Sign in with
Sign up | Sign in
Your question

Rambus question latency vs physical speed

Last response: in Memory
Share
March 26, 2002 7:33:58 PM

I want to sort out some information here, so bear with me. Let me get some things in order first.

1-PC-800 runs at 400Mhz double pumped, meaning that it has a 100Mhz clock pushed four times

2-RDRAM has higher latency than SDRAM/DDR due to its need to be used in a serial mode (you would have to add the latency of 2 sticks of RDRAM to get the actual latency, right?)

2.5-Does RDRAM have lower latency (per stick) than DDR or SDR?

3-Because it is serial, EVERY port has to be filled identically, meaning that in no possible way, you could interlace different size modules (ie 128 and 256)

4-the PIV only tops out its current bandwidth with 3.2Gbs of RDRAM (1.6 X 2 sticks)

5-What is the highest bandwidth the Athlon can be fed by? (I'm guessing 2.1Gb/s?)

6-If someone can out with some magical DDR running at 400Mhz double pumped, it would run at 800Mhz synonymously like RDRAM, albeit with less latency, would this special brand of DDR be able to outperform the PIV with RDRAM?

I still have yet to see why Intel risked the PIV on Rambus, and is soon to withdraw from them. As I can gather, they plan on going northwood with 533FSB and PC-1066 RDRAM. If they do go with DDR later on, what can we expect for other memory types? Is DDR really the future?

"When there's a will, there's a way."
March 26, 2002 8:04:34 PM

“1-PC-800 runs at 400Mhz double pumped, meaning that it has a 100Mhz clock pushed four times”

PC800 runs at 400MHz “double pumped.” The FSB and the Memory Bus are two different things. The FSB is the bus between the CPU and the Northbridge. The memory (data) bus is between the memory modules and the Northbridge. The MCH (Memory Controller Hub) is incorporated in the Northbridge. The FSB runs at 100MHz with four channels. (100MHz x 4 = an effective 400MHz.) The Memory is running at 400MHz utilizing a DDR signal. (400MHz x 2 = 800MHz)


“2-RDRAM has higher latency than SDRAM/DDR due to its need to be used in a serial mode (you would have to add the latency of 2 sticks of RDRAM to get the actual latency, right?)”

The latencies are different because the architecture is different. Thus the timing, signals, and strobes are different. PC100 SDRAM has a latency of about 90ns (nanoseconds) and RDRAM PC800 has a latency of about 67.5ns. DDRSDRAM is a little higher than SDRAM, clock for clock that is, because of the additional timing sets. So PC200, which is 100MHz DDR, would have higher latencies than PC100.



“2.5-Does RDRAM have lower latency (per stick) than DDR or SDR?”

See above.



“3-Because it is serial, EVERY port has to be filled identically, meaning that in no possible way, you could interlace different size modules (ie 128 and 256)”

That is not because it is serial but because the bus is dual channel. Both channels have to be filled with identical twins.



“4-the PIV only tops out its current bandwidth with 3.2Gbs of RDRAM (1.6 X 2 sticks)”

I don’t know if this is a question and if it is what are you trying to ask???




“5-What is the highest bandwidth the Athlon can be fed by? (I'm guessing 2.1Gb/s?)”

I don’t know. The Nvidia Chipset uses a dual channel and the theoretical bandwidth is 4.2GB/s.




“6-If someone can out with some magical DDR running at 400Mhz double pumped, it would run at 800Mhz synonymously like RDRAM, albeit with less latency, would this special brand of DDR be able to outperform the PIV with RDRAM?”

Most likely it would, but we are talking about the abstract. There is no way to test it out until it came out.




“I still have yet to see why Intel risked the PIV on Rambus, and is soon to withdraw from them. As I can gather, they plan on going northwood with 533FSB and PC-1066 RDRAM. If they do go with DDR later on, what can we expect for other memory types? Is DDR really the future?”

DDRII, QDR, and QDRII are the next progressions for SDRAM. PC1200 is the next for RDRAM. Both types are to be in the marketplace some time this year.


Fatburger,

Fill in what I forgot.

<b>"I put instant coffee in the microwave and almost went back in time" - Steven Wright</b> :lol: 
March 27, 2002 6:03:31 AM

Quote:
1-PC-800 runs at 400Mhz double pumped, meaning that it has a 100Mhz clock pushed four times

No, PC 800 is 400MHz double pumped meaning it has a 400MHz clock pushed two times.

Quote:
2-RDRAM has higher latency than SDRAM/DDR due to its need to be used in a serial mode (you would have to add the latency of 2 sticks of RDRAM to get the actual latency, right?)

The two (dual channel) sticks of RDRAM run in parallel and do not double the latency - the additional latency <i>is</i>, however, due to serial processing. Each chip on a RIMM has a set latency and the total latency of the stick is a product of the number of chips; to reach the latter chips the data must pass through the closer ones and the latency is normalized between them all. Adding additional pairs of sticks also significantly increases the latency factorially as the data must now pass through a pathway double or triple the length.

Quote:
2.5-Does RDRAM have lower latency (per stick) than DDR or SDR?

To which Bum JCRules replied:
PC100 SDRAM has a latency of about 90ns (nanoseconds) and RDRAM PC800 has a latency of about 67.5ns. DDRSDRAM is a little higher than SDRAM, clock for clock that is, because of the additional timing sets. So PC200, which is 100MHz DDR, would have higher latencies than PC100.
Quote:


Which is <A HREF="http://www.ee.umd.edu/~blj/papers/memwall2000.pdf" target="_new">incorrect</A>. RDRAM has a much higher latency than SDR or DDR due to the fact that each RIMM has multiple chips on board. Only a single chip of RDRAM approaches the lower latency of SDRAM.

Quote:
3-Because it is serial, EVERY port has to be filled identically, meaning that in no possible way, you could interlace different size modules (ie 128 and 256)

No, the slots have to be filled with identical pairs as the i850 chipset requires dual channel PC800 (400MHz * 2 => 800MHz * 2 RIMMs * 16bits = 3200MBps) for the 100MHz quad pumped 64bitFSB (100MHz * 4 => 400MHz * 64bits = 3200MBps) to run synchronous. You may mix and match any sizes so long as RIMM are added in identical pairs.

Quote:
4-the PIV only tops out its current bandwidth with 3.2Gbs of RDRAM (1.6 X 2 sticks)

Yes, until PC1066 and PC1200 are required for the 533MHz and 600MHz P4s' arrivals.

Quote:
5-What is the highest bandwidth the Athlon can be fed by? (I'm guessing 2.1Gb/s?)

First that would be "2.1G<b>B</b>/s." The Athlon can be fed at any memory rate a chipset can be configured at to support it - many, including myself, have run it at 166MHz (333DDR) FSB (2.7GB/s) or even higher such as at 200MHz (400DDR) FSB (3.2GB/s). It is not a matter of how much the Athlon can be fed, but how hungry it is. The Athlon has a much higher IPC and lower pipeline than the P4; therefore, its L1 and L2 caches tend to be much more efficient. It does not require as much memory bandwith as the P4 to perform equally in most tasks. This is illustrated by the 4.2GB/s nForce chipset that shows very little performance difference (although it is a very sophisticated and valiant first effort for nVidia). Some specific tasks are nevertheless memory bandwidth intensive and are benefited by the increased bandwidth.

Quote:
6-If someone can out with some magical DDR running at 400Mhz double pumped, it would run at 800Mhz synonymously like RDRAM, albeit with less latency, would this special brand of DDR be able to outperform the PIV with RDRAM?

PC3200 (DDR400) (200MHz * 2 => 400MHz * 64bit = 3200MB/s) already exists in small expensive quantities. This runs sychronous to the P4 bus in nearly the same fashion as RDRAM (400MHz * 2 => 800MHz * 2 RIMMs * 16bits = 3200MBps) and more closely matches the P4's FSB (100MHz * 4 => 400MHz * 64bit = 3200MB/s). Performance really depends on the chipset. If Intel had introduced a dual-channel DDR chipset, although slightly more expensive on the chipset and mainboard side, we would already have 5.3GB/s (PC2700 * 2DIMM) available to us. And the higher cost of chipset and mainboard would have offset the prior extreme high cost of RDRAM.

Quote:
I still have yet to see why Intel risked the PIV on Rambus, and is soon to withdraw from them. As I can gather, they plan on going northwood with 533FSB and PC-1066 RDRAM. If they do go with DDR later on, what can we expect for other memory types? Is DDR really the future?

Intel risked it with Rambus because they wanted more control over the memory technology feeding their processors - previously, they were at the whim of the slowly progressing RAM manus. This put a shot in the SDRAM manus' respective arms to get DDR et. al. out the door faster - they didn't/don't want to be controlled by Intel.

Intel is already going with DDR for its server chipset, the e7500. What they do with the Prescott still remains to be seen. The future of RAM is, IMHO, going to progress based on the market. Rambus has shown us one vision of that future, but we'll just have to wait and see.

I thought a thought, but the thought I thought wasn't the thought I thought I had thought.
!