Kyro 2 the killer of nvidia ???

noko · Apr 18, 2001

Positive except for the price somewhat...

If you've got to have a mid-priced graphics card right now, I think the Prophet 4500 certainly justifies its price, but it doesn't come out as being noticeably better value than an MX board for half the money. Buy an MX card instead and you can get another 128Mb of PC133 RAM, and change, or a couple of games.<A HREF="http://www.dansdata.com/prophet4500.htm" target="_new">http://www.dansdata.com/prophet4500.htm</A>

I do believe these cards will have a lower price in the $100 range after several months have gone by. Then It will be an excellent deal. Right now I believe the best deal for the money by far is a Radeon LE. Especially if you are looking at playing Max Payne this summer

. I just can't wait...

noko · Apr 18, 2001

Ouch!!!!!

Edited by noko on 04/17/01 08:16 PM.

Guest · Apr 18, 2001

That review was finally published so I can now post those Kyro II DroneZ numbers:

640x480x16/32 = 59.82fps/59.40

800x600x16/32 = 58.43/58.67

1024x768x16/32 = 56.90/56.79

1280x1024x1/326 = 52.44/52.10

1600x1200x16/32 = 43.21/42.79

FSAA x2 tests:

640x480x16/32 = 58.58/58.52

800x600x16/32 = 55.60/56.47

1024x768x16/32 = 50.69/50.60

1280x1024x16/32 = 38.90/38.60

FSAA x4 tests:

640x480x16/32 = 54.73/54.99

800x600x16/32 = 46.57/46.99

1024x768x16/32 = 34.19/34.20

Now you will notice that this game is CPU limited in lots of resolutions but not nearly CPU limited enough to make performance at all bad, 1024x768x32 with 4xFSAA at 34fps in a DX8 vertex shader game with 4 texture layers with the Kyro II, now thats impressive, and at 1024x768x32 without FSAA it gets 56fps which is a really nice framerate, if this is a sign of how well DX8 games will run on the Kyro II then it doesn't look bad at all does it?

oh and BTW the test system was a P3 700mhz.

Guest · Apr 18, 2001

I looked in the "What sort of system will be required to play Doom 3?" section and there was no mention of the requirment of a 700Mhz cpu but a high end graphics card from Carmack.

EDIT: Oh right now I see it, I was looking in the blue bits cos I thought you meant that Carmack actually said this.

Edited by Teasy on 04/17/01 11:09 PM.

noko · Apr 18, 2001

Allright, I am limiting my discussion to the Kyro2 with a known fillrate of 350million pixels per sec (mp/s) with one texture applied to an 3d object. As the number of textures per object increases, the Kyro2 fillrate decreases proportionaly. Meaning, at two textures per object the fillrate goes down to 175mp/S. 3 textures -> 117mp/s. 4 textures -> 87.5mp/s..... 12 textures -> 29mp/s.

These pixels being rendered in the Kyro2 pixel pipe lines are what is going to be displayed on the monitor. Every pixel rendered will be displayed on the monitor for your eyes to see. None of these rendered pixels will be thrown away. Now on a typical card after the pixel is texturized it is then compared to a Z-buffer to determine if it is going to be used or thrown away based on if the object or part of object being rendered is occluded or not. Overdraw - rendering of hidden surfaces which your viewpoint occludes.) Assuming that the Kyro2 has zero overdraw then all the pixels will make it to your monitor.

12 texture object

Your in the heat of the battle and you move towards a 12 texture object in a game. As you approach, the object on your monitor it is covering more and more of your screen until the whole object is taking up all the pixels on your display. It starts to disintegrate as you quickly try to move around it. Time is of the essence or be dead.

What will happen on the Kyro2:

. A. Ideally

. . . Scenario: 1280x1024x32 display, 1.53ghz Polimino on a DDR2100 system.
. . . We are assuming the cpu can handle any T&L calculation, game physics calculation
. . . and has unlimited bandwidth to the Kyro2 meaning the AGP bus transfer rate is
. . . infinite and won't limit data to the Kyro2 memory or chip. Not relistic but
. . . lets look at what happens.

. . . The maximum FPS the fillrate of the Kyro2 can support for pixels on your monitor on this
. . . rather ideal system

. . . . . Number of pixels on a 1280x1024 display is: 1280 x 1024 = 1,310,720 pixels
. . . . . Fill rate of a 12 texture object on a KyroII is limited to 29mp/s (million pixels/sec)
. . . . . Calculating Max FPS based on Fillrate:
. . . . . . . . (29000000pixels/sec)/1310720pixels = 22FPS maximum the pixel pipe lines can support.

hmmmmmmmmm

. B. Reality

. . . Scenario: 800x600x32 display, 800mhz Duron on a pc133 system. In this
. . . more relistic setup the cpu causes a conservative 40% stall on T&L calculation due to advance
. . . game physics. Look at my previous benchmarks dealing with HW T&L and SW T&L which makes
. . . this a realistic number. Meaning that the Kyro chip will have to wait in order to receive
. . . the next vertex coordinates as well as lighting information 40% of the time causing the pixel
. . . pipeline to be idle. This is not the only problem that the KyroII is going
. . . to encounter...

. . . . Memory Requirements of 12 textures

. . . . High detailed textures at 32bit take up significant amounts of ram even when compressed.
. . . . Today high detail 2 texture games barely fits into a 32meg card with compression turned on,
. . . . now we are adding 10 more! It would be safe to say that AGP texturing would be used in this
. . . . case. Even on a 64meg card.

. . . . . . . The clevits
. . . . . . . . 1. Kyro2 memory bandwidth is 2.8gb/s which is sufficient for the kyro2 to operate.
. . . . . . . . Except now textures are being placed on main memory. The first Clevit is really the
. . . . . . . . data path from the Kyro2 to the main memory. The rather slow 2x AGP speed with a
. . . . . . . . slow 528mbits/s transfer rate. Some 5.3 times slower than the onboard memory of the Kyro2.
. . . . . . . . Meaning 5.3 times slower to load a texture into the KyroII chip.

. . . . . . . . 2. Hey wait a minute?? This is also the data path for T&L calculations and
. . . . . . . . polygonal data to the Kyro chip which will also have to share with the texture
. . . . . . . . transfers. Oh boy another bottleneck.

. . . . . . . . 3. Another Clevit occurs when AGP texturing is used, CPU performance decreases,
. . . . . . . . reason is that only one device can access memory at a time.

. . . . . . . . 4. Summary of clivits

. . . . . . . . . a. 40% reduction in accelerator performance due to slower AGP textures ( A conservative number.)
. . . . . . . . . b. AGP2x bottleneck reduction? 10%, 20%????(Won't use in final estimated calculation.)
. . . . . . . . . c. 10% reduction in CPU performance (When the AGP device is accessing
. . . . . . . . . the main memory then the CPU is locked out of that memory thus reducing overall
. . . . . . . . . CPU performance)

. . . The maximum FPS that the fillrate of the Kyro2 can support for pixels on your monitor on this
. . . more relistic system

. . . . . Number of pixels on a 800x600 display is: 800 x 600 = 480,000 pixels
. . . . . Fill rate of a 12 texture object on a KyroII is limited to 29mp/s (million pixels/sec)
. . . . . Calculating Max FPS based on Fillrate:

. . . . .{(29000000pixels/sec)/480,000pixels} x 60%(note a) x 55%(note b) = 19.9FPS maximum the pixel pipe lines can support. Not really enough for all the other factors in a game.

Note a. AGP texture effect
Note b. T&L stalls, cpu degradation due to AGP textures

Conclusion
Kyro2 doing 12 textures is not a viable solution. Kinda ridiculus even to think about it. Whats funny is when plugging in 8 textures it still doesn't look to good.

Edited by noko on 04/18/01 09:36 PM.

noko · Apr 18, 2001

Please post link to this review. The small hit when doing FSAA is impressive. Still the numbers don't mean much without comparisons or a feel for what type of game we are dealing with. On another note its kinda funny the review that PowerVR2 did a linked to because it had a KA7 with a 800mhz Atlon Classic. Exactly what I have sitting in my closit waiting to be used again. Any word on Linux support? Dam, actually I would buy a Kyro2 if it had some sort of linux support. Love to use the FSAA on FlightSim2000.

Guest · Apr 18, 2001

We've been discussing that Sweeney comment over at Beyond3d, I think everyone can see he's a guy that has a real vendetta against PowerVR, for a start its a TNT2 class chip??, in what way is it like a TNT2 chip, is it anywhere near the performance of a TNT2 chip......nope its around GTS performance, is it like a TNT2 chip in feature set.....nope, it has Dot3, EMBM, Anisotropic filtering, FSAA, Texture Compression, 8 layer multi-texturing etc so in features and performance its nothing like a TNT2, so what is Sweeney talking about you ask?, well Sweeney has the crazy notion that tile based renderers don't actually have any advantage over a similarly speced traditional, he said in another comment that every tiler thats been released theoretically should be faster then a comparably speced traditional but because a tiler has to capture the all the geometry in ram this somehow negates all the other advantages and means that no tiler has every really been faster then a similarly speced traditiona, this is what he's trying to say here, he beleives that because a Kyro II's raw speces don't look miles away from a TNT2 ultra then the Kyro II is no faster then a TNT2 Ultra, now many of you who have seen Kyro 1 and II reviews will now be saying WTF is he talking about???? because clearly the Kyro II beats not only cards that are speced close to the Kyro II but also cards with 2 or 3 times the raw specs of the Kyro II, so why does Tim Sweeney have this ludecrous notion?, well there's a few possibilities here:

1 Tim Sweeney has never used a PowerVR card and is merely theorising that this will be the case (I know this isn't true)

2 Tim Sweeney has used early PowerVR like the first PCX1 which was released way back before the Voodoo1 and maybe even the Neon250 back in 1998, these cards had a few problems with performance in some games which he maybe beleived to be related to this geometry storage and so he never again even tried a PowerVR card?, so he's never used a Kyro 1 or II and also never saw a review or heard anything about the card being faster then the MX and also the GTS at high res????, in other words he's been berrying his head in the sand for years now???

3 Epic signed an agreement with Nvidia, Tim already doesn't like PowerVR and so he's attacking the Kyro II for that reason?.

Make up your mind which, but whichever it is he is wrong in the extreme and his comments really show incredible ignorance of the Kyro II, he says the Kyro II doesn't have some DX7 features, he emphasises this as if to make fun of the Kyro II, Cube Mapping is the DX7 feature he's refering to, well the Geforce 2 Ultra doesn't have a DX6 feature called EMBM (yes 6)

), and of course his comments that the Kyro II is not a viable card for the market its aimed at speaks for itself, the Kyro II 32mb TV-Out costs £85 here in the U.K inc VAT, the cheapest MX 32mb here is £75, the cheapest GTS 32mb is around £150, when the Kyro II 32mb is released in the U.S by a cheap board maker expect a $80 price tag online, not viable at at such a cheap price with such great performance? as you can see Tim hasn't really took any time to think about what he's said he's just tried to ridicule the Kyro II as much as possible in order to mislead people, well unless you buy the fact that he's never actually seen a PowerVR card since the Noen250 was made in 1998? and never ever seen any reviews of either the Kyro 1 or II?.........Nope I don't think anyone can beleive that.

noko · Apr 18, 2001

Downloading the 66.6 meg avi file for DroonZ. (Weird number, trying to remember what that number means??)

Guest · Apr 18, 2001

Now do the same calculation with the MX or GTS and use a normal overdraw value of around 2.5 (the GTS and MX has to draw 3 and a half times the amount of pixels as the Kyro II), remember to use the single texture fillrate scores in SS as there starting point to and also remember the extra bandwidth hit that will be taken by all those extra passes as well as all the AGP bandwidth wasted by having to pass each polys through 5 times per pixel for 10 layers of textures and 6 times per pixel for 12 layers, thats if you have time of course. Also when I was talking about 10 and 12 multitexturing I was talking about the card iself not the card in any perticular system, also another thing you have to think of is the Kyro II only has the have textures stored in ram that are actually going to be visible, with a overdraw level or 2.5-3 the GTS or MX or any other traditional need to store 3.5-4 times the amount of textures as the Kyro II and the Kyro II also has more space for textures because it needs no z-buffer in ram, so texture sotrage space on the Kyro II isn't as much of an issue is it is with traditional cards, its 3-4 times less of an issue infact in most games.

Oh and thanks for taking the time to do those Kyro II calculations, they were interesting.

Edited by Teasy on 04/18/01 00:13 AM.

noko · Apr 18, 2001

Tomorrow I will calculate a Radeon pipeline. Using the 3 texture rate in 3dMark2001 as a value. I made a correction to my calculation, I had a math error, now the Kyro2 has a higher final FPS. Using one texture value vice the capability of the video card would be the wrong way to calculate doing multitexturing. Plus the Kyro2 is much easier to calculate for since the pixel pipeline represents what the monitor receives. I can extrapolate data from posted benchmarks and make a few assumptions. We will see how it works out. Also I will do some further research on this matter. Watched the high quality AVI file of DroonZ, a game I wouldn't buy but colorful. Would like to see benchmarks of other cards doing this game.

Edited by noko on 04/18/01 00:22 AM.

Guest · Apr 18, 2001

The benchmark was from the only just released Beyond3d Kyro II preview (http://www.beyond3d.com/previews/vividxs_kyroii/index1.php), unfortunately you'll be unhappy to find out that the preview does not include any other cards except for a Kyro 1 so you can't see any comparison with DroneZ, but my main point was that a not yet released DX8 Vertex shader game thats been endorsed by Nvidia as being one of the real next gen DX8 games runs great on the Kyro II, just look at the framerate at 1024x768x32 with 4xFSAA = 34fps, and at 1024x768x32 without FSAA = 57fps, thats incredibly impressive for this sort of game.

noko · Apr 18, 2001

Thanks for the link, I will read tomorrow. GoodNight that is from the states.

HolyGrenade · Apr 18, 2001

Cool Link. I might get the book sweeney mentioned. Thanx for the link.

"My name is Ozymandias, King of Kings:
Look on my works, ye Mighty, and dispair!"

Guest · Apr 18, 2001

I got this info from IMGTEC on linux Kyro 1 and II drivers:

"We are developing Linux drivers which are currently in early stages. I don't have a time frame at the present."

So yes they deffinately are making Linux drivers, I have access to Kyro 1 and II drivers before they become public so if I see a linux driver ready then I'll tell you.

Guest · Apr 19, 2001

wow, I don't think I've ever seen a thread go this long before, and only half of it is flaming!

noko · Apr 19, 2001

Early stages doesn't sound like a linux driver will be coming in the near future. Oh well I can't have everything. Been refreshing on the different filtering methods (point, Bi and Trilinear, anistropic. Which affects texture bandwidth of a video card. Fillrates of the Radeon can be found at:
<A HREF="http://www.beyond3d.com/reviews/Radeon64MBTIVO/index1.php" target="_new">http://www.beyond3d.com/reviews/Radeon64MBTIVO/index1.php</A>
Also at ATI:
<A HREF="http://www.beyond3d.com/reviews/Radeon64MBTIVO/index1.php" target="_new">http://www.beyond3d.com/reviews/Radeon64MBTIVO/index1.php</A>

Also the new review you linked to came to a similar conclusion dealing with the texture bandwidth limitation of the KyroII. Still no clear reviews on DVD playback and 2d quality of the Kyro2. The Kyro2 does have some great advantages and some disadvantages as well. Hopefully everything will be clearly laid out on the table for someone to make a wise decision.

Would like to see some Vulpine benchmarks. Version 1.1 cleared up the problems with the Kyro/2. A high detail OpenGL benchmark using extremely high polygonal models.

Negaverse23 · Apr 19, 2001

This thread has only just begun.

=
Are you just lazy or incredibly stupid? -King of the Earth

noko · Apr 19, 2001

Correction
Appears that the review at Sharkey extreme found the 2d to be allright on the Kyro2 3dProphet 4500 at 1600x1200. Very good news concerning 2d. DVD was comparable to a GF2 card but not as good as a Radeon.

Edited by noko on 04/18/01 09:42 PM.

OzzieBloke · Apr 19, 2001

- A boy who needs a friend finds a thread that never ends... -

The Neverending Thread

Coming to a forum near you.

Australian PMs are like steer horns; a point here, a point there, & a heap of bull in the middle.

Guest · Apr 19, 2001

Well Like Holygrenade I was gone from this forum for a few days... and it has now gone from 17 pages to 30! Never-ending thread is right...

Since I just spend all night reading what I missed (and reading all the links in all the posts I missed...) I will not take time now for any detailed responses. But I do have a couple remarks:

Ozzie -- You come up with consistently hilarious stuff man. Don't stop!

Noko – Thanks for puting in all that work doing all those benchmarks and related research. Your posts are always well thought out and well researched, which is cool.

Teasy – You have actually helped to change my mind on a point or maybe two, though I remain stubborn on the others.

But like I said I will go into details tomorrow.

Holygrenade – Thanks for that big long post you made when you first got back because you said most of the things I was going to say. Saves me some typing.

Cheers,
Warden

===========
The sum of the IQs on this planet is a constant--only the population is increasing...

Guest · Apr 19, 2001

phsstpok,

I have a question. (My ignorance is really going to show). What is the relationship between FSB speed and AGP speed? I understand that AGP 1X is normally 66 mhz. 2x and 4x are then obvious. Here is where I am confused. This implies the AGP 4x would theoretically be 4 x 66 mhz or 266.7 mhz (66 mhz is really 66.6666...). How can AGP operate this fast if FSB is only 133 mhz?

Unless I missed something, your question never really got answered. The difference is that while the AGP bus is 32 bits wide (4 bytes), the FSB is on SDRAM systems is 64 bits wide (8 bytes). The bandwidth then works out as follows:

AGP 4x: [4 bytes x 66 MHz x 4] = 1.056 GB/Sec
133 MHZ FSB: [8 bytes x 133 MHz x 1] = 1.064 GB/Sec

So AGP 4x almost exactly matches the bandwidth of the most common FSB. Hope this clears it up for you. (Note that the final terms, x4 and x1, account for the number of bits transferred per clock cycle.)

Regards,
Warden

===========
The sum of the IQs on this planet is a constant--only the population is increasing...

phsstpok · Apr 19, 2001

Poster: Warden
Subject: Re: Could AGP2x and Software T&L kill the Kyro2?

phsstpok,

In reply to:
--------------------------------------------------------------------------------

I have a question. (My ignorance is really going to show). What is the relationship between FSB speed and AGP speed? I understand that AGP 1X is normally 66 mhz. 2x and 4x are then obvious. Here is where I am confused. This implies the AGP 4x would theoretically be 4 x 66 mhz or 266.7 mhz (66 mhz is really 66.6666...). How can AGP operate this fast if FSB is only 133 mhz?

--------------------------------------------------------------------------------

Unless I missed something, your question never really got answered. The difference is that while the AGP bus is 32 bits wide (4 bytes), the FSB is on SDRAM systems is 64 bits wide (8 bytes). The bandwidth then works out as follows:

AGP 4x: [4 bytes x 66 MHz x 4] = 1.056 GB/Sec
133 MHZ FSB: [8 bytes x 133 MHz x 1] = 1.064 GB/Sec

So AGP 4x almost exactly matches the bandwidth of the most common FSB. Hope this clears it up for you. (Note that the final terms, x4 and x1, account for the number of bits transferred per clock cycle.)

Thanks, Warden. The picture is starting to get a little clearer.

I had another question and I can't recall if it was in the same post. Basically my question is, what is limiting AGP performance? AGP 4X does not seem significantly faster than 2X. I thought I figured out the answer but, based on what you are saying, AGP 4x and FSB data rates are comparable. Therefor, I must be wrong. Something else is limiting performance.

Your response does bring up more questions. Does the Athlon's DDR FSB double the data rate of 1.064 GB/sec which you quoted for a 133 Mhz FSB, at least in theory? Also, I have an older KT133 motherboard and my maximum stable FSB speed is 106 Mhz. How does this factor into the equations?

Sorry, I know you were just trying to help. Now, I am asking you to teach PC Design, 101.

noko · Apr 20, 2001

There is a multiplier of the Front side bus which controls the AGP bus speed. The AGP bus speed was designed for 66mhz. For a 133mhz FSB the AGP multiplier for the AGP bus is 1/2 of the FSB {133mhz x 1/2 = 66mhz}. For a FSB of 100 the AGP multiplier is 2/3 {100mhz x 2.3 = 66mhz}. Now if I took my IWILL KK266 FSB from its 133mhz FSB speed and raised it up to 150mhz I would also increase my AGP bus 150mhz x 1/2 = 75mhz, which is out of spec but most modern day video cards operate just fine. On Intel BX systems the AGP multiplier is 1 for 66mhz FSB and 2/3 for 100mhz. The BX chipset never had a 1/2 AGP multiplier so when a BX chipset motherboard FSB is overclocked to 133mhz you get the following; 133mhz x 2/3 = 89mhz for the AGP bus, which is even more out of spec but still most video cards today can do it. Really AGP is not a bus but a port to the cpu via the northbridge. Still alot of people call it a bus. Don't get too confused about AGP transfer rates 1x(1byte/clock), 2x(2byte/clock) and 4x(4bytes/clock) and AGP bus speeds. I hope this helps.

Now the difference between 2x and 4x AGP is that 4x can transfer twice the amount of data compared to 2x. Sounds Significant except most modern day applications still don't exceed the 2x bandwidth. On the other side some applications do.

Edited by noko on 04/19/01 05:48 PM.

noko · Apr 20, 2001

Thanks Warden. What else do we need to discuss in this tread prior to kyro2 release?

phsstpok · Apr 20, 2001

Sorry, Noko I knew I would have trouble making myself clear.

I understand the relative frequencies, the 2/3 divider for AGP (1/2 when FSB is at 133 mhz on with a KT133A chipset). What I didn't understand and am still have trouble with is the actual data rates and the architecture of both a Front Side Bus and an AGP port. I was concerned that maybe the FSB was just unable to keep pace with the theoretical speeds for AGP 4X. Without understand the different nature of both "buses" I came to the wrong conclusion. I thought that with AGP clocked at 66 mhz and 4x would give an effective rate of 267 mhz. I thought that a 133 Mhz FSB would be ineffective sustaining the high rate of transfer. I mean it looked like AGP could be twice as fast as FSB. If the FSB is transfering data to the Northbridge at a slower rate than the Northbridge is able to pass data to the AGP port and then onto the video card then the FSB would be a bottleneck.

Warden pointed out that while the FSB tranfers data 8 bytes wide (64 bits) while the AGP port transfers it only 4 bytes wide (32 bits). (Hope I got that right). As Warden said, this would actually make a 133 mhz FSB a good match as the data rate is about the same as that for a 66 Mhz AGP. This all sounds reasonable since the datapath for the FSB is twice as wide as that for the AGP and that AGP is "quad pumped", if you will. (At least that's what I think Warden was saying).

I my reply to Warden I brought up the fact that an Athlon's FSB is, in itself, "double pumped" and doesn't that actually make FSB twice as fast, in terms of data transfer not megahertz, as AGP port?

This situation would be the reverse of what I originally thought and this is where the discussion stands.

All this is still confusing to me.

Kyro 2 the killer of nvidia ???

Distinguished

Distinguished

Guest

Guest

Guest

Guest

Distinguished

Distinguished

Guest

Guest

Distinguished

Guest

Guest

Distinguished

Guest

Guest

Distinguished

Distinguished

Guest

Guest

Guest

Guest

Distinguished

Distinguished

Distinguished

Distinguished

Guest

Guest

Guest

Guest

Splendid

Distinguished

Distinguished

Splendid

Share this page