|  | | Nvidia Desperate for 55nm G200 | 
One point four billion. That's how many transistors the NVIDIA G200 core consists of. If you take that many transistors and try to make a chip out of them with a manufacturing process limited to 65nm lanes, you're going to get a lot of faulty of chips per wafer. Yields were suggested to be low 40% before the launch and that the chip would become extremely expensive, $100-110. Even though this sounded a bit extreme, it might actually have been even better than reality.
The fact is that many NVIDIA partners have been left out after G200, simply because of a lack of chips. And the partners that do get cards are not happy with the margins. NVIDIA is trying hard to get the 55nm "G200b" out as soon as possible, but it doesn't look like it will happen anytime soon. G200 at 55nm should be about 470mm2, which should make room for about 120 dies on a 300mm wafer. Hopefully this will push yields closer to 50%, but NVIDIA is probably going to be satisfied with above 40%.
You can read the entire article at Nordic Hardware. | | 20 Comments | | | man that's rough... it's gotta look like a bee hive over there now, everyone scrambling and working over time to get something pushed out... | | | | It is said that NVIDIA may cancel the G200b project and work on the release of the new G3x0. It will integrate 384SP and it will use 45nm technology. Its power consumption will be lower than the G200. Its clocks will be improved to 850/2000MHz. It will match 1GB 256bit GDDR5 memory and the clock will be more than 4000MHz, yielding far better results than the 4870X2 | | | | Quote:
Originally Posted by Unregistered  It is said that NVIDIA may cancel the G200b project and work on the release of the new G3x0. It will integrate 384SP and it will use 45nm technology. Its power consumption will be lower than the G200. Its clocks will be improved to 850/2000MHz. It will match 1GB 256bit GDDR5 memory and the clock will be more than 4000MHz, yielding far better results than the 4870X2 | That would be a wise course of action but this still begs a question of timeline? Will they get it out before ATI's next Iteration. | | | | you can't have 1GB of memory on a 256bit bus, each Module interfaces at 32bit and is 64mb in size. increasing the size of the module while maintaining the same interface would cause congestion issues.
the G3x0 will be a 512bit card with 1GB DDR5 memory. | | | | Bits per clock cycle has a lot to do with it though as well ,bus width is not everything..... | | | | Actually thats where you are wrong.
The bandwidth of memory is calculated via base speed * DDR * module count * 4 (modules have a 32bit interface and 32bits = 4bytes)
The less modules, the less bandwidth.
256b GDDR5 actually requires more then 2x the speed of 512b GDDR3 to perform at the same Framebuffer output. so not only does the extra memory give more performance because theres more space, it also gives more performance because theres more modules to send data to.
you can clock a 6600GT and 5900Ultra the same, however, the Utra will always have better performance when Antialiasing is applied, because it has faster framebuffer access and write performance. | | | | yes , 32bits = 4 bytes and yes 256 bit would require 2 x the speed of 512 to perform the same framebuffer output per request per clockcycle. However this does not mean it can't be done(as stated in your original post) , it just means it can be clocked higher than the Gddr3 model and whether or not it is bottlenecked or not would still yield better results. I am not saying a 512b bus would not be better, however i am saying you are wrong about using 1gb on a 256bit bus if it is GDD5. | | | | what i said is you CANNOT HAVE 1GB OF DDR ON A 256BIT BUS.
Ram modules do not come in 128 pieces. | | | | But you can , it's only bus width and not indicative of how much data flow and is on the bus... just how much data flow per clock cycle, increase the clocks....like gddr5 and you increase the data flow....... 256bits/per clock cycle increase the clock cycles though to double (doubtful to get that out of gddr5 but you get the point) you would still see an increase in performance of a 1 g model versus a 512m model or a gddr3 model at any speed increase..... 1g of gddr5 would work on a 256bit ring bus......like i said it not to say you would be bottlenecking at the bus compare to a 512 but it is not as bad as you have made it out to be.....and certainly you can have 1GB on a 256bit bus. | | | | no you can't. face it, get over it, and move on. | | | | ok ,whatever man.... but here are the stats on the 4870......and i hate to hurt your feelings but.....1 gig of memory on a 256bit bus.....just cause nv don't do it don't mean it is isn't possible.....
link http://en.hardspell.com/doc/showcont.asp?news_id=3747
sapphires version.....can you read what the memory size is?? how bout the memory interface?? 256 bit you say? http://www.sapphiretech.com/us/produ...gpid=250&grp=3
if you don't believe these links i can give you a few more.....i am not wrong and now i am over it...... |
Last edited by n3omatrix; August 31st, 2008 at 11:54 AM..
| Quote | | | | | There's even a 1GB 4850 on newegg here.
Now I haven't seen it benchmarked but based on your theory here it should take a performance hit on lighter loads compared to the 512MB model right? | | | | comment from site that unixlord posted:
Made a big difference in newer games jumping from a 512MB 4850 to this one with 1GB. GRID framerates completely smoothed out at 1680x1050 where I was getting choppiness on the 512MB version. The 4850 is the best bang for buck right now. Keeps up with the 280 GTX that I have in my main system in many situations
Good example Unixlord......squall face it man,get over it, and move on. |
Last edited by n3omatrix; August 31st, 2008 at 12:38 PM..
| Quote | | | | | Quote: |
The GDDR3 Interface transfers two 32 bit wide data words per clock cycle from the I/O pins. Corresponding to the 4n-pre fetch a single write or read access consists of a 128 bit wide, one-clock-cycle data transfer at the internal memory core and four corresponding 32 bit wide, one-half-clock-cycle data transfers at the I/O Pins. Single-ended unidirectional Read and Write Data strobes are transmitted simultaneously with Read and Write data respectively in order to capture data properly at the receivers of both the Graphics SDRAM and the controller. Data strobes are organized per byte of the 32 bit wide interface.
| according to a inside source, which is backed up by wikipedia Quote: | Hynix Semiconductor has introduced the industry's first 1 Gib GDDR5 memory. It supports a bandwidth of 20 GB/s on a 32-bit bus, which enables memory configurations of 1 GiB at 160 GB/s with only 8 circuits on a 256-bit bus.[6] | 128MB GDDR modules have only recently gone into production (rather silently). However, addressing 128MB of memory across 32bit is slower then addressing 64MB over 32bit. This is the same reason why System DDR controllers have used 64bit interfaces for years, and why Dual Channel controllers are so important not.
Theres no doubt, that bumping the interface to 512bits and using 16x64 modules would provide a much more amusing speed improvement. As for now, this is simply a Frankenstien card, the memory controller is not designed to address 128mb modules, and The limitations will quickly show up. |
Last edited by squall_leonhart; August 31st, 2008 at 01:10 PM..
| Quote | | | | | Quote:
Originally Posted by squall_leonhart according to a inside source, 128MB GDDR modules have only recently gone into production (rather silently). However, addressing 128MB of memory across 32bit is slower then addressing 64MB over 32bit. This is the same reason why System DDR controllers have used 64bit interfaces for years, and why Dual Channel controllers are so important not.
Theres no doubt, that bumping the interface to 512bits and using 16x64 modules would provide a much more amusing speed improvement. As for now, this is simply a Frankenstien card, the memory controller is not designed to address 128mb modules, and The limitations will quickly show up. |
That was not what i was disputing and i NEVER disagreed with you on the fact that a 512 bus would be better.........i don't believe it is any more frankenstein than the 280....but you did say it was flat out can't be done..... that was the only arguement i had, and that GDDR5 would make a serious difference in memory speed on a 256bit bus. Posted by squall_leonhart on August 29th, 2008, 12:35 AM
you can't have 1GB of memory on a 256bit bus, each Module interfaces at 32bit and is 64mb in size. increasing the size of the module while maintaining the same interface would cause congestion issues.
the G3x0 will be a 512bit card with 1GB DDR5 memory. Posted by squall_leonhart on August 30th, 2008, 04:14 AM
what i said is you CANNOT HAVE 1GB OF DDR ON A 256BIT BUS.
Ram modules do not come in 128 pieces.  It is nothing personal ,you seem to be a very intelligent guy ,i hope we can walk away friends......... |
Last edited by n3omatrix; August 31st, 2008 at 01:24 PM..
| Quote | | | | | Quote:
Originally Posted by n3omatrix That was not what i was disputing and i NEVER disagreed with you on the fact that a 512 bus would be better.........i don't believe it is any more frankenstein than the 280....but you did say it was flat out can't be done..... that was the only arguement i had, and that GDDR5 would make a serious difference in memory speed on a 256bit bus. Posted by squall_leonhart on August 29th, 2008, 12:35 AM
you can't have 1GB of memory on a 256bit bus, each Module interfaces at 32bit and is 64mb in size. increasing the size of the module while maintaining the same interface would cause congestion issues.
the G3x0 will be a 512bit card with 1GB DDR5 memory. Posted by squall_leonhart on August 30th, 2008, 04:14 AM
what i said is you CANNOT HAVE 1GB OF DDR ON A 256BIT BUS.
Ram modules do not come in 128 pieces. It is nothing personal ,you seem to be a very intelligent guy ,i hope we can walk away friends......... | 1GiB GDDR3 is new, so new theres not even any information on it yet. There wasn't a big advertisement of 1GiB GDDR5 either. Can't say either of us are wrong since you had information that i didn't lol. | | | | Lol , eh well ,it was a fun arguement while it lasted.......so how bout them (enter cities hockey team here)? | | | | I'll keep an eye out for a in depth review of the 1GB GDDR3 4850 but as I recall there have been 1GB GDDR3 8800GTs for quite a while. | | | | Wow, that
ASUS GeForce 8800GT 1GB 256-bit looks awesome | | | | Quote:
Originally Posted by Unixlord I'll keep an eye out for a in depth review of the 1GB GDDR3 4850 but as I recall there have been 1GB GDDR3 8800GTs for quite a while. | I stand corrected....i guess Nvidia has one as well....nice card though. | |
|
