'Fusion' cards

[Rob M.]

<font face="Verdana, Arial, Helvetica" size="2">Originally posted by dZeus:
You know Rob and DGost, not that I want to stop you from having a discussion about 3D Hardware architecture and stuff, but at the forums of www.beyond3d.com there are several experts on this, including former 3Dfx employees... I bet you would find some interesting discussion partners over there (This absolutely is not, by any means, a way of me to ask you to bugger of out of this thread! So if you feel like it, just go on! )</font>

heh, I think we both know what were talking about, it's just that when we catch a mistake in the other's arguement we go on rants to show how each other is wrong. Problem is, when we rant, we're continuing to make mistakes...

[DGhost]

Very very very true rob... i do acctually believe we have been talking about the same thing for quite a while...

just can't say it right enough to please each other =)


Jord, who said anything about leaving????

[Rob M.]

<font face="Verdana, Arial, Helvetica" size="2">Originally posted by DGhost:
hehe... point taken Jorden

Rob, if you wish to discuss/argue/quabble/etc further, email me instead. my addy is in my profile...</font>

It's not fun if we can't publicly make each other look like idiots

...Can't wait for superfly's response

[DGhost]

lol

[frankymail]

Hi,
IMO, dual channel is still the only possible and viable alternative for an high performance 3D graphic chip for now and the near (6 months) future...

Sure, QDR-Sdram looks good on paper, but due to the added latencies of asynchronous memory settings, it would only provides a 65-70%, were DDR-SDRAM scores a nice 80-85% efficacity of its theorical maximum bandwidth (if you've overclocked a GeForce 2 Ultra, you've seen that running the core at 250 MHz and the memory at 250 MHz too is faster than running the core at 252 MHz and the memory at 255 MHz... this due to the fact that in the first case, the chip and memory cycles coincide more often than in the second case... waisting a 4-bits cycle is much worst than waisting a 2-bits cycle).

Also, not a single memory manufacturer has even announced availability/close-availability of products based on quad-data-rate technology. Take the example of NVIDIA's upcoming GeForce 3... the price will start at an astounding $600 US !!! For that price, you can get 600 Big Macs for that price at big M (Those of you that also reside in the province of Quebec all know that the "M" doesn't stand for McDonald, but for Mouettes . The card will use very expensice DDR-SDRAM clocked at250 MHz (you must know that high speed DDR chips' prices are out of this world, so imagine the prices of QDR-SDRAM chips...) So, QDR-SDRAM is not a viable option for the upcoming next-generation cards like ATI'S Radeon 2, NVIDIA's GeFOrce 3 and Matrox G800.

About the added cost of doubling the traces... 3Dfx showed us with the V5 5500 that it's possible to have two 128-bits buses on one card without pushing the price very high (The V5 5500 have always retailed for a little less than GeForce 2 GTS-based cards, which is still affordable). The price of the two VSA-100 chips would be approximatly that same as a bigger G800 with two 128-bits buses would. The V5 5500 still uses a 4-layers PCB (so no need for a 6-layers design), and have you seen how much PCB space is waisted? So, a graphics card with a big (as big as the GeForce 1/2/3 were...) G800 chip paired with a Dual-Channel memory architechture with a decent price tag is something very possible.

(BTW, I think I'm growing impotent because of all that G800-induced stress; last night, my gal and I only danced the Grand Mambo once and then both fell asleep watching "The Late Late Show with Craig Kilborn"...)

Francis Beausejour
Frankymail@yahoo.com

----------------------------------------
- Intel P3 850 MHz Retail O/C 1.133 GHz
- Alpha FC-PAL6030
- Asus CUSL2-M
- 256 MB TinyBGA PC150 (2 DIMMs)
- Matrox Millennium G400 DH 32 MB O/C 175/233
- Sony E400 + Sony
- Quantum Fireball LM 10 GB + AS 40 GB
- Teac W54E CD-RW + Panasonic 8X DVD-ROM
- Sound Blaster Live! Value
- 3Dfx VoodooTV 200
Francis: 19/Male/5'8"/155Lbs/light blue eyes/dark auburn hair/loves SKA+punk+Hard/Software... just joking, I'm not THAT desperate (again, Marie-Odile, if you ever happen to read this (like THAT could happen...), I'm just joking and I REALLY, REEEAAALLLLLYYY love you!))

[Rob M.]

<font face="Verdana, Arial, Helvetica" size="2">Originally posted by frankymail:
About the added cost of doubling the traces... 3Dfx showed us with the V5 5500 that it's possible to have two 128-bits buses on one card without pushing the price very high (The V5 5500 have always retailed for a little less than GeForce 2 GTS-based cards, which is still affordable). The price of the two VSA-100 chips would be approximatly that same as a bigger G800 with two 128-bits buses would. The V5 5500 still uses a 4-layers PCB (so no need for a 6-layers design), and have you seen how much PCB space is waisted? So, a graphics card with a big (as big as the GeForce 1/2/3 were...) G800 chip paired with a Dual-Channel memory architechture with a decent price tag is something very possible.</font>

The problem is that the chip packaging out there currently just isn't capable of handling that many leads. There's good reason why we havn't seen any single chip dual-channel sdram solutions yet.

I've seen a dual channel sdram motherboard using two memory controllers too, however it took up quite a large amount of pcb. There's going to be a break-even point where using a smaller 6-layer pcb board is going to be cheaper than going with a large 4-layer board.

However, it's not the PCB that's the biggest expense, it's the GPU. Even expensive 250mhz ddr memory is cheaper, especially if your yeilds are low.

As for qdr, I wasn't mentioning it because I thought it might come out with next-gen technology, I just said it was future technology. Faster memory in the near future may be that 'FCSDRAM' that matrox has been rumoured to be playing with in this thread. I agree with you that qdr isn't 4x the speed of standard sdram, but I'm it'll probably still be worth it once implemented, especially compared to a multi-gpu solution.

[dZeus]

I still don't really understand the difference between a so called 'dual-channel' 256 bit memory bus and a normal 256 bit bus... do they both require the same amount of pins/traces on the card? If so, it will be increadibly expensive to make, if not too expensive to make it available for consumers, as Rob M. pointed out.

[Rob M.]

The matrox dual channel bus internal. The external memory bus is still 128 bit. The reason for the 'dual channel' name is that 128bits are dedicated for reading, the other 128 are dedicated for writing. A true 256 bit bus can read or write data, 256 bits at a time.

[This message has been edited by Rob M. (edited 06 February 2001).]

[frankymail]

Hi,
I must disagree with some of your statements; First, a dual-channel memory subsytem doesn't mean that one bus is dedicated to write, one to read; it means two completely independant buses; one can read while the other writes, or both can read or write at the same time: they're INDEPENDANT !

Also, you said that the GPU is what drives cost up; I completly disagree with you... Have you ever spoken to someone in the graphics card industry ? Currently, chips like the Radeon and the GeForce 2 Family (Geforce 2 GTS/PRO/Ultra) cost between $30-40 per unit, while DDR-SDRAM costs more than $1 per megabyte, even more for the faster parts...

And about the PCB not being a major factor in the graphics card prices: have you read the multiple interviews with Abit when the Athlon 512 KB came out... they said they would not produce (and they did not) K7 motherboards until AMD provided them with a 4-layers design. Why, because the high production cost would have meant too high retail prices... (mentionned in the article...)

Again, I agree that textures compression, hidden surfaces removal, z-buffer compression and on-die RAM have to be implemented to maximise performances.

Oh, and about the pin counts, some of VIA-S3 chipsets (and do you know the pin count of an AMD 761 northbridge%) have MORE than 600 pins, where current ~35mm BGA chips like the Radeons the various GeForces (excluding the GeForce 2 MX) have 400-500 pins, and a lot of them are "reserved" (which means "not used"...

Wow, that was significantly shorter than my usual posts...

Notice I've decided to change my signature; I acknowledge it was a "little" too long, we wouldn't want to kill the forums . . . or would we ? ? ?


------------------
What was necessary was done yesterday;
We're currently working on the impossible;
For miracles, we ask for a 24 hours notice ...

[This message has been edited by frankymail (edited 06 February 2001).]

[frankymail]

To everyone; (and I hope Matrox employees are reading this!)

Now that this tread is more than 6 month old (and the sightings of "G800" and "Fusion" in those darn .DLLs are now more than 8 months old...), I wonder if this tread will reach the "one year" mark ? ? ? Excuse me M (Matrox, not Maggi), but that would socks big, BIG time! (notice I'm cautious about the words I'm using? . I just hope that the Mpire will come out in broadlight soon to reveal what every Matrox fan has been waiting for:

- 4 pix pipes, 3 tex per pipe
- Dual-Head
- Two High-speed on-die RAMDACs
- EMBM and other texture effects
- FSAA (the real hardware thing like 3Dfx, not the el cheapo emulation of theGeForce serie)
- 0.13/0.15µ die
- 200-250 (or higher) MHz core speed
- 4-8 MB of embedded RAM 512-bits wide (like the BitBoys design called for) @ core speed or higher
- Real AGP-4X usage...
- Textures and Z-Buffer compression, hidden surfaces removal (in hardware, please!)
- A 32 and a 64 Version available, and maybe a slightly stripped down version to boost OEM sales)
- The possibility to add a Rainbow Runner and TV-Tuner


------------------
What was necessary was done yesterday;
We're currently working on the impossible;
For miracles, we ask for a 24 hours notice ...

[superfly]

AGP 4x can transfer data up to 1 gig sec,but pc 133 certainly can't provide that much on a continuing basis,so the limitation is still there.

The only memory type so far that can acheive and even surpass 1gig/second of sustained memory bandwith are p4 systems using rambus memory

Need proof???...just look no further than the widely used sandra memory benchmark,where even the fastest DDR sdram acheives an average of about 700/750 meg/sec SUSTAINED memory speed,while standard sdram usually does no better than 500~550 megs second.

It was david perez(head nvidia pr guy)who first said when the first Geforce were anounced(the 50 000 poly limit),but since the quote was made nearly 18 months ago,it would be hard to give you a working link anyways.

You're also confusing my last quote,try to really understand it before offering a reply,ok.

It seems to me that i was only talking about the fill rate/bus bandwith limitation on the video card itself when you push it hard enough(resolution,color depth and depth complexity issues),not increased bus traffic in terms of geometry etc...

But here's a situation that will explain the system bus bandwith limtation in an even more obvious way.

i'll use q3 as an example,mostly because you can find plenty of benchmark data to be able to confirm it yourself,if you wish.

Take a p3 600,use it with a Geforce card and when you run the q3 benchmark at normal settings at 640*480 16 bit,you'll usually get about 100fps(+/- 5 fps)


Now use a 1200 mhz cpu(athlon or overclocked p3) with the same Geforce card at normal settings at 640*480 16 bit,and the most you'll ever see out of q3 is about 155/160 fps.

Assume both systems are using standard sdram,which according to you,has more than enough bandwith to go around.

Now there is no way that a Geforce is fill rate limited at that resolution and the 1200 mhz athlon IS twice as fast as the 600 p3 without a shadow of a doubt,so why didn't the scores double???

However use a p4 system,witch is no faster than the 1.2ghz athlon,at least until optimized software beconmes more widely available,and under the same test,you'll gain at least a 35 fps increase above that of the 1.2ghz athlon(nearly 200 fps).

Care to explain that???....

[This message has been edited by superfly (edited 07 February 2001).]

[This message has been edited by superfly (edited 07 February 2001).]

[Rob M.]

<font face="Verdana, Arial, Helvetica" size="2">Originally posted by superfly:
AGP 4x can transfer data up to 1 gig sec,but pc 133 certainly can't provide that much on a continuing basis,so the limitation is still there.
Need proof???...just look no further than the widely used sandra memory benchmark,where even the fastest DDR sdram acheives an average of about 700/750 meg/sec SUSTAINED memory speed,while standard sdram usually does no better than 500~550 megs second.
</font>

So why aren't developers using up this 500 megs/second?? AGP 1x is max 256 mb/s and it makes no difference if I use 2x or 1x.

<font face="Verdana, Arial, Helvetica" size="2">It was david perez(head nvidia pr guy)who first said when the first Geforce were anounced(the 50 000 poly limit),but since the quote was made nearly 18 months ago,it would be hard to give you a working link anyways.
</font>

Yeah, I'm gonna trust your memory. I assume NVIDIA PR person Derek Perez released some kinda statement though...

<font face="Verdana, Arial, Helvetica" size="2">You're also confusing my last quote,try to really understand it before offering a reply,ok

It seems to me that i was only talking about the fill rate/bus bandwith limitation on the video card itself when you push it hard enough(resolution,color depth and depth complexity issues),not increased bus traffic in terms of geometry etc...</font>

<font face="Verdana, Arial, Helvetica" size="2">This means that even if you're running a game at low resolution and color depth as well as in depth complexity,in which video card bandwith or fill rate is a non issue,if the developer uses too many polys in order to make everthing look as good as possible,you risk turning what would normally be a smooth game into a slide show,because all that extra vertex data will choke the system bus quite easily. </font>

<font face="Verdana, Arial, Helvetica" size="2">
But here's a situation that will explain the system bus bandwith limtation in an even more obvious way.

i'll use q3 as an example,mostly because you can find plenty of benchmark data to be able to confirm it yourself,if you wish.

Take a p3 600,use it with a Geforce card and when you run the q3 benchmark at normal settings at 640*480 16 bit,you'll usually get about 100fps(+/- 5 fps)

Now use a 1200 mhz cpu(athlon or overclocked p3) with the same Geforce card at normal settings at 640*480 16 bit,and the most you'll ever see out of q3 is about 155/160 fps.

Assume both systems are using standard sdram,which according to you,has more than enough bandwith to go around.

Now there is no way that a Geforce is fill rate limited at that resolution and the 1200 mhz athlon IS twice as fast as the 600 p3 without a shadow of a doubt,so why didn't the scores double???

However use a p4 system,witch is no faster than the 1.2ghz athlon,at least until optimized software beconmes more widely available,and under the same test,you'll gain at least a 35 fps increase above that of the 1.2ghz athlon(nearly 200 fps).

Care to explain that???....
</font>

You know what, it can take me a few hours to compile my quake3 maps when they're finished. That's a cpu-intensive process, id uses an 16-way xeon computer. We MUST be CPU limited, not bus limited.

No wait. I went to defrag my hard drive last night, and it took an ENTIRE HOUR!! We've GOT to be HARD DRIVE limited. I want speeds of at least 300 dps (defrags per second).

I could keep going, but I'll sum up: Yes, you're bus limited in that example, however there's ways to find the limits of every component in your system. Asking for more than 155 fps is ludicrous, if you're getting anything more than half that you're fine. IMHO, the bus speed is doing pretty damn well!

Fact is, while you can at best go no lower than 155 fps with a PC100 system, I can throw on quake3 at 1600x1200x32 with FSAA you'll get can get less than 1 fps on even the fastest video card on the market.

So, where would YOU say the biggest problem is?

[DGhost]

No matter how you look at a system you can find a bottleneck. the only way to remove bottlenecks is to 1) give each device its own share of bandwidth and 2) make everything run the same speed as the processor. and even then the whole system would be a bottleneck.

what difference does it make when the games look good and run faster than your eyes tell the difference?

wanna know what the biggest bottleneck in the system is?

the interface between the keyboard and the chair. computers now do things faster than a human can, and even while gaming spend a large chunk of time waiting on the user. that is evident by being able to get 155fps.

[Rob M.]

<font face="Verdana, Arial, Helvetica" size="2">Originally posted by frankymail:
Hi,
I must disagree with some of your statements; First, a dual-channel memory subsytem doesn't mean that one bus is dedicated to write, one to read; it means two completely independant buses; one can read while the other writes, or both can read or write at the same time: they're INDEPENDANT !
</font>

ROFLMAO. Sorry, not just at you in particular franky, or that statement, it's just been building this entire thread...

gotta good link for you dZeus on that mtx dual-channel thing. Lucky for you I got challenged. From here:

<font face="Verdana, Arial, Helvetica" size="2">With the new G400, Matrox has taken one step further and expanded the dual 64 bit bus to a dual 128 bit bus totaling in as much as 256 bit depth of the I/O interface. Both bus subsystems are unidirectional, meaning that we are looking at one 128 bit input bus and one 128 bit output bus</font>

<font face="Verdana, Arial, Helvetica" size="2">Also, you said that the GPU is what drives cost up; I completly disagree with you... Have you ever spoken to someone in the graphics card industry ? Currently, chips like the Radeon and the GeForce 2 Family (Geforce 2 GTS/PRO/Ultra) cost between $30-40 per unit, while DDR-SDRAM costs more than $1 per megabyte, even more for the faster parts...
</font>

ok, I admit, it's not the GPU in itself that's expensive relative to other components, but...
let's be nice. $30 for the gpu, $1.50 per mb DDR sdram
So. Single GPU 64mb card: $126
Double GPU 64mb (effective 32mb) card: $156
Double GPU 128mb (effective 64mb) card: $252

Notice even the double gpu 64mb card costs nearly 20% more than the single gpu card. I don't know if you've talked to manafacturers, but 20% is a LOT.

Now let's not be so nice, $40 per GPU, $1.00 per mb of memory:

Single GPU 64mb card: $104
Double GPU 64mb (effective 32mb) card: $144
Double GPU 128mb (effective 64mb) card: $208

Now if you have a next-gen video card with only 32 mb effective of onboard ram, you're just not gonna cut it. It's rumoured the nv20 will have a 128mb configuration.

<font face="Verdana, Arial, Helvetica" size="2">And about the PCB not being a major factor in the graphics card prices: have you read the multiple interviews with Abit when the Athlon 512 KB came out... they said they would not produce (and they did not) K7 motherboards until AMD provided them with a 4-layers design. Why, because the high production cost would have meant too high retail prices... (mentionned in the article...)</font>

Interesting logic, but you have one flaw. Abit isn't the only mb manafacturer. In fact, they're only a 2nd rate manafacturer. Had you said it wasn't worth it for ASUS or Gigabyte to produce AMD boards, that may have carried some weight. In reality, many other low volume mb manafacturers produced those boards, so the cost couldn't have been THAT much.
BTW, I never said PCB cost wasn't a consideration, I just said that chip prices were a bigger one. I just love a good arguement though

I guess I shouldn't have to mention the fact that most mb's require a little more pcb than a video card as well, possibly to the amount where it's no longer viable to produce boards on a 6 layer process (oh wait, I am repeating myself)...

<font face="Verdana, Arial, Helvetica" size="2">Oh, and about the pin counts, some of VIA-S3 chipsets (and do you know the pin count of an AMD 761 northbridge%) have MORE than 600 pins, where current ~35mm BGA chips like the Radeons the various GeForces (excluding the GeForce 2 MX) have 400-500 pins, and a lot of them are "reserved" (which means "not used"...</font>

Who's manafacturing these chips? What process? How many extra pins does it require to implement a 2nd sdram channel?

Oh, and English probably isn't your first language, so I forgive you:

<font face="Verdana, Arial, Helvetica" size="2">re·served (r-zûrvd)
adj.

Held in reserve; kept back or set aside.
Marked by self-restraint and reticence. See Synonyms at silent. </font>

[Rob M.]

woah. this anandtech article that came out today is pretty relevant. The DDR board didn't perform as bad as I figured, although both it and the dual-channel sdram board seem to be very beta..