Specs for r300 and rv250

[Tygrus_Artesiaoa]

http://www.3dcenter.org/artikel/2002/06-07_english.php

Yep a bargain part for rv250...seemingly a little bump in speeds with 2 pipeline cut out. Wonder how much cheaper that will make it and how much it will suffer for performance. Hopefully and it doesn't appear they will cut out TnL or TLC engine and turn to 64 bit ddr...but as many flavors of its cards ATI has...I wouldn't doubt it if they had a version like that
As for r300...interesting specs...I heard or read NV gonna use some blazing fast ram for its cards.
I wonder if it's gonna be similar to r300 and gf3 where r300 is clocked higher but not necessarily outperforming it. Course gf4 4600 is clocked higher than radeon 8500

 

[nggalai]

Just for clarification's sake (Leo didn't feel like translating the introductory notes into English ):

this article is an accumulation of RUMOURED specs as found in AnandTech's article, old ATi roadmaps, and some own speculation / calculations. so take them with a grain of salt. I should also add that, in the German version, it is made very clear that at least some of the "specs" such as 256bit bus and .15um fab are considered "very questionable indeed." Pity that got somewhat lost in the translation.

ta,
.rb

 

[Tygrus_Artesiaoa]

true its just a sometin' of a rumour

true its just a sometin' of a rumour

ain't true till ya figured out past the PR as what a card says and does after testing in a system

 

[KimB]

Eight rendering pipelines in addition to 64-bit color is probably not going to happen on 107 million transistors, unless, perhaps, it only effectively has four pixel pipelines when running in 64-bit. Still, that may be rather inefficient...

 

[Typedef Enum]

The one thing that will really be interesting to watch is just how efficient the P10/R300/NV30 turn out to be...

As we know, it's not just a matter of doubling the pipelines...but doing so in an efficient manner as well.

Look back @ the GF256/GF256DDR benchmarks...

Without the bandwidth to back it up, the SDRAM version would end up tanking in 32-bit performance (vs. G400, TNT2, etc.)

 

[Pete]

So will halving the R200's pipelines mean the RV250 will be much slower at FSAA? Could ATi possibly be switching the RV250 to MSAA?

 

[Tygrus_Artesiaoa]

Hmm dude it already is...

Radeon 8500 apparently is already slow at FSAA. Looks good in quality mode, yes but slow to a crawl. Perhaps good to use in older games but nothing really current I'd say.
Anisotropic will probably take a hit as well...seeing as it could only fetch half the texels per clock then. More clocks yes but only a small amount more. 25% more clockspeed for the core as well as memory bandwidth. Still the memory bandwidth part of it is nice- take a look at the gfmx 4 440 and 460. Still since there method seems somewhat adaptive it would not be as perf killing as if they were using NV method.
It's bloody late and I don't want to hear about my sentences being out of order or poor grammar...if I was trying to be *cool* with that, then it would be annoying. It was just free flow thought typing.
Still if they cut out 2 of the pipelines and it still had hydravision going for it-it would be a drastic step up from a radeon ve/7000. They could easily make a cheap part then.

 

[Soccerman]

it may very well be that 2 pipelines are cut out.. but then it wouldn't make sense for ATi to not release NOW it if it's not competitive with the Radeon 8500 performance, it would be marketed (rightly so) as a low end part, becuase even though it would have maybe 500 MPixels/second fillrate (assuming 250mhz), the HyperZ 2 engine allows the 8500 to get alot closer to it's fillrate numbers (I think I read close to 700MPixels/second). therefor, the RV250 would be slower (assuming the RAM is unchanged, though they could always do something like the Radeon 7500 or Radeon 8500DV and have slower RAM than the core is at..

my point is, I don't think ATi would have delayed the RV250 so long if it was going to be a low end part compared to the R200. releasing it LATER only means that they wanted to allow R200 sales to continue longer (thus getting more money for the R&D that was done on it), and by the time the RV250 DOES come out, it WILL be a lower end part, compared to the R300 at least...

my question is, what ARE they going to change on the RV250, if it's not the pipes or mhz...

as for the R300, I doubt we'll see 8 pixel pipelines, or whatever the equivalent is with the VPU being so programmable (read the P10 article if you don't know what I'm talking about). unless they're taking away the dual texture units per pipe and replacing them with the pipes, it doesn't make sense, like John Carmack said. even then, I still doubt the 8 pipes spec..

 

[Doomtrooper]

The Rv 250 is a value part hence the V , people think the Rv 250 is going to be a R200 screamer are mistaken. Depending on clock speed it may be slightly faster than a current 8500 but priced alot lower with Dx 8.1 support.

 

[Hellbinder]

no simply not true, Even just this week 3rd party Ati vendors stated that the Rv250 is the successor to the 8500.

I just dont see how a 2 piped chip is going to do that. it is a 4 piped chip.

there are many, many issues with those combined rumors IMO.

 

[Doomtrooper]

Trust me...

 

[Soccerman]

umm.. well considering the TIMING of the release (which will be around when the R300 is announce or released obviously seeing as it hasn't already come out to replace the RV200), then you have to consider what parts ATi will have as their top end (R300).

if the power of the R300 is good enough to power Doom 3 at a fairly good pace, I suspect we'll be looking upon todays cores as somewhat low end.

 

[Doomtrooper]

Low end for serious gamers but high end for the general public, there really isn't a true Dx 8.1 value product on the market either, and I for one WANT the general public to own a Dx 8/ 8.1 card so we as 3D freaks and our expensive cards can finally see games being made that use these features. Finally maybe, just maybe the user base will be significant enough to justify developers to get off their ass and not be able to use the old excuse that DX7 cards are still the prominent card.

I do know that the Rv250 was never designed to be a performance part, the R300 will come in many flavors that will cover the Mid to Very high end.

 

[Dolemite]

The Rv 250 is a value part hence the V , people think the Rv 250 is going to be a R200 screamer are mistaken. Depending on clock speed it may be slightly faster than a current 8500 but priced alot lower with Dx 8.1 support.

I can't see how the RV250 would fit into the current pricing scheme if it was anything slower/cheaper than the R200/8500. With the 8500 going for around $100 and all the 7000-series models under $100 except for the 7500 AIW, there's simply no room for another "value card" in the Radeon line.

Everything I've seen prior to this 3dcenter BS says that its 4-pipe, so I see no reason to believe otherwise. With that in mind, it looks to be a slightly faster R200 considering the logical clock speed increase and any other potential improvements they've squeezed in, and should be priced accordingly -- somewhat higher than the 8500. Of course anything pricier than an 8500 will have to compete with the low end of the GF4 line, but the 8500 has been doing that fairly well anyway, so more speed at a price shouldn't hurt. I would say that the RV250-based card will be the 7500 of the 8000 series but I don't think it will offer quite the speed increase that the 7500 did over the original Radeon without a die shrink and corresponding clock speed increase.

 

[KimB]

Don't forget that only around now are games starting to demand a GeForce as a minimum requirement. It will likely be another two years yet before a GeForce3-class product is demanded as the minimum.

 

[Dave Baumann]

The quicker DX8 capable hardware is pushed into the low end the quicker developers can target DX8.

 

[Randell]

Don't forget that only around now are games starting to demand a GeForce as a minimum requirement. It will likely be another two years yet before a GeForce3-class product is demanded as the minimum.

maybe, but how many games this year run like crap at 1024x768x32bit with less than a Gf3 and 1 gig CPU? Morrowind, GTA 3 to name but 2. I reckon in the last year developers have accelerated requirements. pity that doesnt inlcude many DX8 effects.

 

[Doomtrooper]

The Rv 250 is a value part hence the V , people think the Rv 250 is going to be a R200 screamer are mistaken. Depending on clock speed it may be slightly faster than a current 8500 but priced alot lower with Dx 8.1 support.

I can't see how the RV250 would fit into the current pricing scheme if it was anything slower/cheaper than the R200/8500. With the 8500 going for around $100 and all the 7000-series models under $100 except for the 7500 AIW, there's simply no room for another "value card" in the Radeon line.

Everything I've seen prior to this 3dcenter BS says that its 4-pipe, so I see no reason to believe otherwise. With that in mind, it looks to be a slightly faster R200 considering the logical clock speed increase and any other potential improvements they've squeezed in, and should be priced accordingly -- somewhat higher than the 8500. Of course anything pricier than an 8500 will have to compete with the low end of the GF4 line, but the 8500 has been doing that fairly well anyway, so more speed at a price shouldn't hurt. I would say that the RV250-based card will be the 7500 of the 8000 series but I don't think it will offer quite the speed increase that the 7500 did over the original Radeon without a die shrink and corresponding clock speed increase.

You will see the DX7 line of cards downsize and the Rv250 take over that market...The RV250 for speed will be in the range as a Radeon Retail @ 275, faster in some cases.

These specs are VERY close..

ATi RV250:
ca 35-45 million transistors (estimated)
0.15 µ process by TSMC
256 bit architecture
2 rendering pipelines with 2 texture units each (?)
6 textures per pass
Single Vertex Shader engine
complete DirectX 8.1 compatible
Vertex Shader 1.1
Pixel Shader 1.4
128 bit DDR-RAM memory interface
Core clock: alleged at 300 MHz and more
Memory clock: probably the same as core clock = 300 MHz and more
HyperZ II
HydraVision
= a Radeon 8500 (R200) with only 2 rendering pipelines, but higher clock rates

 

[Joe DeFuria]

I can't see how the RV250 would fit into the current pricing scheme if it was anything slower/cheaper than the R200/8500. With the 8500 going for around $100 and all the 7000-series models under $100 except for the 7500 AIW, there's simply no room for another "value card" in the Radeon line.

You are looking at this from the end user perspective, not the manufacturer and seller perspective.

Some OEM versions of the 8500 might be going in the $100 range, but I can guarantee you that the profit margin on those parts will be considerably lower than a "2 pipe RV250" (about half the transistors) selling at the same price.

In other words, the RV250 would pretty much replace the 8500 LE cards, and possibly the 8500 Retail parts, depending on the exact performance / clock of the RV250 line. It's purely a move by ATI to lower the cost of the 8500 series...not a move to bring a new level of performance to the "mainstream" market.

If there's one thing we've seen from both ATI and nVidia, it's that neither of them have built a board with more than 2 pixel pipes on the 0.15 process, and targeted that part for the "value" market. 2 pixel pipes = approximately 30-40 million transistors, and that seems to be about the size limit on 0.15 to expect "value" pricing.

Now, once they shift their product lines over to 0.13, then I might expect value parts to start including 4 pipes. But that won't happen until 0.13 matures more...probably spring '03 or possibly even fall. By then, we should start seeing 4 pipe DX8 boards being sold as "value", sub $100 cards en masse.

So to the point: I don't know if RV-250 is in fact a 2 pipe, DX8 board...but it would make perfect sense to me for ATI to introduce such a part.

 

[Squidlor]

Saw this over at Rage3D and thought that some might be interested (even if it is more rumor).

ATi's PCB demarkation system (the "PN" code) follows no set pattern. They produce about original 100 items a year and every single one receives a PN code, whether it be a dongle/remote control/graphics card etc. "109" indicates a board product.

The ASIC revision is not related to the board number at all. Newer ASICS just get loaded into the same board over the course of development to incorporate changes in memory type, board layout, special vendor requests etc. The few that make it through development we see on the shelves

Computex card: "PN 109-94200-00A"

"00A" indicates an initial test version. This is incremented with each new revision. 00A>>>00B>>>etc.

"942" is the R300. It uses a 256-bit memory interface (confirmed) and is fully DX9 compatible (I guess we know that already). JIGGAHERTZ posted a good explanation regarding PCB layers, and why the board at first looks like it could use either a 128-bit (32-bit share) or 256-bit memory bus, but is in fact a 256-bit bus'ed R300 for sure:

"Signal integrity is very important for high-speed SSTL_2 signals. However, adding extra layers is just a quick fix for poor manufacturing/design. With proper PCB layout, it can accomplished. The R300 942 board is only 8-layers, I believe. The overseas companies have poor quality control, and fear they may have to resort to 10 layers, even if they do follow ATI's reference designs - but they dont want to do this as their only edge in the market is cost.

It is true that memory sometimes shares a bus. As for proof of a true 256 bit interface (which it is - no sharing) - take a look at the datasheet for the memory bga packaging. Then visualize 2 memory chips on the board. Then look at the traces (some are on internal layers) and the resistor packs (4 SMT resistors in one package) inline with the traces (serial and active memory termination). Well, since you have already spent a few hours on photoshop, you might as well spend a few more so you can verify it for yourself. Take a regular 128MB board (AIW 8500 836 etc., whatever works for you) and start counting these resistor packs. You should end up with approximately twice as many with the 942. Thats your proof.

This simply means twice as many terminated memory lines between memory and GPU. If there was any sharing, you would see the chips memory lines tied together and then routed to the GPU - only 128 bit interface."

The only other board based on the R300 in development is an AIW model (as far as my source was able to find out... there may be more). ATi are keen to retain this as their "flagship" card. 942 will be used for desktop sales only but may also surface in the FireGL line. We have seen PN 857 (the R200) pop up as the FireGL 8700 already so it has been done before.

The soldering pads on the Computex card that look like they may be for a ribbon-type connector are to facilitate the connection of a daughtercard to provide additional multimedia features. This information was published in the Inquirer not long ago and is correct (which is pretty good going for The Inquirer, lol). The daughtercards can interface with either RV250 or R300's equipped to do so. This is the reason that neither card features a Rage Theater chip... CVBS/S-VHS TV and secondary VGA output are all handled on-die.

I couldn't find out alot about the GPU internals as the source is not familiar with the specifics (pipeline/TMU config etc). The hardware development team is a fairly isolated group. The "1000 Monkeys Sitting at 1000 Typewriters" that we like to call the ATi software development team are given more than enough information (regarding registers etc) to work with, without knowing the specifics of the hardware. 8/4 (pipes/TMUs) looks likely for R300, if not then 8/2 - DX9 only calls for 16 textures per pass (I think). 2/2 makes sense given RV250's current clockspeed and performance estimates (very close to R200 in most situations). The RV250 is a "toned-down" R200 core w/TV-out and a secondary RAMDAC on die. Whether this toning down involves the removal of two rendering pipelines I do know for sure.

According to feedback on current yields:
The R300 is capable of ~<250 MHz core w/0.15u.
The RV250 is capable of ~<325 MHz core w/0.15u.

Regarding memory; supply is currently a problem (not that it isn't available, it's just that fast memory is in short supply). The R300 was designed with 400MHz memory I/O in mind. For Computex, the fastest memory available was used (350MHz DDR SGRAM Samsung BGA) - there were a few other qualification samples to be tested. By fall, 400-450MHz will be more widely available so may be seen on the production card.

PN 109-9****-*** is an RV250-based product that uses TSOP memory. PN 109-9****-*** is an RV250-based product that uses BGA memory. Cannot give you the exact numbers but they have been posted here if you were quick enough to get them before I deleted the post.

The memory on the TSOP and BGA versions of the RV250 runs at 200MHz and 300MHz respectively. This is a good indicator that ATi will phase out the 7500 and 8500 altogether (the standard models at least) although the parts may be available to 3rd parties for sometime. There may also be huge stockpiles so watch out for 7500/8500 boards dropping in price.

The main push at ATi at the moment IS to beat the NV30 to the market by a considerable margin. Orton insists on it to remain competitive and gain back some market share. Things look set for a fall release as planned.

AFAIK, all hardware engineering is completed and the code monkeys are currently slaving away on the drivers. The base code is similar to previous generations but alot of work needs to be done to efficiently use the new memory subsystem. I guess the marketing team are also getting pretty frenzied around this time, although I hear they have yet to decide on a name for either card ("Porgy" and "Bess" are current favourites).