Main memory, GDDR3, XDR, 1T-SRAM, whats your choice?

I would like some specifics, bandwidth, MHZ, die size(not sure if this is the correct wording), cost.

How about embedded memory, which would take up less space on a chip?

Would you go for speed over quantity for effenciency? Nevermind, I already know the answer, just thought about it, quantity would come first. Maybe both if possible.
 
DeanoC said in a old ( dont remenber wich ) thread that 512MB with 25GB of BW or 256MB with 50GB of BW he would take the second, it is more important keep this things (very parallel, like CELL) "feed".

I do not have sure about the number but it was something in those ratio
 
bandwith about 100 times bigger than size

gddr3 1000 mbit/sec
xdr 3200 mbit/sec/

512 mbit gddr3 chip has 7 GB/s with 64 bit bus
512 mbit xdr chip has 6 GB/s with 16 bit bus

4 chip has 25 GB/s = 256 MB
8 chip has 50 GB/s = 512 MB

console's CPU , GPU dont store to ram only READ , in every frame system read the full main ram , this is 60-100 * main ram bandwith

ps2: 32MB,3.2 GB/s
xbox: 64MB,6.4 GB/s
gc: 24MB,2.4 GB/s

xenon: 512MB , 52 GB/s
ps3: 512MB , 52 GB/s or 768MB , 75 GB/s
rev: 384MB, 40GB/s or 512MB , 52 GB/s
 
Well, if it was not shared memory... e.g. if the 256 MB at 50 GB/s vs. 512 MB at 25 GB/s question was in regards to the CPU alone... I'd agree. Hell, I don't really think the *code* part of things is going to demand even 128 MB. The only thing that is rather concerning is the fact that next-gen (even if not right away) is going to be loaded full of content. Massive worlds and massive numbers of textures given that things like normal mapping, parallax mapping, gloss mapping, multipass rendering will probably all be status quo... And I think no hard drives (at least not standard) is also a safe assumption, so you have to be bugged by the awful slowness that is optical drives, which are probably not going to be fast enough to stream massive quantities of data (Bluray is still about an order of magnitude slower throughput than a hard drive).

Well, I guess you can basically say that capacity and bandwidth will never be substitutes for each other.
 
And I think no hard drives (at least not standard) is also a safe assumption, so you have to be bugged by the awful slowness that is optical drives, which are probably not going to be fast enough to stream massive quantities of data

Both rumored preliminary Revolution specs. that were circulating a while back both included 15gb HDDs. E3 will show if there was any validity to these rumors.
 
pc999 said:
DeanoC said in a old ( dont remenber wich ) thread that 512MB with 25GB of BW or 256MB with 50GB of BW he would take the second, it is more important keep this things (very parallel, like CELL) "feed".

I do not have sure about the number but it was something in those ratio
Hey not me, I'd go with more slower memory every time... Low bandwidth I can work around (by just not going out to memory as much).

Low memory and I have to start cutting stuff or get crazy arse compression systems...
 
Come now Deano, just admit it, crazy arse compression systems are a lot more fun then not going out to memory as much :p
 
Brimstone said:
SiliconPipes memory controller technology crushes everything by offering 12.8 Gbits per pin. This is some of the best technology I know of that takes ordinary DDR chips and applies differential signaling like Rambus XDR has
How could it take ordinary DDR chips which do not support differential signalling, and turn them into something XDR-like? It'd be (to use a bad car analogy) trying to turn a 2-liter engine into a 4-liter engine by hammering in larger diameter pistons into the cylinders... What you must mean (as I can't be arsed to read your link :p) is this is some kind of external memory controller that hooks up to many quite ordinary DDR memory chips in parallel and then connects to an ASIC of some kind through another, separate differential signalling interface with 12.8Gbit/s/pin I/O.

Doesn't sound like a very smart idea to me, not only does one lose the benefit of an integrated on-chip memory controller, one loses the benefit of XDRs low pin-count too (as all those DDR chips still connect with the same crazy number of pins as all other DDR memory subsystems in use today). There will be the added latency of additional buffer FIFOs to the external memory controller and the parallel-to serial conversion registers on the ASIC, then the serial-to-parallel conversion on the off-chip memory controller, then ITS buffer FIFOS too, etc... It adds up to several quite unneccessary buffer stages that will introduce additional latency.
 
Guden Oden said:
Brimstone said:
SiliconPipes memory controller technology crushes everything by offering 12.8 Gbits per pin. This is some of the best technology I know of that takes ordinary DDR chips and applies differential signaling like Rambus XDR has
How could it take ordinary DDR chips which do not support differential signalling, and turn them into something XDR-like? It'd be (to use a bad car analogy) trying to turn a 2-liter engine into a 4-liter engine by hammering in larger diameter pistons into the cylinders... What you must mean (as I can't be arsed to read your link :p) is this is some kind of external memory controller that hooks up to many quite ordinary DDR memory chips in parallel and then connects to an ASIC of some kind through another, separate differential signalling interface with 12.8Gbit/s/pin I/O.

Doesn't sound like a very smart idea to me, not only does one lose the benefit of an integrated on-chip memory controller, one loses the benefit of XDRs low pin-count too (as all those DDR chips still connect with the same crazy number of pins as all other DDR memory subsystems in use today). There will be the added latency of additional buffer FIFOs to the external memory controller and the parallel-to serial conversion registers on the ASIC, then the serial-to-parallel conversion on the off-chip memory controller, then ITS buffer FIFOS too, etc... It adds up to several quite unneccessary buffer stages that will introduce additional latency.

Apparently SiliconPipe has some ex-Rambus workers and this technology is going to compete against XDR. While the chipset and DIMM slots remain the same, a change to the way the motherboard is manufactured has to be done along with putting the memory controller on the board. How much cost this add's I don't know. Also I don't think any technical hurdles exist to prevent GDDR being used with this approach.

It does seem very similar to Intel FB-DIMM, so it will probably cause some increase in latency. DDR-II compared to XDR at 3.2 Ghz, Rambus XDR probably has a performance advantage latency wise. But GDDR is optomized for a really high clock speed and seeing how they compare would be intresting. XDR at 3.2 is the dram is clocked at 400 mhz and GDDR-3 the dram is clocked in the 1 Ghz range.



In general, the faster data can be retrieved, the faster the host processor can execute its algorithms. Internal parallelism on the processor side requires ever-faster data-transfer rates from external memory to keep parallel execution units active. In the DRAM world, this has translated into the move from single-data-rate synchronous DRAMs (SDRAMs) to double-data-rate (DDR) SDRAMs down one path and to the use of Rambus' DRAM (RDRAM) down a second path. A third fork for specialized low-latency applications, such as that found in network subsystems, has produced a reduced-latency DRAM (RLDRAM) that's optimized for real-time packet-handling subsystems.

Already, the first-generation DDR interface, which peaked at 333 to 400 MHz and memory densities of 512 Mbits (although at least one company is sampling 1-Gbit devices), is giving way to the second-generation version, DDR II, which offers data-transfer rates starting at 400 MHz. Over the next two years, DDR II data rates will peak at 667 to 800 MHz, and available memory densities will increase from today's 256- and 512-Mbit devices to 1 Gbit/chip. A third-generation DDR interface will offer still higher data rates. However, samples probably won't be available until late 2004 or early 2005.

Yet many designers still look for even faster data-transfer rates. Answering that demand, several companies developed new memory-interface architectures that promise to boost transfer rates to more than 6 Gbytes/s. Later this year, Toshiba and Infineon Technology both expect to sample DRAMs based on Rambus' extreme-data-rate (XDR) interface.

The XDR approach requires a full revamping of the DRAM interface and a new point-to-point memory subsystem that may take a while to be accepted by mainstream systems. Another approach, slated for introduction this quarter by Silicon Pipe Inc., employs a controller chip that sits between the host processor and the memory array and multiplexer chips that connect the controller to the memory array. This scheme, which can use standard DDR memory DIMMs, achieves throughputs of up to 12.8 Gbytes/s.


.....

LOOK FOR A HIGH-SPEED MEMORY interface from Silicon Pipe as an intriguing alternative to the XDR interface. The novel interface design employs a controller that converts the parallel bus of a host system to serial interfaces. In turn, the interfaces connect to small multiplexer circuits that talk to standard DRAM modules.

http://www.elecdesign.com/Articles/ArticleID/7035/7035.html


Here is the other article so you can read it.



Avoiding the need to change the native memory interface, a novel controller and buffering scheme pushes data-transfer rates for standard SDRAMs to 3.2 Gbits/s/pin and up to 12.8 Gbits/s/pin for first-generation double-data-rate (DDR) SDRAMs. A controller and multiplexer/demultiplexer buffer make up this two-chip solution.

Silicon Pipe's Chaniplexer chip set, targeted at desktop and server systems, can use standard dual inline-memory modules (DIMMs). For applications demanding higher performance, the company's Seriplexer requires a redesigned DIMM to achieve data rates of 24 Gbits/s per pin and beyond.

In a Chaniplexer system, a controller connects to the host processor or DRAM controller's address/data bus and converts the parallel data into multiple differential serial channels. The multiple channels are then connected to the multiplexer/demultiplexers, which are positioned between the DIMM connector and the differential signal traces (see the figure).

The differential signals are more immune to noise than standard parallel buses. This adds more flexibility in the pc-board layout with respect to the DRAMs' proximity to the memory controller. Though some pc-board modifications are required for the existing memory-controller data path, no changes are needed to the host processor or chip set, or to the memory DIMMs.

Based on the Grand Canyon differential signaling scheme, serial channels virtually eliminate signal disruptions between a memory controller and its memories. The serial channels can transfer bit streams at data rates starting at 3.2 Gbits/s and scaling up to 8 Gbits/s and beyond, depending on the memory interface. When used with existing DIMMs and system logic, the Chaniplexer can implement systems with word widths of up to 128 bits and support as many as 32 DIMMs. When used with DDR DIMMs, data rates start at 12.8 Gbits/s and can scale to over 20 Gbits/s.

The Seriplexer scheme requires custom DIMMs optimized for the approach. With it, data transfers should hit 24 Gbits/s and higher. Silicon Pipe says that the Seriplexer has virtually no scaling limits, so it can work in systems that concurrently employ different memory technologies.

http://www.elecdesign.com/Articles/ArticleID/7169/7169.html



Figure_01.gif
 
pc999 said:
DeanoC said in a old ( dont remenber wich ) thread that 512MB with 25GB of BW or 256MB with 50GB of BW he would take the second, it is more important keep this things (very parallel, like CELL) "feed".

I do not have sure about the number but it was something in those ratio

pc999 is right I did say that here... http://www.beyond3d.com/forum/viewtopic.php?p=319583&highlight=#319583

but I think I've changed my mind. At least for this week...
 
Acert93 said:
DeanoC said:
pc999 is right I did say that here... http://www.beyond3d.com/forum/viewtopic.php?p=319583&highlight=#319583

but I think I've changed my mind. At least for this week...

:oops: I hear rumblings in the force! Does this mean there may be some last minute changes... not that you would know, so anything you say would be pure speculation and therefore our welcome ears would just love to speculate right along with you ;)

The gods of console hardware design will do whatever they want ignoring my musing, I just meant which I personally would prefer change on a weekly basis...

It depends on whether I'm feeling techy (more bandwidth is cool because everything will be optimised etc.) or gamey (more RAM is good because it means we spend more time on the gameplay rather than compression systems...)
 
Acert93 said:
DeanoC said:
pc999 is right I did say that here... http://www.beyond3d.com/forum/viewtopic.php?p=319583&highlight=#319583

but I think I've changed my mind. At least for this week...

:oops: I hear rumblings in the force! Does this mean there may be some last minute changes... not that you would know, so anything you say would be pure speculation and therefore our welcome ears would just love to speculate right along with you ;)

He did not want to confirm that Xbox 2/Xenon was upgraded to 512 MB of RAM in a UMA configuration ;).


Well, willing and acting are nice subjects ;).
 
We all know Deano really wants 1Gbof ram with 104GB/s of bandwith.
If so, then I'd pretty much concur (unless that lowercase "b" is intentional). 1 GB will surely deal with the crazy content demands unless your name is Tim Sweeney. And 104 GB/sec shouldn't be anything to complain about bandwidth-wise... although how the theoretical relates to the practical limits is still questionable.
 
Apparently SiliconPipe has some ex-Rambus workers

And they're making an XDR-like technology? And Rambus isn't supposed to sue them?

BTW, how about the quad channel ddr thing that VIA was working on about a year or so ago?
 
Fox5 said:
Apparently SiliconPipe has some ex-Rambus workers

And they're making an XDR-like technology? And Rambus isn't supposed to sue them?

BTW, how about the quad channel ddr thing that VIA was working on about a year or so ago?


Nononono you don't understand. Rambus would sue you for saying the name Rambus.
 
Back
Top