Dial-A-Rate MEMORY

[MadRat]

Would it be possible to make a signal controller to control data transfers from DDR-II equivalent modules to operate as SDR, DDR, or DDR-II upon command?

My feelings are that it would be impressive to see the CPU be able to control whether it needs SDR, DDR, or DDR-II like transfers. The latency could be narrowed to the absolute minimum if the memory operated in SDR mode by default, but kicked into the other two modes as necessary. The CPU would basically decide whether to turn on the afterburners for raw bandwidth when necessary. I seriously doubt many people would bitch about dual-channel SDR200 @1-1-1 settings if it was available for AMD chips rather than the DDR400 @2.5-2.5-2.5, being DDR @2-2-2 vs. SDR @1-1-1 was a 15-20% increase in real bandwidth but ONLY a 3-7% gain in most real world performances. Being able to slice the latency down at default to its minimum then ramp transfer speed (cost = higher latency in order to boost transfer rate) would be pretty effective way to get the most of both worlds.

 

[Sahakiel]

One question :

Why would SDR vs DDR affect how fast a memory cell operates? As far as I know, most of the latency increases comes from the physics of the memory cell itself and how it is accessed. In other words, changing whether or not data is transferred on the falling edge or not does not change the access latency. The only thing that changes would be the burst latency between each data transfer.

 

[MadRat]

DDR-II has four times the latency of SDR.

 

[Sahakiel]

Originally posted by: MadRat
DDR-II has four times the latency of SDR.

Hm...I didn't know that. I would have thought DDR-II would be advertised as having 1/4th the latency of SDR of the same clock frequency.

 

[Sahakiel]

Hm... I just checked Samsung's spec sheets for DDR-II, DDR, and SDR at 133 MHz all 512Mb. If I'm reading the data correctly, DDR-II actually has a lower latency than DDR and SDR at any speed. Of the few latency specs I could find, DDR-II is faster by 5 - 10 ns. DDR and SDR both have the same latency ratings.

 

[MadRat]

Link? As far as I've read, its not any lower latency. To be more geniune, DDR-II\533 should be compared to SDR @533MHz and DDR @266MHz.

 

[Sahakiel]

512Mbit DRAM
512Mbit DDR DRAM
512Mbit DDR-II

 

[andreasl]

Originally posted by: MadRat
DDR-II has four times the latency of SDR.

No it does not. The latency should be pretty simular since they run about the same speed at the core. There is no such thing as SDR running at 533MHz. It cannot run that fast, that is why they went to DDR and now DDR-II. It's like saying Intel should never have bothered with the P4 and clocked the P3 at 2GHz instead. But everyone knows the P3 could not run at anything close to 2GHz.

 

[MadRat]

I saw a 4-bit prefetch for DDR-II, but nothing to lean towards it being better latency-wise. I guess I'm going to have to ask you to translate what I missed.

 

[andreasl]

Originally posted by: MadRat
I saw a 4-bit prefetch for DDR-II, but nothing to lean towards it being better latency-wise. I guess I'm going to have to ask you to translate what I missed.

SDR has CAS latency of 2-3 cycles at 133MHz
DDR has CAS latency of 2, 2.5 or 3 cycles at 133-200MHz
DDR-II has a CAS latency of 4, 5 or 6 cycles at 200-333MHz (-400MHz??)

4 cycles latency at 266MHz is the same amount of time in nanoseconds as 2 cycles latency at 133MHz.


EDIT: Fixed a typo in cycletime for DDR-II

 

[MadRat]

So there is nearly no overhead for DDR, let alone DDR-II? I have a hard time believing it. High-performance SDR has always outperformed high-performance DDR in terms of quickness when it comes to same internal clock speeds, but at a cost of losing the race in raw bandwidth. DDR and DDR-II are the bandwidth winners and SDR will always be the raw quickness winner.

From what I see is that DDR-II will have twice the latency at the same internal speed at a minimum, and it uses raw clock advantage to overcome this overhead. Well that same clock advantage when given to SDR would then make SDR look even better in terms of quickness against DDR-II then... I also sincerely hope that nobody thinks that SDR cannot run at the exact same frequencies as the internal clocks of DDR and DDR-II. We aren't talking RDRAM vs. SDRAM here, we're talking SDRAM vs. SDRAM.

 

[andreasl]

Originally posted by: MadRat
So there is nearly no overhead for DDR, let alone DDR-II? I have a hard time believing it. High-performance SDR has always outperformed high-performance DDR in terms of quickness when it comes to same internal clock speeds, but at a cost of losing the race in raw bandwidth. DDR and DDR-II are the bandwidth winners and SDR will always be the raw quickness winner.

From what I see is that DDR-II will have twice the latency at the same internal speed at a minimum, and it uses raw clock advantage to overcome this overhead. Well that same clock advantage when given to SDR would then make SDR look even better in terms of quickness against DDR-II then... I also sincerely hope that nobody thinks that SDR cannot run at the exact same frequencies as the internal clocks of DDR and DDR-II. We aren't talking RDRAM vs. SDRAM here, we're talking SDRAM vs. SDRAM.

What is it about this that is so hard to understand? The reason they introduced DDR to replace SDR was because SDR could not be clocked higher and still be manufactured with acceptable yeilds. The memory manufacturers operate on razorthin margins because the DRAM memory is a commodity. Faster DRAM does exist but it is custom designed and not tailored for mass market.

Latency can be expressed in two ways. Either in nanoseconds or in clock cycles. The important metric is nanoseconds. In nanoseconds there is not much difference in latency between SDR, DDR and DDR-II. In terms of clock cycles there is a difference between DDR and DDR-II. Namely what I listed above. DDR-II base frequency is twice that of DDR so the CAS latency becomes twice as many cycles when they are using the same DRAM core.

The latency numbers I listed for DDR-II are the numbers from the official specification. If you don't believe them, take it up with JEDEC

 

[MadRat]

Back in 2000 the review sites were hitting over 220MHz with the SDRAM on bleeding edge TNT2U-based graphics cards. In 2001 the videocards were pushing 260MHz with SDRAM. You can't tell me that SDRAM couldn't make over 300Mhz today for the commodity market today if they moved the RAM closer to the memory controller and shrunk the packaging's overall layout like they do for video. (Video memory, because it requires much stricter timing, is located right next to the memory controller - totally unlike in the system!) Desktop SDRAM was pushing 166MHz back in the day when the comparable DDR was struggling to do 133Mhz, making me wonder if the overhead for DDR isn't killing SDRAM's improved latency as the clocks rise. From what Tomshardware seems to say about DDR2 performance, its really nothing to write home about with DDR-II's higher latency. Looked like the 4-bit prefetch was its lone banner to really brag about.

 

[Sahakiel]

Originally posted by: MadRat
Back in 2000 the review sites were hitting over 220MHz with the SDRAM on bleeding edge TNT2U-based graphics cards. In 2001 the videocards were pushing 260MHz with SDRAM. You can't tell me that SDRAM couldn't make over 300Mhz today for the commodity market today if they moved the RAM closer to the memory controller and shrunk the packaging's overall layout like they do for video. (Video memory, because it requires much stricter timing, is located right next to the memory controller - totally unlike in the system!)

I was just about to place bets on when you'd be bringing up that topic again...
I don't seriously doubt the ability of legacy SDRAM to run at 300MHz on a video card. Of course, that would be the limit of the technology. However, I hope you realize that SDRAM at 300MHz will get thoroughly beaten by DDR at 300MHz on any decent video card. Perhaps the only video cards to really not see much benefit would be budget cards but for the cost of 300MHz SDRAM it's much cheaper to get 150MHz DDR and a few percentage points less performance or even 200MHz DDR and a few percentage points better performance.

Desktop SDRAM was pushing 166MHz back in the day when the comparable DDR was struggling to do 133Mhz, making me wonder if the overhead for DDR isn't killing SDRAM's improved latency as the clocks rise. From what Tomshardware seems to say about DDR2 performance, its really nothing to write home about with DDR-II's higher latency. Looked like the 4-bit prefetch was its lone banner to really brag about.

Repeat: DDR-II latency is equal or better than SDRAM. SDRAM doesn't have improved latency. Latency at this point seems to be a function of process technology. Decreases in latency at the same density is the same as increasing clock speeds with the same lithography. Both memory formats benefit, but DDR and DDR-II also have the added bonus of raw bandwidth clock for clock.

New info : the 4-bit prefetch is merely another way of saying QDR. DDR SDRAM data sheets list 2-bit prefetch because, you guessed it, DDR transfers data on the rising and falling edge of the clock. Essentially what it means is that for DDR-II, for every clock period for the memory cell, four bits of data are transferred across the output pin. The DDR-II standard doubles DDR output per clock cycle by doubling the speed of the output logic and also transfers data on rise and fall for an effective 4x increase in bandwidth vs 2x. The memory cells still run at the same speed regardless, it so happens that more than one cell is accessed during one clock for each pin. DDR memory cells run at the same speed as standard SDRAM and have the exact same latencies.

 

[MadRat]

DDR-II and DDR is just big memory makers attempt to recycle old tech. They push ahead the transfers per clock one generation and move two steps back in clock rate each time. They did it with EDO (70ns 5v 168-pins from 50ns 5v 72-pins; recycling 50ns, 60ns, and 70ns), they did it with DDR(PC166 @166MHz to PC1600 @100MHz; recycling 100MHz, 133MHz, 166Mhz), and now they'll do it with DDR-II. Lest we forget we've been past the 100MHz mark in memory a couple of times already...

DDR-II\533 isn't hardly different than DDR400, so why would it be 4x better than PC133?? DDR wasn't but 7-15% different in raw streaming performance at the same internal clock speeds, so why would DDR-II suddenly be so much more efficient? The truth is that DDR-II will at best be around 50% efficient, no where near as compared to SDR's mid 70-ish percent efficiency. DDR was only in the mid-50's, another reason its impossible to say DDR is actually double the performance; at 133MHz SDR is around 600MB/sec and DDR is around 1GB/sec. When it came to asynchronous reads and writes the SDR memory had an advantage.

SDR, DDR, and DDR-II may use identical SDRAM technology, but the controllers for each one are different. Interlaced SDR memory was running pretty close to DDR performance, but with better quickness. DDR could read and write relatively close to SDR, but there was a slim difference between the two. That and it took DDR much longer to catch up to the speeds of SDR on a per clock basis; SDR was at PC166 when DDR266 was just coming to market. DDR-II is going for dual-channel and DDR-II's QDR-style of transfers to boost their Prescott performance in the Alderwood/Grantsdale chipsets if I understand right, and the reviewers were given PC4200 when the initial mainstream rollout will likely be only PC3200. I seriously doubt AMD would want PC3200 in any of their machines!

Raw bandwidth is good for processes that demand bandwidth, like multimedia processing, but for performance computing its not a benefit. Dial-A-Rate MEMORY would span the technologies to bring about the best of each as necessary.

 

[Sahakiel]

Originally posted by: MadRat
DDR-II and DDR is just big memory makers attempt to recycle old tech. They push ahead the transfers per clock one generation and move two steps back in clock rate each time. They did it with EDO (70ns 5v 168-pins from 50ns 5v 72-pins; recycling 50ns, 60ns, and 70ns), they did it with DDR(PC166 @166MHz to PC1600 @100MHz; recycling 100MHz, 133MHz, 166Mhz), and now they'll do it with DDR-II. Lest we forget we've been past the 100MHz mark in memory a couple of times already...

Strangely enough, you seem to know more about the history of memory than I do yet you learned nothing from it.

DDR-II\533 isn't hardly different than DDR400, so why would it be 4x better than PC133?? DDR wasn't but 7-15% different in raw streaming performance at the same internal clock speeds, so why would DDR-II suddenly be so much more efficient? The truth is that DDR-II will at best be around 50% efficient, no where near as compared to SDR's mid 70-ish percent efficiency. DDR was only in the mid-50's, another reason its impossible to say DDR is actually double the performance; at 133MHz SDR is around 600MB/sec and DDR is around 1GB/sec. When it came to asynchronous reads and writes the SDR memory had an advantage.

Hmm.... perhaps you should take another look at how the efficiency is calculated. If it is done the way I think it is, then DDR is more efficient overall than the legacy SDRAM it replaced. DDR transfers two bits per clock, but at 50% efficiency, that would be at least one bit of valid data per clock. Legacy SDRAM at 70% efficiency would be the same as saying 30% of the time, SDRAM came up short and sent nothing useful. 50% efficiency at twice the speed is a lot better than 70% at the reference speed.

SDR, DDR, and DDR-II may use identical SDRAM technology, but the controllers for each one are different. Interlaced SDR memory was running pretty close to DDR performance, but with better quickness. DDR could read and write relatively close to SDR, but there was a slim difference between the two. That and it took DDR much longer to catch up to the speeds of SDR on a per clock basis; SDR was at PC166 when DDR266 was just coming to market. DDR-II is going for dual-channel and DDR-II's QDR-style of transfers to boost their Prescott performance in the Alderwood/Grantsdale chipsets if I understand right, and the reviewers were given PC4200 when the initial mainstream rollout will likely be only PC3200. I seriously doubt AMD would want PC3200 in any of their machines!

Um.. I think you forget AMD is still using PC2100 and PC2700. PC3200 is still high-end for AMD.
Anyways, legacy SDRAM never made it to PC166. Sure, you could overclock parts to 150 or 166 and sell it as such, but you could do the same with DDR.
Also, DDR controllers, if I remember correctly, seriously sucked. It took how long before a decent controller hit the scene? Legacy SDRAM controllers had been on the market for years and the technology was well understood. In other words, comparison of mature vs new technology.
I have a feeling the reason interleaved legacy SDRAM did about as well as the DDR counterparts is because internally, DDR is interleaved into 2 banks minimum.

Raw bandwidth is good for processes that demand bandwidth, like multimedia processing, but for performance computing its not a benefit. Dial-A-Rate MEMORY would span the technologies to bring about the best of each as necessary.

Well, stupid me, in that case, higher density hard drives must not improve performance at all.

There are two ways to increase microprocessor performance : reduce latency or increases bandwidth. Seeing as how latency isn't increasing anytime soon (on the order of 10% / year for the past twenty years) it looks like bandwidth is the way to go.
I see a couple major problems with your Dial-A-Rate memory idea. First off, very little benefit for a lot more cost. The percentage gains are at best on the order of 10% in best case scenarios, but the cost of extra logic on the memory alone is going to cost money and, more importantly, latency.
Second problem involves system design. I don't recall any CPU which can specify which memory type it wants to run. Doing so destroys compatibility with other systems and future memory types. You'll end up with early generations running only a couple memory types while the new CPUs mop the floor with new memory formats. It's like supporting old ISA slots and with about as much benefit.

 

[andreasl]

MadRat, do you really think you are smarter than the hundreds of engineers working for the DRAM companies? You seem to have problems grasping even the most fundamental issues of most things you are talking about. IMHO you should take a more humble approach to technology and stop pretending you know everything. Ask questions instead of making grand statements.

There is a difference between memory in graphic cards and for PCs. For PCs you stick the memory chips on modules and connect them on a bus that must support several modules in parallell. On a graphic card the memory is connected on a point to point bus which allows much higher frequencies and smaller tolerances. Not only that but memory for PCs must be produced in HUGE quantities which means yeilds can't be sacricied. Memory for graphic cards do not need the same volume since not every PC comes equipped with a Radeon 9800XT. Thus these cards can use the top end binsplits of the DRAM that is turned out.

Until there is a significant change in the way the DRAM cells are designed we are going to continue seeing the incremental improvements in memory. Large increases in bandwidth but not much decrease in latency.

Your idea is simply not possible in the way you imagine it. You cannot trade bandwidth for latency with DDR-II. I agree with you about the efficiency of SDR vs DDR. SDR is much more efficient than DDR in terms of delivered bandwidth to max bandwidth. I am not sure what steps are done in DDR-II to improve this, if there are any at all. We will soon find out how good (or bad) DDR-II really is.

 

[Peter]

Latency, in real time measures (read: nanoseconds) has neither improved nor gotten worse with the move from SDR to DDR. That's the whole point of why DDR does not equal twice the speed - only the data transfer got faster, the addressing and housekeeping did not.

DDR266 still operates at 133 MHz, and e.g. a CAS latency of two clocks still takes 15 nanoseconds, just like on 133 MHz SDR. Only the 8-word databurst following the latency completes in 8 half-clock cycles not 8 full cycles. That's where the perceived effectiveness gap comes from - it's not that DDR is worse at latency, it's SDR's being worse at throughput, relatively dwarfing the addressing latency.

Unbuffered PC133 RAM on a really good chipset and lightly loaded bus takes 12+8 cycles to complete a databurst - worst case, from scratch, full row and column addressing. DDR266 in the same situation takes 12+4. Now, should we say that the addressing overhead got worse in DDR just because addressing takes three times as long as moving the actual data, while on SDR it was only 1.5x as long? No. The addressing overhead is exactly the same.

And if you did pump existing SDRAM technology to e.g. 266 MHz, then the silicon would still need the exact same amount of real world time to address the RAM cell you want to read/write. So, at twice the frequency, you'd need 24 instead of 12 cycles of addressing overhead for your 8 datawords. Now, SDR266 taking 24+8 cycles 3.75ns each then would be exactly as fast as DDR266 RAM taking 12+4 cycles 7.5ns each - with the extra headache of having to manage twice as high a clock signal frequency on the board. Great move.

 

[MadRat]

Sahakiel-

Strangely enough it never stopped you from dumping in my threads. I understand your messages fair enough, but trust me, the abrasiveness is revolting.

andreasl-

I'm not sure where your personal blast was directed at, but again, please keep your thread dumps elsewhere. If you got a problem with the idea then stick to the subject. Leave out ad hominum. You don't like me bringing up video ram and that seems to always set you off for some reason. If the industry moved to a two-tier memory solution then your argument would not hold water at all. Saying it won't happen is not going to alter history - alot of things have been done that experts across the industry said were stupid or the changes simply didn't make sense. Every once in awhile we get a change that makes us, another year or two down the road, look back and say that old stuff sucked.

Peter-

So you are saying that SDR and DDR should operate identically as far as performance goes when it comes to all memory transaction delays. I can relate to that, which its commendable you were able to deject my argument with raw argument and no ad hominum. Thanks.

There are some situations where SDR is faster than DDR, tough, which is why I asked if a dynamic controller was possible.

OT: I realize that 100MHz PC1600 memory was relatively equal to PC166 (which really wasn't a JEDEC standard, true, but it was sold - as was PC200...) in bandwidth, but in some things the comtempary SDR standard (PC133) absolutely crushed the PC1600 DDR. In all honesty, in some benchmarks even PC100 was better than PC1600. Generally the SDR could only compete in activity that required alot of quick address jumps and only performing dynamic reads of single random bytes from single random dwords from single random pages. (Not sure if those were the correct terms, but it sounds right.) The old performance PC reviews by review sites, though, don't do SDR justice because they always compared overclocked SDR and DDR, which meant the DDR cleaned up. It really wasn't a fair comparison once overclocking was done back in the day because the SDR was at 3.3v and the DDR at 2.5v, which meant the DDR memory was - on the microarchitecture level - at a huge advantage for raw speed. Which seems to contradict something I wonder about - if reducing the process that they build the memory on if this actually makes you have to recharge memory more often? That could explain why DDR never was truly double-fast and why its getting harder to get quality high-speed DDR memory with decent latency these days as clock speeds ramp.

 

[Sahakiel]

Originally posted by: MadRat
Sahakiel-

Strangely enough it never stopped you from dumping in my threads. I understand your messages fair enough, but trust me, the abrasiveness is revolting.

The history comment came about because your posting is rubbing off as coming from one who knows numbers but not theory.
As for the "dumping," I can honestly say I'm not "dumping" in your threads. Your thread (this one) asked for feasibility based on flawed data. I corrected your data, cited the various reasons from my analysis, but for some reason you ignored it. I repeated the information, you bypassed it. Peter posted the exact same data and suddenly your view has changed. Conclusion: Anything I post which differs from your viewpoint is dumping whereas the same information from others is seen as perfectly valid. Trust me, your abrasiveness is revolting.

OT: I realize that 100MHz PC1600 memory was relatively equal to PC166 (which really wasn't a JEDEC standard, true, but it was sold - as was PC200...) in bandwidth, but in some things the comtempary SDR standard (PC133) absolutely crushed the PC1600 DDR.

In the first sentence, you're comparing 166MHz legacy SDRAM to 100MHz SDRAM. Obviously, there will be differences.
In the second sentence, you're comparing PC133 to PC1600, 133MHz SDRAM to 100 MHz SDRAM. Again, there will be differences.
The key issue being bandied about in this thread is the comparison of legacy SDRAM to DDR SDRAM to DDR-II SDRAM at the same clock speeds. The reasoning is due to all three formats using essentially the same memory cells which run at the same speeds. The sole exception would probably be DDR-II which doesn't have a specification beyond 133MHz which is why my analysis has so far been done using 100-133MHz clock speeds.

In all honesty, in some benchmarks even PC100 was better than PC1600. Generally the SDR could only compete in activity that required alot of quick address jumps and only performing dynamic reads of single random bytes from single random dwords from single random pages. (Not sure if those were the correct terms, but it sounds right.)

Again, I question as to whether the problem was due to the memory technology or the memory controller. I am having a hard time seeing how two memory sticks running at the same speed can operate at a significantly measurable difference when the only control variable is bandwidth. I am more inclined to believe that the memory controller has a more difficult time dealing with the excess data. The assumption here is that given a DDR memory controller which simply discards the second bit the performance of legacy SDRAM and DDR SDRAM will be equal. The data I have seen on spec sheets has been equal latencies for both technologies.

The old performance PC reviews by review sites, though, don't do SDR justice because they always compared overclocked SDR and DDR, which meant the DDR cleaned up. It really wasn't a fair comparison once overclocking was done back in the day because the SDR was at 3.3v and the DDR at 2.5v, which meant the DDR memory was - on the microarchitecture level - at a huge advantage for raw speed. Which seems to contradict something I wonder about - if reducing the process that they build the memory on if this actually makes you have to recharge memory more often? That could explain why DDR never was truly double-fast and why its getting harder to get quality high-speed DDR memory with decent latency these days as clock speeds ramp.

It is quite possible that more refreshes are required. I believe DRAM cells contain on the order of 100 electrons or less. However, this also means refreshes take less time. My guess is maintaining the data doesn't affect performance at all but the faster refreshes allow faster reads hence the faster clocks.
Also, I for one would've been surprised if DDR performed twice as fast as its predecessor. Spatial locality in is high, but nowhere near 100%.

 

[MadRat]

I didn't ignore what you said, I just didn't percieve your message. Peter is a little better at speaking down to layman's terms. Hope there is no hard feelings.

 

[Pudgygiant]

At the risk of getting flamed, two-tier graphics card memory? Elaborate pelase.

 

[MadRat]

Graphics cards require raw bandwidth to feed their parrallel behavior. CPU's require much less bandwidth, but quick access to their data is a bonus. If one was to mate CPU and GPU functions together then what would one settle for in memory, GPU-like speeds or CPU-like speeds? The answer would be to do both, running a top tier of high speed memory that allows the GPU to dominate all or just some of it. A second tier of much slower memory would be for the CPU to dominate, but as with current graphics cards I'm sure the GPU would need some limited access to it. That is why I said the hybrid technology's practicality could probably be demonstrated using HT links between a dedicated GPU and CPU. The problem is that the two processors need to learn how to play together, first, something neither currently was setup to do.

 

[Pudgygiant]

If CPU's require much less bandwidth, why is the bus speed relatively faster than that of a GPU?

 

[MadRat]

Its not. GPU's frequently use 256-bit connections at full memory speed, with the memory substantially faster than the memory found on the typical high performance motherboard. The next generation high end cards will be pushing 500-600MHz whereas DDR-II will debut at 100MHz, and I'm not for certain if its dual-channel. (DDR-II @ 100MHz = 3.2GB/sec memory bandwidth for a single channel.) The video memory would be pushing four to six times the raw MHz and using basically four channels to gain its extreme performance. Big difference in brute force.