| SRAM vs. DRAM |
| May 1, 2002, 9:32 PM CST by pat777 | | Which is better for supercomputing overall? I think it depends on the situation. If you need more memory I say DRAM. If you need fast memory I say SRAM but I don't know which is better overall for supercomputing. |
|
| May 1, 2002, 10:51 PM CST by pat777 | | Anyone? |
|
| May 2, 2002, 12:51 AM CST by Pichu | SRAM uses six Transistors per bit whilst DRAM uses one transistor (BJT, MOSFET whatever) and two capacitors per bit.
The capacitors in DRAM are used to store the signal temporarily. Once in a while, does that those capacitors require to be recharged, so this slow down overall memory performance. Thus, the word dynamic is derived because the data is dynamically refreshed. Since SRAM uses six transistors, all the data are stored permanently when there are electric flow, so the memory speed will be a lot faster than DRAM because the data doesn't need to be refreshed periodically (the period is dependent of the capacitance). That is how static comes about.
Since capacitor in silicon (simple one is like two plates) can be a lot smaller than a transistor (PNP junction etc), SRAM will be in general six times larger than DRAM.
Supercomputing requires memory crunching and a vast amount of memory, yes? If there are 128MB DRAM chips available, the biggest SRAM that you can get is only 20MB. I can't exactly give you the answer, but I think SRAM is waste of silicon, so DRAM is a better choice. You could, however, use SRAM as cache. |
|
| May 2, 2002, 4:12 AM CST by khaid | I put some thought into it.
The first thing I want to do if I ever get my hands on a supercomputer, is play a good game of nibbles. |
|
| May 2, 2002, 7:35 AM CST by David_South | There are three other types of SRAM to concider now, (Two new types of DRAM,) and one all together new type still coming.
3.2ghz Rambus SRAM
EDIT: The correct name is: 3.2GHz RDRAM.
SigmaRAM for high speed network and servers.
When I was interviewing Sony's rep for this form of RAM I asked if this might appear in home use some day. He gave a good answer. ;D
(GameCube) 1T-SRam: this is the unknown freak of SRAM. There still hasn't been an independant bench to show how much faster the enitial responce of this is over other SRAMS.
Still in the works at IBM is MagRAM.
A very stable form of ram that stores information by shifting the polar alignment of small magnetic bits within the ram. This means that power loss with not hurt the stored information.
Don't most servers use SRAM?
I understand the size issue, But isn't it much faster after the ential latentcy then DRAM. Right?
Where does this eDRAM & VideoRAM stuff fit in performance?
Is VRAM even a real type of RAM, or just what they call it when it runs Video memory? |
|
| May 2, 2002, 11:37 AM CST by cd36 | 1T-SRAM isn't really SRAM its a cross between DRAM and SRAM more towards DRAM i think
and its not just for the GCN its made by the company MoSys the GCN just used it |
|
| May 2, 2002, 11:40 AM CST by Pichu | What he just said are only examples of DRAM. They are not Static Random Access Memory.
Servers depend heavily on Static RAM. They used it as cache, like 8MB L2 cache or higher. |
|
| May 2, 2002, 4:07 PM CST by chairmansteve | 1T-SRAM is SRAM only by name (marketing). It requires the same refreshing as any DRAM. MoSys claims latency comparable to SRAM by "hiding" the refreshing. But SRAM can still have a much higher clock speed. Would be nice to have a PC system with 1T-SRAM DRAM, so benchmarks can be run -- to compare to other DRAM.
eDRAM = embedded DRAM = the DRAM and processor are on the same chip.
VRAM or VideoRAM depends on how the term is used. It can either mean any RAM (usually DRAM) that is used for video storage, or VRAM can refer to the VRAM standard that has the ability for the RAMDAC and video processor to access memory simultaneously. |
|
| May 2, 2002, 5:06 PM CST by khaid | | eHAMW = embedded hamster & wheel? |
|
| May 2, 2002, 5:23 PM CST by chairmansteve | | Yes, eHAMW™ technology will revolutionize the computer industry. |
|
| May 2, 2002, 5:23 PM CST by Chipaku | here's a little info from MoSys on 1T-SRAM from their website. http://www.mosys.com/product_set.html
(to get more info you need to register and they will email the info.)
MoSys’ patented 1T-SRAM technology offers a combination of high density, low power consumption, high speed and low cost unmatched by other memory technologies. The single transistor bit cell used in 1T-SRAM technology results in the technology achieving much higher density than traditional four or six transistor SRAMs while the innovative 1T-SRAM architecture allows the use of standard logic manufacturing processes without requiring any process changes. 1T-SRAM technology also offers the familiar, refresh-free interface and high performance for random address access cycles associated with traditional SRAMs. In addition, this technology can reduce operating power consumption by a factor of four compared with traditional SRAM technology, contributing to making it an ideal technology for embedding large memories in SoC designs. 1T-SRAM technology is in volume production both in stand-alone memories available from MoSys as well as in SoC products from MoSys’ licensees.
1T-SRAM Characteristics
Silicon-Proven
Qualification programs completed; products in volume production.
Ultra-High Density
1T-SRAM provides 2x to 3x the density of traditional embedded SRAM.
Standard Process
No special processes required; uses standard CMOS logic process.
Lower Power
1T-SRAM can reduce power consumption to one-fourth of traditional SRAM.
SRAM Performance
1T-SRAM delivers sustained random access cycle times to below 3ns.
High Yield
Standard process, small area & built-in redundancy give enhanced yield.
Low Testing Cost
High speed and MoSys proprietary sector redundancy reduce test costs.
SRAM Interface
Standard interface for high performance and ease of design.
Macro Sizes
Custom macros can range from under 1 megabit to over 50 megabits. |
|
| May 2, 2002, 5:47 PM CST by chairmansteve | Dang, copying over MoSys marketing stuff? They make it seem like they are making something faster than SRAM. 1T-SRAM is more like DRAM than SRAM.
Look at this: http://www.mosys.com/news/idsoc.pdf
Those types of DRAM with "transparent refresh" are sometimes called PSRAM (pseudo-SRAM), but they use DRAM cells. Some other companies make PSRAM too. |
|
| May 2, 2002, 5:58 PM CST by David_South | Thanks guys. Btw I knew Nintendos use wasn't exclusive.
I just couldn't remember the details (like even the name-RAM).
Does anyone know about the SigmaRAM for servers and this new MagRAM?
>>What he just said are only examples of DRAM. They are not Static Random Access Memory.<<
What did you mean by this Pichu?
I'm thought Sigma, Rambus, & T1-SRAM (T1 I now know better) were all SRAM memories.
There are also two new types of DRAM in the work. I really don't know anything about them. One I thought was Quad clocked, QDR-RAM (Quadruple Data Rate RAM). Then there is a 400mhz FSB DDR-RAM in the works. We have just made the move to the triple stuff, DDR333. But these other two have been lightly mentioned in my presence.
Any clues or info would help, and be much appreciated. |
|
| May 2, 2002, 6:14 PM CST by Pichu | I have to do a search to clarify my statement. Yes, they are all Dynamic RAM (DRAM), not SRAM. If you've taken an electronic course, you'll understand how SRAM and DRAM are designed or you could just read my post #3 to understand what is the real difference between SRAM and DRAM.
SigmaRAM: http://www.sigmaram.com (it appears to be SRAM - not very clear though as I have to read the datasheet in depth)
1T-SRAM (still DRAM or maybe Pseudo-SRAM is possible)
RAMBUS http://www.rambus.com (a company name - it has no relationship with the memory modules or if you refer to RDRAM, then it's acceptable. RAMBUS doesn't even make SRAM chips) |
|
| May 2, 2002, 6:19 PM CST by Pichu | | Yeah, SigmaRAM is SRAM, but the size is very small (wow, they got 18Mb SRAM just 2.25MB SRAM!) |
|
| May 2, 2002, 7:04 PM CST by Chipaku | I never said it wasnt DRAM, I just paste that since you guys didnt know what it was. I already knew what it was.
supposedly it has SRAM performance at DRAM price, thats whats good about it. |
|
| May 2, 2002, 7:12 PM CST by pat777 | | MoSys said there's eSRAM. Thats got to be fast. I heard that SRAM can be instant. |
|
| May 2, 2002, 7:14 PM CST by chairmansteve | Well, SigmaRAM is meant for telecommunication and network data buffering. Perhaps large amounts aren't needed in those fields.
Dave, thatsalotta terms you got there. DDR and QDR can apply to any memory, but normally those refer to variations of SDRAM (Synchronous DRAM) to increase bandwidth.
Then there is a 400mhz FSB DDR-RAM in the works. We have just made the move to the triple stuff, DDR333.
FSB and RAM are separate. FSB is data bus from CPU to northbridge -- sometimes called system bus. Maybe you mean DDR400 and DDR333? That's 400MHz (used in Xbox) and 333MHz DDR-SDRAM -- actually 200 and 166, but doubled.
Chip, I was talking about the info you posted from MoSys, not what you said -- nowhere in that MoSys marketing stuff does it say it's DRAM. They making it seem like some super-duper memory. If it's better than standard DRAM, that'll be good enough. I doubt the performance is like SRAM. |
|
| May 2, 2002, 7:16 PM CST by chairmansteve | | pat, that's probably embedded 1T-SRAM (used in GameCube's Flipper GPU), which is a type of embedded DRAM (eDRAM). |
|
| May 2, 2002, 7:16 PM CST by Pichu | | by the way, what does "1T-SRAM" stand for? 1 transistor "SRAM"? Hmm... |
|
| May 2, 2002, 7:19 PM CST by chairmansteve | | Yep, 1T means one-transistor. |
|
| May 2, 2002, 7:22 PM CST by Pichu | | that explains. duh, DRAM also uses one transistor! I'm sure that it's patented, so I might be able to look at the design to see if it's really Static RAM. |
|
| May 2, 2002, 7:46 PM CST by chairmansteve | MagRAM looks interesting. I read about it a couple years ago and did a quick search right now.
Magnetic RAM, eh? Present SRAM and DRAM are electronic memories, while present hard disks and floppy disks are magnetic mediums and CD/DVD are optical. MagRAM has the promise of replacing present ROM, RAM, HDD, and FDD technologies. I used "present" instead of "current" to avoid a pun (current and electricity). :)
Could be good stuff, but probably a long way from being used widely. |
|
| May 2, 2002, 7:54 PM CST by chairmansteve | And for anyone interested in general difference between RDRAM (Rambus) and SDRAM:
RDRAM - serial (allows higher bit rate)
SDRAM - parallel bus
What improvement does DDR-II SDRAM and DDR-II+ SDRAM add to allow higher bit rates (clock speed) over regular DDR-SDRAM? Should I look it up myself, ye bastids!? :) |
|
| May 2, 2002, 8:10 PM CST by cd36 | | i've never really understood whats the difference between a single channel and a dual channel RAM? |
|
| May 2, 2002, 8:20 PM CST by chairmansteve | Dual channel uses two channels! :)
I assume you are using the term correctly and don't really mean DDR, yes?
Dual channel is like having two memory paths that work together. It would kind of be like having two FSBs with two CPUs, or like two AGP buses with two GPUs. That making the sense? |
|
| May 2, 2002, 8:28 PM CST by cd36 | | so dual channel RDRAM would have one channel going to one memory module and the other channel to the other and thats why RDRAM has to be used in pairs? |
|
| May 2, 2002, 8:42 PM CST by chairmansteve | Yep, that's why RDRAM are used in pairs with dual-channel systems.
I think it's also possible to have dual channel with a single memory stick, if the memory is designed for that purpose (i.e. memory module designed to accept two channels). |
|
| May 3, 2002, 2:17 AM CST by Gman | Has anyone heard the news.
Today i read in my Maximum Pc mag, and it says,that a memory company is developing "Qdr ram".Qdr ram will be 2x faster than ddr ram .Qdr ram would also be used in a regular ddr slot.I have no other details on QDR ram,maybe Hank or Steve can break down the info,but i do know you can use QDR ram,in your ddr slot.
So i guess not the slot will be changed,but the memory module.Thats neat!!!! |
|
| May 3, 2002, 2:31 AM CST by khaid | QDR = Quad Data Rate
Uses 2 DDR banks on a single DIMM? |
|
| May 3, 2002, 3:44 AM CST by chairmansteve | Well, I just looked into DDR-II, and I see it's the same thing as QDR-SDRAM. So that's how DDR-II allows increased bit rates per pin. Actually, I remember that about DDR-II now from reading some articles many months ago, but I suppose I misplaced that piece of info and needed some refreshing (damn DRAM!).
You sure MaxPC said QDR will be used in the same DDR slots out now? From what I've seen, DDR-I can be used in the new DDR-II boards, but not the other way around -- and it'll only be DDR-I modules that are designed to work in the DDR-II boards. Thus, the DDR-I modules out now wouldn't be compatible.
Anyhow, being able to use QDR in a current Athlon DDR board would be worthless -- it would be throwing money away (just like dual channel memory on nForce without using the IGP).
QDR (DDR-II) could be useful for video cards and Pentium 4 systems (high FSB). |
|
| May 3, 2002, 3:53 AM CST by chairmansteve | There is also QBM (Quad Band Memory) that also doubles DDR bandwidth, but works differently from DDR-II. QBM uses existing DDR-I memory.
http://www.quadbandmemory.com |
|
| May 3, 2002, 5:55 AM CST by Pichu | | it's possible that they both utilize the same banks. Anything beyond 184 pins is just superfluous. Remember also that Via P4X600 suppose to support QDR as well as DDR-II and DDR? Well, it might have been the same case for those chipsets with SDR and DDR SDRAM support... |
|
| May 3, 2002, 12:25 PM CST by khaid | What pichu is talking about.
VIA P4X600 chipset for the Pentium 4 processor is expected to be released in Q2. The P4X600 will be based around the P4X333, but is expected to support dual channel DDR SDRAM (previous rumours suggested QDR SDRAM support instead (Quad Data Rate, implemented by operating 2 banks of DDR SDRAM - on a single DIMM - at a 90deg phase to each other), but this is unlikely). The VIA 8235 South Bridge features support for USB2.0 and ATA133. |
|
| May 3, 2002, 3:48 PM CST by chairmansteve | Speaking of MagRAM or MRAM, Cypress Semiconductor recently formed a subsidiary to work on MagRAM development.
Cypress Semiconductor Corp. has formed a subsidiary to develop magnetic RAMs using technology licensed from a start-up company located in Eden Prairie, Minn.
The creation of Silicon Magnetic Systems in San Jose was disclosed today at Cypress' annual meeting. The product line has the potential of developing into a $100 million business over the next few years, said T.J. Rodgers, chief executive officer.
Cypress will be competing with a number of companies, including IBM, Motorola and Infineon, to commercialize large density MRAMs. "They don't exist yet, and nobody's quite sure how to make them exist," Rodgers said. "I believe it's going to work and I believe we'll be first to make it work."
http://www.ebnews.com/story/OEG20020502S0040
|
|
| May 4, 2002, 4:05 PM CST by pat777 | | Bump |
|
| May 5, 2002, 8:46 PM CST by pat777 | | How many transistors does gamecube's 1T-SRAM use? |
|
| May 5, 2002, 8:52 PM CST by chairmansteve | Flipper = 51 million transistors (around half or 25.2 million of those are for the 3MB embedded 1T-SRAM)
For the 24MB 1T-SRAM system memory, that's a little over 200 million total (not all on one chip).
For memory, transistors and size are related. |
|
| May 8, 2002, 6:47 AM CST by David_South | Thanks for the MagRAM news.
Here's a little from me.
"Low latency network DRAM."
(I'm following this because I believe this type may appear as eDRAM for the PS3.)
http://www.eetimes.com/printableArticle?doc_id=OEG20020506S0026
MagRAM in the works.
http://www.eetimes.com/printableArticle?doc_id=OEG19990817S0088
http://www.eetimes.com/printableArticle?doc_id=OEG20010110S1009
And a very general look at RAM in the news.
http://www.eetimes.com/search?action=FilterSearch&site=2465&hidesummary=false&filter=o_filter.hts&queryText=RAM&sort=match
You should read page two of the results. Ferrolmagnetic RAM. |
|
| May 8, 2002, 9:38 AM CST by pvcHH | | I don't understand any of this. I'm just going to add a post to get one closer to not being "Frag Bait" anymore ... The post will be clever though. Ready? QDR-SDRAM is A-OK 2 me. Aren't I funny? |
|
| May 8, 2002, 10:08 PM CST by pat777 | | I can't believe it! Nintendo could of made the Flipper a lot better if they took away the 1T-SRAM and made something out of the 25 million transistors(vertex and pixel shaders, and more DDR Memory). Hey Steve you said that the speed is only useful for rare situations. The other 200 million transistors could be use for DDR memory and an extra GPU. No wonder PCs don't use eDRAM. |
|
| May 8, 2002, 10:13 PM CST by Gman | | How come pc's dont use eDRAM PAT??? |
|
| May 8, 2002, 10:23 PM CST by Pichu | ...
Hey Steve you said that the speed is only useful for rare situations. The other 200 million transistors could be use for DDR memory and an extra GPU. No wonder PCs don't use eDRAM.
200M transistors = 200Mbits = 25MBytes of Memory. Wait, I thought common systems already equips with 512MB of DDR Memory, which uses more than 4.2 Billion Transistors, distributed in sixteen memory chips, so each chip already has 268 Million transistors. PC don't use eDRAM because there is just no point in using it.
How come pc's dont use eDRAM PAT???
Where do you want to put your eDRAM? Every CPU (Pentium-4) already equip with 512kB of L2 SRAM cache, which uses 25 million transistors in the processor. |
|
| May 8, 2002, 10:25 PM CST by Gman | | REally....Thats neat,Thanks Pichu |
|
| May 8, 2002, 10:26 PM CST by chairmansteve | vertex and pixel shaders
It would be possible.
and more DDR Memory
Nope. Know the difference between embedded memory and external memory? If the memory is not embedded, then it doesn't occupy transistors on the processor (Flipper in this case).
Hey Steve you said that the speed is only useful for rare situations.
Where did I say that?
The other 200 million transistors could be use for DDR memory and an extra GPU. No wonder PCs don't use eDRAM.
That 24MB 1T-SRAM system memory on GC is not embedded on any processor. DDR SDRAM memory uses the same number of transistors as 1T-SRAM. However, I think 1T-SRAM takes up about 1.1x to 1.2x as much space as DDR SDRAM (I'd have to verify that though). |
|
| May 8, 2002, 10:27 PM CST by chairmansteve | | Gman, you didn't know that processors have cache? |
|
| May 9, 2002, 12:17 AM CST by David_South | I know about three levels of cache. L1 L2 and L3(Banks) then there is 'write cache' that is stored on the hard drive.
What is write thru cache? Is that the same as the previous?
The one thing I want most is optimized bandwidth L3.
A set up where a ball grip RAM module plugs into a HyperTransportGate that is a fourth the MegaHz of the CPU.
I also want to see PCs or Motherboards that accept ball grip GPUs.
If upgrading was as simple as a Processor Chip, pricing would be better.
What I haven't been able to figure is the 'cache size' chips need.
Presently the largest eDRAM caches are about 18mb, on chip. unless you've heard different.
Would more on chip cache slow down sreaming funtions?
Less cache will take more time for large functions,
it has to access the next memory level, slower level.
I came up with some trial formulas but no clear pattern for how they decided it. |
|
| May 9, 2002, 12:52 AM CST by Pichu | DS:
I know about three levels of cache. L1 L2 and L3(Banks) then there is 'write cache' that is stored on the hard drive.
What is write thru cache? Is that the same as the previous?
All I know is that cache is faster than the transfer rate of the device/chip/whatever. If cache were slower, than would it be wise to implement caches?
The one thing I want most is optimized bandwidth L3.
A set up where a ball grip RAM module plugs into a HyperTransportGate that is a fourth the MegaHz of the CPU.
I also want to see PCs or Motherboards that accept ball grip GPUs. If upgrading was as simple as a Processor Chip, pricing would be better.
From GPUs to GPUs, the pins layout aren't even compatible with one another. Take Geforce SDR and Geforce2 GTS, Geforce 3 and Geforce 4. They all have different pins. The only ones that kept the same are Geforce 2/GTS, Geforce 2/Pro, and Geforce 2/Ultra, Riva TNT and TNT2, etc. But as we have discovered before, the speed difference from Geforce 2/GTS and Geforce 2/Ultra isn't too greatly as we compared between Geforce 3 and Geforce 2. Again, other chips use their own designs, i.e. Radeon 8500 and Geforce 3's pins are completely different.
What I haven't been able to figure is the 'cache size' chips need.
Presently the largest eDRAM caches are about 18mb, on chip. unless you've heard different.
Is it 18MegaBytes (MB) or 18 Megabits (Mb)? Last we heard that there's 6.5MB (52Mb) SRAM using 0.09 microns fabrication. eDRAM is suppose to be six times smaller than SRAM, so that would equivalent to about 408Mb or 51MB of eDRAM.
Would more on chip cache slow down sreaming funtions?
Less cache will take more time for large functions,
it has to access the next memory level, slower level.
I came up with some trial formulas but no clear pattern for how they decided it.
Actually, the answer is YES. Take the following scenario for examples:
In the old days, for Socket 7 systems, each mainboard has its own 256kB, 512kB, 1024kB, and 2048kB of SRAM off-die L2 cache. Since these cache are off-die, this would mean that the L2 cache is slow, but comparing to the internal bandwidth, these caches are still relatively fast. Here's the problem. Comparing a 1024kB L2 cache system to a 256kB L2 cache system, the 1024kB system beats 256kB in most benchmarks. However, if the cache is increased to 2048kB, the performance degrades. Why? It simply means that these SRAM started to get very slow at processing large amount of data. Therefore, faster cache is desired. Also, there's another problem with Pentium-II XEON 2MB L2 cache as these large amount of L2 off die cache doesn't seem to help a lot in comparison to on-die cache of Coppermine. Anyhow, it's better to build all L2 cache on die to have faster cache, which will advent in Galatin cores (Pentium-4 XEON).
Analogously, if the processor and FSB speeds continuously to increase and the SRAM cache size continuous to raise, then don't you want a very fast cache to be able to process the data from all the places in the cache? Otherwise, slower SRAM = smaller cache. |
|
| May 9, 2002, 12:57 AM CST by chairmansteve | Yep, L1, L2, and L3 cache are three levels of cache. Itanium and upcoming Xeon chips have some L3 cache -- meant for servers. AMD K6-III systems had L3 cache on the motherboard. There can be L4, L5, etc. too, but only the necessary levels of cache should be used (2 levels for current desktop system).
Write Cache is a buffer for temporary storage before writing to the hard disk, so that the number of times that memory is written to the hard disk can be decreased. Read-Ahead Cache is similar, but for the opposite direction (reading from hard drive).
Write Thru Cache is something else. It's a method of how write operations are handled. With write-thru cache, writes go immediately to memory/disk -- cache and memory/disk will have the same data. With write-back cache, writes are stored temporarily in a buffer and then transferred to memory/disk later when the system is ready or when necessary -- cache and memory/disk won't always have the same data. Hard Disk Cache is usually a type of Write-Back Cache.
Ball Grip? Do you mean Ball Grid Array (BGA)? |
|
| May 9, 2002, 1:30 AM CST by chairmansteve | Last we heard that there's 6.5MB (52Mb) SRAM using 0.09 microns fabrication. eDRAM is suppose to be six times smaller than SRAM, so that would equivalent to about 408Mb or 51MB of eDRAM.
How you got 408Mb? 52Mb * 6 = 312Mb. Did some other bits sneak in from the back door? 68Mb * 6? |
|
| May 9, 2002, 1:42 AM CST by David_South | Yes, I meant Ball Grid Array (BGA).
& Yes, IBM has .013u eDRAM = 1MB > 18MegaBytes sizes.
EDIT: http://www-3.ibm.com/chips/techlib/techlib.nsf/techdocs/025B939A31655A4687256B910056E8C1/$file/emb_dram_cu11_00.pdf
Here is a fact sheet on the new eDRAM from IBM. They plan on using this in games later on in their road map. I have this technology bookmarked on the chance that they mean the PS3.
The name was something like CU_11 eDRAM.
I'll try to remember my Capitals(MB)when I do that again.
BGA, It seems to me like this would be needed for memory banks to interface at about 1/4 or 1/2 of future CPU clocks.
It just seems like a better memory interface. Also with the creation of MagRAM or FerrolRAM, this seems like the best way to have it connect to the board.(Only because I like the pin grid pattern & it seems like it might work better access wise.)
I've seen that Rambus created , 3.2GHz RambusDRAMs test sticks.
But it just doesn't seem right to me (i.e. it's a little to complicated for me). |
|
| May 9, 2002, 1:50 AM CST by chairmansteve | BGA has been used for a while. GeForce4 cards come with BGA DDR memory. Graphics processors and mobile CPUs also use BGA connection.
That 3.2GHz Rambus memory is really 400MHz with Octal Data Rate (ODR), using Rambus' codename Yellowstone technology. |
|
| May 9, 2002, 2:53 AM CST by David_South | Whooops. I meant to add the new information as a seperate post.
It is now part of the above edit.
Since you let my error alone in your reply it shouldn't matter. Thanks!
I never new that BGA was in use for the GF4. Does this mean I can swap memory sizes? |
|
| May 9, 2002, 3:08 AM CST by Pichu | | no, it's just the way that the pins are layout, which is defined by BGA, PGA, etc. |
|
| May 9, 2002, 3:12 AM CST by David_South | Ball Grid Array? I guess I meant Ball Pin Grid Array.
There was a demo a while back where I saw memory that socketed like a CPU.
PGA? is that Plastic or Pin for the 'P'. |
|
| May 9, 2002, 6:14 AM CST by chairmansteve | Dave, so you want BGA sockets on PC motherboards? BGA itself does not imply a socket, but it only means the chip uses solder balls for connection instead of leads (pins) like in PGA (Pin Grid Array).
BGA chips used on current PC hardware are mounted onto the board and are not meant to be removed. Only the main processor on a desktop uses a PGA ZIF socket, so chips can be replaced. |
|
| May 9, 2002, 6:19 PM CST by pat777 | | Steve, you said that extra memory speed is only useful in rare situations in another thread. |
|
| May 9, 2002, 6:29 PM CST by chairmansteve | Pat, if you can find what you are talking about, I can maybe clarify. You might be taking something out of context, or maybe misunderstanding.
Was it about small amounts of eDRAM on PC GPUs? |
|
| May 11, 2002, 11:45 AM CST by kenneth9265 | so for us less than degree-like geeks....what would you all say is
future standard for a high proformance machine for playing games
on you PC.....DDR2/3 or RAMBUS? |
|
| May 11, 2002, 6:03 PM CST by Pichu | 32 bit RAMBUS is getting very interesting, improving the bandwidth of the old 16 bit rambus twice. And, there will be 64 bit RAMBUS. As for DDR-II/QDR, the speeds will increase slowly when comparing to CPU speed.
For example,
Pentium 133MHz uses SIMM and sometimes PC66 DIMM. We'll standardize everything to DIMM.
Now, we got Pentium-4 2.533GHz, which is roughly about 2533/133 or nineteen times faster than Pentium 133MHz, okay let's say 20 times faster. Now, consider the memory. Let's pick the faster ones, say RAMBUS PC1066 dual channel or 4.2GB/s. PC66 has only 533MB/s, so dividing the two will give you 4200/533.333 = 7.875 times faster or just eight times faster.
Over the past six or seven years years,
CPU has increased 20 times faster.
Memory has increased 8 times faster.
In conclusion, if DDR/SDRAM is continued to be used, the speed increment will be much lower than the processor (GPU, VPU, CPU), thus some architectural enhancements are requried to be made, like LMA II, etc. This enhancement is not necessary for early developments of the GPU though. Even the eDRAM, which is about roughly five times faster than the current memory, will increase very slowly while the processor speed continue to increase rapidly. Then, after a few years, eDRAM will be too slow to catch up the processor speed just like the DDR memory now. Thus, what is the point of shifting to eDRAM immediately since it doesn't really show any necessity??? Of course, it is still required to focus on architectural enhancements now before it's too late. So, stop whining to get the current fastest memory available as it will serve no real purpose two years later. |
|
| May 11, 2002, 10:30 PM CST by kenneth9265 | dig! so for the next five years or so you would recommend RAMBUS over DDR2/3,
at least till the "memory gods" come up with something better? |
|
| May 11, 2002, 11:04 PM CST by Pichu | Actually, I won't recommend anything just yet. Have you ever seen the benchmarks of PC3200 (DDR400) vs. Dual Channel PC800 RDRAM (3.2GB/s) on a Pentium-4? Well, it was shown that PC3200 is somewhat faster than the dual channel PC800 RDRAM even though they have the same speeds.
In addition, PC1066 has only a bandwidth of 2.1GB/s, similar to DDR266 (PC2100), and the 32 bit RDRAM will have a bandwidth of 4.2GB, comparable to DDR-II at half the bus size (32 bit vs 64/128 bit). Furthermore, you're just comparing two types of memories, RDRAM and SDRAM - both functioning a little bit different. Perhaps, MagRAM is something interesting to be seen since as technology improve, the performance will improve drastically. |
|
| May 12, 2002, 12:43 AM CST by kenneth9265 | | when do you think MagRam will come available to the pulblic? |
|
| May 13, 2002, 2:13 AM CST by David_South | Hey Kenneth9265,
I was just thinking of you yesterday.
"Wonder if Ken is still alive?" Guess so.
Hey, don't be shy about speaking up.
How long for MagRAM?
Not for a few years I think (my guess: 2 years).
Even though I love the idea of stable system memory it will be a while.
Most of the ‘spintronic’ theory is a set for the creation of it, but the engineering of manufacturing technology and production reliability needs work.
Right now it takes a lot of custom lab work to create a small functioning test piece.
Here's an earlier MagRam.
Crap!!! The file isn't responding.
Oh,well.
Along time ago, way back in the year 1956 there were two scientists (one Asian, one American) that created "Core Storage" RAM.
Core storage consisted of 1/16 inch magnetic donuts on a wire network that looked like a crochet weave.
Each donut could be magnetically polarized by running electricity thru the corresponding wires.
A few short years later IBM bought all of the research and license for $500,000 dollars.
A hefty price back then.
This technology was then used by the military and early space flight programs for critical memory storage.
I'm also told that in some form or another this type of RAM memory is still being used by the government.
Sorry about the file but those are the main highlights. Have a good day.
I hope to lookup the original backup and add some of the other memory stuff I've read. |
|
| May 13, 2002, 2:42 AM CST by David_South | There are three memories I plan to look up and learn more about.
GRAM (Graphic RAM)
VRAM (Video RAM)
WRAM (Window RAM)
I've listed them in increasing performance speeds(?).
Or at least this is the way I remember reading them.
The referance article I'm recalling is from the April 2000 'Computer Desktop Encyclopedia'.
It only mentioned each for about two line to say this is faster than that.
If anyone knows more about the them let me know.
Also I found that there is another type of "E" DRAM.
eDRAM is 'embedded', but EDRAM is 'enhanced'.
Does anyone know the responce time/cycle for the 'enhanced' one.
SRAM is something like 10-30(nano/pica?)seconds and DRAM is the greater than 30 from what I recall.
Then BiPolar(?) is less than 10****?seconds.
With recent advances I'm curious to know just how fast DRAM is getting.
Again I'm a little fuzzy on the details because this is all recalled from casual reading. |
|
| May 13, 2002, 12:29 PM CST by Pichu | Actually, there's no GRAM, not what I've heard since 1990.
There's SGRAM (Synchronous Graphics Random Access Memory) and SDRAM (Synchronous Dynamic RAM), and of course these are different due to how they handle the calls. |
|
| May 14, 2002, 5:26 PM CST by David_South | P.S. Each magnetic donut was only a single bit of info or 'core storage'.
So you can imagine how big this would get with mutiple layers and such.
Think metal frames, with small cuban bead seats, as screens.
Stacked or stood in layers.
It was huge and heavy. |
|
| May 14, 2002, 5:28 PM CST by Pichu | | Magnetic = Electromagnetic Emission? And this can cause cancer. I hope that they have proper shield in these new devices. |
|
| May 14, 2002, 5:44 PM CST by David_South | This was back in the late 50s and up until the 70s.
I don't think they cared or knew. |
|
| May 14, 2002, 6:08 PM CST by Pichu | | you're wrong. Cellular phones have been very bad to the health without proper shield, and they already cause brain tumour, a.k.a Cancer. Perhaps, the emission isn't that strong in these cores. |
|
| May 14, 2002, 7:06 PM CST by David_South | Well many people have different feelings about cellphones.
The reason they are a danger is that they generate tons of EM pulses to send messages.
With the cores they are just magnetic washers on a screen.
The amount of EM field comming from them shouldn't be any more significant than a Magnatite boulder.
P.S. Do you know what the responce times are for the differant memories now-a-days?
I don't have any real world times for this stuff. |
|
| May 14, 2002, 7:09 PM CST by khaid | http://khaid.hypermart.net/uglyface.gif
This is your face after use of cell phones. Any questions? |
|
| May 14, 2002, 8:39 PM CST by Pichu | | khaid, did you just use too much of your cell phone? |
|
| May 14, 2002, 8:42 PM CST by khaid | | You know, it's funny that you ask that. I've been text messaging more on my cell phone than making voice calls. I guess I'm safe from the cancer? |
|
| May 14, 2002, 8:54 PM CST by Pichu | | hehe... I was suppose to ask you did you just talk too much... but then I changed the wording, thinking that you might have used it for other purposes. ^^ |
|
| May 28, 2002, 11:27 PM CST by pat777 to Pichu | Is 1T-SRAM like having 6 logical transistors in one physical transistor?
Note:I wasn't replying to Pichu. I really wanted to click on add reply. |
|
| May 29, 2002, 12:26 AM CST by pat777 | | Anyone? |
|
| May 29, 2002, 12:28 AM CST by chairmansteve to pat777 | | I was about to answer, but since you bumped the thread, I changed my mind. :) |
|
| May 29, 2002, 12:29 AM CST by pat777 to chairmansteve | | I didn't bump but I did something like a bump. |
|
| May 29, 2002, 12:30 AM CST by chairmansteve to pat777 | | If it walks like a bump and talks like a bump, it's a bump. |
|
| May 29, 2002, 12:47 AM CST by khaid to pat777 | | I'll bump ya on the head if you're saying that wasn't a bump. |
|
| May 29, 2002, 12:47 AM CST by chairmansteve to pat777 | | What would logical transistors be? 1T-SRAM is a type of DRAM. It has 1 transistor per cell and requires regular refreshing. |
|
| May 29, 2002, 12:57 AM CST by pat777 to khaid | | I was just joking. |
|
| May 29, 2002, 1:00 AM CST by pat777 to chairmansteve | | Never mind. I was thinking about Intel's hyper threading technology and figured that transistors are the smallest part of an computer and you can't have more than one logical transistor in a physical. |
|
| May 29, 2002, 1:06 AM CST by chairmansteve to pat777 | | HyperThreading provides two logical processors on one physical processor. It requires extra transistors for this ability. It doesn't create two logical transistors out of one physical, in case you were thinking so. |
|
| May 29, 2002, 1:20 AM CST by pat777 to chairmansteve | | I wasn't thinking that hyper threading did that. I was thinking that it might be possible to have a similar technology to do the same for transistors. |
|
| May 29, 2002, 1:26 AM CST by chairmansteve to pat777 | It's called "make the transistors smaller." Then, two of the new can fit in the place of one of the old.
Maybe quantum computing is what you're after. I'm not sure of the details, but I think it allows several states per transistor. Perhaps in a way it can be considered as having several logical transistors. I've never heard it being called that though. |
|
| May 29, 2002, 1:27 AM CST by Pichu to pat777 | Note:I wasn't replying to Pichu. I really wanted to click on add reply.
Thanks, I guess that you don't want me to join the discussion, so I won't. |
|
| May 29, 2002, 1:53 AM CST by pat777 to Pichu | | I mean't I wasn't only replying to you. |
|
| May 30, 2002, 1:56 AM CST by David_South to pat777 | Dude you are having a hard time without any help.
Let me know if that changes. ~ ^'^ |
|
| May 30, 2002, 2:03 AM CST by Gman | | Isnt HyperThreading and HyperTransport like the same technology? |
|
| May 30, 2002, 2:05 AM CST by chairmansteve to Gman | Is that what a co-worker said? No, they are not the same. They aren't even similar -- other than the names.
HyperThreading = multi-threading on single processor
HyperTranport = point-to-point link |
|
| June 19, 2002, 7:18 PM CST by David_South | The Price for Memory.
We haven't compared with real numbers:
-the prices of SRAM vs DRAM?
-latency speeds & other performance differences?
-the ability, use, or sizes for embedding these memories?
Extra News.
U.S. investigates DRAM makers 'anticompetitive practices'
http://www.nytimes.com/reuters/technology/tech-tech-micron.html?pagewanted=print&position=bottom
How many remember ultra low prices that made 1GB of RAM seem reasonable for any system?
I do, & boy was I angry when all I had bought was 512MB. |
|
| June 19, 2002, 11:00 PM CST by Gman to chairmansteve | So whats the diffrence between my bus(Muti-Tol=1.2gb/s) Hypertransport=800mb/s?
Wouldnt my SIS Multi-TOL Chip be more faster and effecient? |
|
| July 4, 2002, 5:03 PM CST by pat777 to David_South | | 512mb is enough. |
|
| July 4, 2002, 8:51 PM CST by Pichu to pat777 | | Actually, I think I'm running out of memory. My system is sometimes a little bit sluggish with 512MB of RAM. I always exceeded, for some reason. Well, I'm going to wait for 1GB DIMM to drop the price further. |
|
| July 4, 2002, 9:00 PM CST by pat777 to Pichu | | For a proffesional(which I think you are)I would say at least 1 GB. For an average gamer 512 MB should be enough. |
|
| July 5, 2002, 7:06 PM CST by pat777 | | I'm leaning torwards DRAM because you can have a lot of it. |
|
| © 2000-2005 pcvsconsole.com |