will "tank out". But because it's a 3dfx people are going to buy the hype.

I've pretty much decided a TennMax heatsink/fan, some heatsinks on the memory chips, and some judicious overclocking of my MAX are in order. This will help me hold out for the G800! Six months is going to "seem" like a long time. But I can run most of my games at pretty high resolutions, with decent framerates, right now! A little tweaking is all I really need to do. Then I'll just "have fun", that's supposed to make "time fly".

Here's a little story to illustrate the point. One day a philosopher visited a farmer, and asked: "Do Cows think?" The farmer replied: "Yes, they think! But they have a problem. They don't think things through!" It seems obvious that some people have a "Cow Problem"!! Don't be one of them. Think things through!! End of "sermon"!

[Proton]

Hmmm, seems like nobody really likes my theory of a dual memory bus with G450/800.

AJ: The march rumours about the "Condor" I + II were clearly talking about a dual 64bit memory bus. I checked http://www.kbench.com, were this info originally came from and they say that the 16MB G450 will use 166MHz DDR or 200MHz FC DDR and 64bit "dual memory bus." I know that these are rumours or at least unofficial information, but this is the only information we have and it is what we are discussing here. You can find the "dual memory bus" quote at http://www.ixbt-labs.com .cgi/archives.cgi?category=1&view=3-00 (the march 2000 news archive) and at the MURC News under the title "Flight Of The Condor". IMO it's obvious that this "unofficial information" is about a dual memory bus@64bit. The core is rumoured to be basically the G400Max core on an 0.18micron manufacturing process and this means to me an internal dual bus@128bit. It would make no sense to downgrade the core to a 64bit dual bus, thus the rumoured 64bit dual bus should be the memory bus.
There are more reasons for me to think that this 64bit memory bus can only be a dual one:
The G450 is rumoured to ship with 166MHz DDR Ram and probably 200MHz FC DDR. If the G450 will have a 64bit "single" memory bus, the bandwidth of 166MHz DDR version would be the same as the "vanilla" G400:
-G450 64bit "single" memory bus with 166MHz DDR Ram: 166(memory clock)*2(DDR)*8(64bit/8=8byte)=2,66 GB/s
-G400 "vanilla" 128bit memory bus with 166MHz SDR Ram: 166(memory clock)*1(SDR)*16(128bit/8=16byte)=2,66 GB/s
With the G450 moving to an 0.18micron manufacturing process, you can definitely expect higher clockspeeds for the chip itself, since it looks like the internal architecture will be equal to the G400's. Therefore it wouldn't make any sense at all to combine a chip@~200MHz clock speed, with the memory bandwidth of the "vanilla" G400(core@125MHz). If this would be true, the G450 wouldn't be any faster than the "vanilla" G400 in games@1024*768/32bit color/32bit textures. I'm sure Matrox wouldn't present a successor to the G400 Series that would be slower than a G400 Max!
Now let's look at the version with FC DDR Ram@200MHz. AFAI understood it, FC DDR has a higher effective bandwidth as regular DDR-Ram but the peak bandwidth stays the same which would translate into a peak bandwidth equal to the G400Max(3.2GB/s) if the G450 would be equipped with a "single" 64bit memory bus. Although this setup would benefit from the higher effective bandwidth of the FC DDR, it still wouldn't be enough for a G400 core@~200MHz and certainly wouldn't be worth the redesign of the chipset and wouldn't justify the use of FC DDR-Ram, which shouldn't be too cheap, to say the least.
To sum it all up: I'm convinced that the G450/800 will use a dual memory bus, but I'm definitely not sure, so I'm still interested in other peoples opinions on this and the pros and cons of this suspected "dual memory bus".

Ole

------------------
My Sytem: P3 450@558, Asus P3B-F(Bios 1005), 128 MB PC-100(2,3,3,4, fast), G400 Max@174/218 (MGA-Tweak:2,5/2/2,5,435) PD 5.52wTGL1.30, 13GB Quantum Fireball CR, Pioneer 40*CD-Rom, SBLive!Value (Liveware 3.0)

[CannyOne]

Proton I like your "theory"!! It sounds reasonable. I've read the same articles, though some of them seem a bit vague in places. So until I can get some more information from Matrox. I don't want to draw conclusions. However, I agree with you that Matrox wouldn't produce a "successor" that is slower than the MAX is now! That does seem sure.

I guess I did digress in my first reply. It's just that I logged on to the forum to find a topic with a link to 3dfx V5 info!?! I was still in "rant mode" when I started posting. I apologise! I'll try not to let it happen again.

[dneal]

Proton, the idea of a dual-channel memory bus is really nice. I've always wondered why Matrox has such a revolutionary core config with dualbus but the memory interface was so "conventional."

With a dual-bus memory setup I'm sure they can match that up with the internal dual-bus for some excellent results........assuming that's what they're up to :P

If I'm correct, a dual-bus memory setup wouldn't only nearly double bandwidth but also halve overall memory-core latencies!


------------------
ABIT BF6, Pentium III SL35D 450MHz -&gt; 630MHz, 192MB PC100 SDRAM @ 140MHz, Toshiba 6X DVD,
Matrox Millenium G400 DH @ 160/200, Creative SBLive Value, 3Com Fast Etherlink XL PCI
Supermicro SC701A ATX 300watt TurboCool PC Power & Cooling PS, Panasonic Panasync S17

[AJ]

Proton, it's just tough to figure out what is going on with the memory bus. The wording used in the various rumors is vague, and my knowledge in this area is poor.

But I do note a few things:
- 16MB frame buffer vers : 2 x (2M x 32 DDR)
- 32MB frame buffer vers : 4 x (4M x 16 DDR) or 4 x (2M x 32 DDR)


So, the 32MB version seems to have a 4*32bit configuration, and that implies 128 bits for the external bus.

The whole dual 64 bit _memory_ bus thing eludes me, whether they mean internal (makes no sense) or external (why, unless it cuts down on signal noise somehow).

-AJ

[Jutai]

well it loks like all the non-technical people are interested in this thread how unfotunate...
well i'll add my not knowin much bout this 2 cents. typically matrox has seen the memmory bandwidth as one of the biggest limmiting factors on their video cards they have always made every effort to take advantage of the latest and greatest memmory technology available. One might think of this as a possible "value" card but i think the rumors are quite contrary to that with charts showing the discontinuation of the MAX boards with the unveiling of the G-450's and the rumor of double the fillrate of the G-450 who would boast about double the fillrate of their "value" card coming in six months. i think the dual 64bit bus theory is interesting it would show the vision of the dual bus techology to lower costs on the pcb allowing for more money spent on better memmory and such while gaining bandwidth. Matrox enigneers are the bomb imo they know whats up and i think they've got some special stuff up they're sleeves i think the G400's were just a taste of Matrox's move on the high end gamers market and next cycle they will open the whole can of whup A$$ on everyone
-Chris k.