X1700: Was'nt this supposed to appear about now?

[Killrose]

Remember reading about this 256bit vers of the X1600 quite sometime ago and thought it was to be released about now. Any word on it?

 

[BernardP]

I had also read about this on The Inquirer web site, but it was fairly speculative. Now, we know that the X1800GTO with 12 pipes will be coming out. I am wondering if this card will take the X1700's place.

 

[Extelleron]

Here's what I think ATi should come out with if they will make a X1700 series:

ATi X1700XT
MSRP: $199
Pixel Pipelines: 4
Pixel Shading Processors: 12
Vertex Shaders: 5
Core Frequency: 675 MHz
Memory Amount: 256MB
Memory Frequency: 745/1490 MHz Effective
Memory Interface: 256-Bit, 512-Bit Internal Ring Bus
Memory Type: GDDR3
Interface: PCI-Express

ATi X1700 PRO
MSRP: $139
Pixel Pipelines: 4
Pixel Shading Processors: 12
Vertex Shaders: 5
Core Frequency: 525 MHz
Memory Amount: 256MB
Memory Frequency: 475/950 MHz Effective
Memory Interface: 256-Bit, 512-Bit Internal Ring Bus
Memory Type: GDDR3
Interface: PCI-Express

 

[Munky]

If the x1800gto sells for ~$200, then it would eliminate the need for a x1700. I was assuming that the x1700 would be a 24 PS, 8 TMU derivarive of the x1900, but a 12 PS, 12 TMU x1800gto would also be a decent card in that price range, even more so if it could unlock the other quad.

 

[Extelleron]

A 4pp/12ps X1700XT with slightly higher clocks as I listed above, and a 256-Bit Memory Interface, would blow away any nVidia card in the price range, be it the 7600GT or 6800GS.

 

[crazydingo]

Originally posted by: Extelleron
Here's what I think ATi should come out with if they will make a X1700 series:

I made some corrections on what I think the X1700 will be:

ATi X1700XT
MSRP: $199
Pixel Pipelines: 8
Pixel Shading Processors: 24
Vertex Shaders: 5
Core Frequency: 600 MHz
Memory Amount: 512MB
Memory Frequency: 700/1400 MHz Effective
Memory Interface: 128-Bit, 256-Bit Internal Ring Bus
Memory Type: GDDR3
Interface: PCI-Express

ATi X1700 PRO
MSRP: $149
Pixel Pipelines: 8
Pixel Shading Processors: 24
Vertex Shaders: 5
Core Frequency: 500 MHz
Memory Amount: 256MB
Memory Frequency: 600/1200 MHz Effective
Memory Interface: 128-Bit, 256-Bit Internal Ring Bus
Memory Type: GDDR3
Interface: PCI-Express

Launch at Computex '06.

 

[Extelleron]

Originally posted by: crazydingo

Originally posted by: Extelleron
Here's what I think ATi should come out with if they will make a X1700 series:

I made some corrections on what I think the X1700 will be:

ATi X1700XT
MSRP: $199
Pixel Pipelines: 8
Pixel Shading Processors: 24
Vertex Shaders: 5
Core Frequency: 600 MHz
Memory Amount: 512MB
Memory Frequency: 700/1400 MHz Effective
Memory Interface: 128-Bit, 256-Bit Internal Ring Bus
Memory Type: GDDR3
Interface: PCI-Express

ATi X1700 PRO
MSRP: $149
Pixel Pipelines: 8
Pixel Shading Processors: 24
Vertex Shaders: 5
Core Frequency: 500 MHz
Memory Amount: 256MB
Memory Frequency: 600/1200 MHz Effective
Memory Interface: 128-Bit, 256-Bit Internal Ring Bus
Memory Type: GDDR3
Interface: PCI-Express

Launch at Computex '06.

I highly doubt that. ATi has already established a 4pp/12ps architecture with the X1600 and it would be much easier for them to simply have a 256-Bit Memory Interface and slightly higher clockspeeds (the clockspeeds I listed would be easy overclocks for X1600XT/X1600 PRO. And in fact, the performance when comparing my "guess specs" to yours isnt what you think it'd be:

MunkyMark 06
(Close to F.E.A.R. performance, 1280x960, 4x AA/8xAF, no SS

Extell's X1700XT: 34 FPS -- crazydingo's X1700XT: 30 FPS
Extell's X1700 Pro: 23 FPS-- crazydingo's X1700 Pro: 25 FPS

So "my X1700XT" would perform slightly BETTER, and your X1700 Pro would perform slightly better (only 2 FPS however) but yours would cost more, due to having ALOT more transistors with double the Pixel pipelines AND double the pixel shaders. All for around the same performance as a simple overclocked X1600 with a 256-Bit MI.

 

[crazydingo]

Originally posted by: Extelleron
I highly doubt that. ATi has already established a 4pp/12ps architecture with the X1600 and it would be much easier for them to simply have a 256-Bit Memory Interface and slightly higher clockspeeds (the clockspeeds I listed would be easy overclocks for X1600XT/X1600 PRO. And in fact, the performance when comparing my "guess specs" to yours isnt what you think it'd be:

MunkyMark 06
(Close to F.E.A.R. performance, 1280x960, 4x AA/8xAF, no SS

Extell's X1700XT: 34 FPS -- crazydingo's X1700XT: 30 FPS
Extell's X1700 Pro: 23 FPS-- crazydingo's X1700 Pro: 25 FPS

So "my X1700XT" would perform slightly BETTER, and your X1700 Pro would perform slightly better (only 2 FPS however) but yours would cost more, due to having ALOT more transistors with double the Pixel pipelines AND double the pixel shaders. All for around the same performance as a simple overclocked X1600 with a 256-Bit MI.

"Your" X1700XT will cost more than "mine" 1700XT. 256bit vs 128bit

Increase in transistors isnt as costly considering ATI went from 16 to 48 pixel shaders in just 60 million transistors. Going from 12 to 24 shaders will be around 20-30 million transistor increase.

For a gpu to be 256bit, it should be big enough for enough pin connections. I dont know how big the die size of X1700 is going to be, but 190-200 million transistors at 80nm sounds like 128bit to me.

And 128bit memory will cost significantly less than 256bit memory, duh ! bringing the overall cost of the card down even more. "My" X1700 looks very cheap and about the same performance as "your" costly X1700.

I think Pete should share some thoughts on this topic.

 

[Pete]

I've been summoned!

I haven't been keeping up very recently, but I also think 8-1-3-2/1 sounds more natural for a GPU to slot in b/w the 4-1-3-2 RV530 and 16-1-3-1 R580, especially if speed yields are good (meaning most core can do 500+MHz). The question is one of memory bus, and the answer seems to be how high GDDR4 can scale relative to the GPU. RV560 may well be 128-bit, if only for economy (I'm thinking notebooks, too).

I'm expecting cores in the range of current cards, so probably 650Hz and below, but I don't know when GDDR4 will debut and at what speed. IIRC, RV560 won't be for sale until mid-summer, if that helps with the RAM timetable. It sounded to me like the new memory controller was designed to scale for higher speeds, so 128-bit GDDR4 may be in the cards (groan ).

Remember, this is all supposition and regurgitation. I believe NV's and ATI's emphasis on SLI/Gemini may seal the deal with 128-bit (that's half the RAM chips to accomodate on a single dual-core card).

Extell, two things to consider. One, 256b RAM conxn sounds high for just 4 TMUs and 4 ROPs (remember, the 6600 feeds 8 TMUs and 4 ROPs w/128b, and I think the 7600 will feed 12 TMUs and 4 or 8 ROPs with just 128b). And both ATI and NV are probably focusing more on shader power than AA perf for anything but the "enthusiast" class. Quad-disabled enthusiast GPUs like ATI's GTO or NV's GS are probably the way to go for AA perf on the cheap. Two, if RV560 is a higher-clocked RV530, that kind of puts RV530 out of business. But I think there's rumors of a RV535 floating around, probably an 80nm RV530, so it doesn't look like that particular 4-1-3-2 config is going anywhere. In that case, it doesn't seem likely that ATI will have two 12PS/4TMU GPUs going at the same time in the same gen, especially since most sites and games don't default to testing with AA, so the 256b bus won't be as obvious as extra shader or texturing power.

But I may be wrong. I've been away the past few days, and that's a long time this close to launch.

Edit: This hints at shader power over bandwidth.

 

[ronnn]

As soon as Ati dumps all their unsold x1800 gpu's, x1700 and x1900xl will likely appear.

 

[jayanath]

u know its impossible to see 512mb card for $200

 

[apoppin]

latest old news:
RV560, X1600 re-spin to be named X1700...

As far for the mainstream part goes, RV560 is working quite nicely - the sources suggest introduction at same time as Green Goblin's 90 nanometer products: 7200, 7300 and 7600. Q1'06. The RV530 core was thoroughly modified in order to increase the performance, so it's no wonder the company has decided to call it X1700.

i with 'dingo on his suggested specs.

 

[evolucion8]

quote:

--------------------------------------------------------------------------------
Originally posted by: crazydingo

quote:

--------------------------------------------------------------------------------
Originally posted by: Extelleron
Here's what I think ATi should come out with if they will make a X1700 series:
--------------------------------------------------------------------------------


I made some corrections on what I think the X1700 will be:

ATi X1700XT
MSRP: $199
Pixel Pipelines: 8
Pixel Shading Processors: 24
Vertex Shaders: 5
Core Frequency: 600 MHz
Memory Amount: 512MB
Memory Frequency: 700/1400 MHz Effective
Memory Interface: 128-Bit, 256-Bit Internal Ring Bus
Memory Type: GDDR3
Interface: PCI-Express

ATi X1700 PRO
MSRP: $149
Pixel Pipelines: 8
Pixel Shading Processors: 24
Vertex Shaders: 5
Core Frequency: 500 MHz
Memory Amount: 256MB
Memory Frequency: 600/1200 MHz Effective
Memory Interface: 128-Bit, 256-Bit Internal Ring Bus
Memory Type: GDDR3
Interface: PCI-Express

Launch at Computex '06.
--------------------------------------------------------------------------------



I highly doubt that. ATi has already established a 4pp/12ps architecture with the X1600 and it would be much easier for them to simply have a 256-Bit Memory Interface and slightly higher clockspeeds (the clockspeeds I listed would be easy overclocks for X1600XT/X1600 PRO. And in fact, the performance when comparing my "guess specs" to yours isnt what you think it'd be:

MunkyMark 06
(Close to F.E.A.R. performance, 1280x960, 4x AA/8xAF, no SS

Extell's X1700XT: 34 FPS -- crazydingo's X1700XT: 30 FPS
Extell's X1700 Pro: 23 FPS-- crazydingo's X1700 Pro: 25 FPS

So "my X1700XT" would perform slightly BETTER, and your X1700 Pro would perform slightly better (only 2 FPS however) but yours would cost more, due to having ALOT more transistors with double the Pixel pipelines AND double the pixel shaders. All for around the same performance as a simple overclocked X1600 with a 256-Bit MI.

Well, that depends, did you remember about the upgrade from the 9600XT based X600 into the X700, different architectures, more power in the latter. But we all know that the X1600 PRO is too far from being bandwidth limited with it's 128-Bits Ringbus memory bus, so, overclocking it and adding a wider memory bus will simply add a very complex termination circuitry that may not fit in such small chip.

 

[Pete]

Shocking (to me) update! Apparently RV570 will be 256-bit. Rumors are still 12-1-3-1, tho now possibly -2 (like RV530, double Z-ops per clock).

Pretty interesting to think RV570XT's been rumored to be 10% faster than X1800XT (in 3DM, anyway). That's not going to happen by doubling the memory bus on RV530's 4-1-3-2 architecture, but now I'm left to wonder if it'll happen with 8-1-3-2 or 12-1-3-2. The question is, at 80nm, will they hit closer to 650 or 550MHz?

 

[Steelski]

Originally posted by: Pete
Shocking (to me) update! Apparently RV570 will be 256-bit. Rumors are still 12-1-3-1, tho now possibly -2 (like RV530, double Z-ops per clock).

Pretty interesting to think RV570XT's been rumored to be 10% faster than X1800XT (in 3DM, anyway). That's not going to happen by doubling the memory bus on RV530's 4-1-3-2 architecture, but now I'm left to wonder if it'll happen with 8-1-3-2 or 12-1-3-2. The question is, at 80nm, will they hit closer to 650 or 550MHz?

at 80nm, i think that 550 is way too conservative.... the X1600 XT already runs from 580-600 stock. I would really stick my head out and say 675+

I just read more closley and i read XT!!!!!!!!!!!!!!!!!! OMG. if this is an indication of the R600...i am very interested.

 

[coldpower27]

The X1700 would have to be a 12-1-3-1 part minimum on 80nm to be able to support 256Bit Wide Memory Interface, if it's 8-1-3-2 then it will almost definitely be 128Bit if it's built on the 80nm process. The cut off limit for 256Bit is 200mm2 +/- 5% judging from prior GPU's.

Think about it using current die size estimates.

90nm R580 16-1-3-1is 353mm2 right now.
80nm R590 16-1-3-1 is going to be about 234mm2 if it's a simple optical shrink.

G71 is 196mm2 24PS/24TMU/16ROP
G73 is 127mm2 12PS/12TMU/8ROP

~ 65% as big.

So a 12-1-3-1 design should be about 83% as big give or take so 194mm2 which is just around good enough for the 256Bit Wide Memory interface.

Ah 8-1-3-2 design would be down to 152mm2 on the 80nm node. This is around as large as NV43 or RV410 and those have 128Bit Wide Memory Interface. This is just my prediction however, so take with a grain of salt.

 

[Pete]

Interesting points, both. Steelski, I just didn't think ATI'd push GPU speeds with a half-step like 80nm. I assumed they'd stay within their previous envelope of 590-650MHz, or maybe within the 7600GT's 560MHz. coldpower, I'm not quite sure how you got 234mm^2 for an 80nm R580. If you just did 353*(80^2)/(90^2), I get 279mm^2. I'm also not sure why 12-1-3-1 would be specifically 83% of 16-1-3-1. Mind sharing your formula? Good point about minimum die size for 256b. I can't find one for Parhelia.

Oh, and HKEPC speculates RV560 is 8-1-2-1 and RV570 is 12-1-3-1 (Chinese link). Then again, they say they're 128-bit, which goes against The Inq's soft and Xbit's seemingly official announcements. It also seems to me to go against that slide, which shows both cores lining up with the X800 in the Performance segment. I'm thinking that segments are now defined primarily by bus width (64b at the bottom, 128b in the middle, and 256b at the top). That would make it odd (for me) if both are listed as performance but either one or both are 128b. Synthesizing (badly) lots of tidbits, maybe RV570 is 12-1-3-2 and 256b at TSMC and RV560 is 8-1-3-2 and 128b at UMC.

 

[LW07]

Originally posted by: jayanath
u know its impossible to see 512mb card for $200

http://www.newegg.com/Product/Product.asp?Item=N82E16814102630

A 512mb card for less than $200. :laugh: Although, it's not all that good.

http://www.newegg.com/Product/CustRatingReview.asp?Item=N82E16814121216

Another 512mb card for less than $200. This one is actually a decent card. Provided you get a good one as the reviews complain alot about DOAs and incompatibilities.

 

[coldpower27]

Originally posted by: Pete
Interesting points, both. Steelski, I just didn't think ATI'd push GPU speeds with a half-step like 80nm. I assumed they'd stay within their previous envelope of 590-650MHz, or maybe within the 7600GT's 560MHz. coldpower, I'm not quite sure how you got 234mm^2 for an 80nm R580. If you just did 353*(80^2)/(90^2), I get 279mm^2. I'm also not sure why 12-1-3-1 would be specifically 83% of 16-1-3-1. Mind sharing your formula? Good point about minimum die size for 256b. I can't find one for Parhelia.

Then again, they say they're 128-bit, which goes against The Inq's soft and Xbit's seemingly official announcements. It also seems to me to go against that slide, which shows both cores lining up with the X800 in the Performance segment. I'm thinking that segments are now defined primarily by bus width (64b at the bottom, 128b in the middle, and 256b at the top). That would make it odd (for me) if both are listed as performance but either one or both are 128b. Synthesizing (badly) lots of tidbits, maybe RV570 is 12-1-3-2 and 256b at TSMC and RV560 is 8-1-3-2 and 128b at UMC.

Sorry about that I was using the 110nm to 90nm conversion as a template and got 67% so that how I got my figure down so low.

Your right though, 353mm2 - 90^2mm2 / x mm2 - 80^2mm2 is indeed 279 mm2. Which is the size of the R590 Core, which is still quite large, basically the same yield level as the R520 on 90nm.

I was just assuming because bascially a G73 is half a G71 in terms of pipes and such and has a die size of 65% then the RV560/RV570 would be similar along those lines assuming it were half a R580 for the moment, so things you don't half of course like the AVIVO technology, or in Nvidia's case the PureVideo so hence you don't get a 50% die size.

100 - 65% = 35% / 2 = 17.5% so 83% if the product was 3/4 of a R580, just basically some VERY ROUGH estimates.

Let's look at it another way though, RV530 is 149mm2 and is 1/4 of a R580 bascially.

So 204mm2 of die space for 36 Pixel Processors, 12 TMU, 12 ROP's & 3 Vertex Shaders.

So 68 mm2 of die space for 12 Pixel Processors, 4 TMU, 4 ROP's & 1 Vertex Shader.

Now normalizing for the 80nm process it's 54mm2 of space on that.

So from 279mm2 take way 108 mm2 so 171mm2 die for a 24 Pixel Processor, 8 TMU, 8 ROP's, 6 Vertex Shader Part.

I think a 8-1-3-2 part on 80nm might be too small for a 256Bit WIde Memory Interface, though a 12 Pipe Part can, though that would basically mean the die would definitely be over 200mm2, which can do 256Bit Interface, but not all that profitable for ATI.

Regarding clockspeeds at 80nm, it would really depend if there is low-k dielectric on it or not, if it doesn't have it then expect it to maintain parityo clock rates, if so then there is room for higher clockspeeds.

 

[Pete]

Thanks, cp. I believe all nodes from 90nm on down will have low-k (I guess as an option, b/c I don't know if NV uses it at 90nm, or even how complicated it is to use with its existing high-k [?] designs).