Printable View
Price will not be high.. DDR is now more expensive then Rd at my Canadian vendor.. with 1066 edging out in performance.. price is not an issue anymore.. and with a quad controler.. the latency would be SWEET... Bandwith is irrelevant if it is more then the cpu needs.. therefor the speed is the real factor.. and DDR 400 usually clocked at 3XX to increase latency.. can't complete with 1200 mhz... ofcourse ddr 600 and stuff will come out but thats still 1/2
I heard that Rambus will have PC-1200 64bit memory out which would give it 9.6gbs in a single channel.
The price increase I was referring to is the price of the controller ASIC, not the ram. I do expect RDRAM prices to climb as demand falls off - even PC1066 is no longer the fastest stuff on the market, and intel will be putting out no more RDRAM chipsets, so demand will fall off considerably. But even if the RAM stayed the same price, putting 4 memory controllers on a chip will increase the price of the chip.Quote:
Originally posted by MAVrecK
Price will not be high.. DDR is now more expensive then Rd at my Canadian vendor.. with 1066 edging out in performance.. price is not an issue anymore.. and with a quad controler.. the latency would be SWEET... Bandwith is irrelevant if it is more then the cpu needs.. therefor the speed is the real factor.. and DDR 400 usually clocked at 3XX to increase latency.. can't complete with 1200 mhz... ofcourse ddr 600 and stuff will come out but thats still 1/2
Also, while more controllers will decrease the latency from the CPU to the memory controllers, it will not affect the latency of the controllers to the RAM at all - and that's where RDRAM has latency issues. DDR 400 is almost always low latency, and even most PC3500 is CAS 2.5 or less now (mushkin, kingston hyper-x, corsair).
Bandwidth is (clock speed) x (bus width). Bandwidth is the total amount of data the memory can send to the controller in a certain amount of time. Latency is how long it takes from a request for data until it is sent on the memory bus. Clock speed is irrelevant except in how it affects bandwidth and latency. RDRAM is high clockspeed, low bus width. DDR SDRAM is high bus width, low clock speed. The result? They perform about the same. 1200 MHz RDRAM at 16 bits bus width (standard 1-channel) would be far slower (less than 1/2 after latency) than 600 MHz DDR SDRAM at 32-bits (standard 1-channel).
But of course, DDR will be used in dual-channel, and RDRAM in quad channel, so it all evens out, roughly.
64 bit single channel RDRAM will never happen - certainly not at 1.2 GHz. The reason RDRAM can run at such high clock speeds is that they can keep the trace count down and terminate and isolate each one. A 64-bit bus at 1.2 GHz would be nearly impossible with current PCB techniques. I suspect that what you heard was the quad-channel PC1200 info, which is 4 16-bit busses and has 9.6 GB/sec of bandwidth.Quote:
By Voodoo 5 5500 I heard that Rambus will have PC-1200 64bit memory out which would give it 9.6gbs in a single channel.
Of course, I could be wrong!
:D
Benchmarks? I would think it would be faster actually, due to the efficient architecture of RDRAM. Granite Bay has more BW and lower latency than RDRAM and still gets beat by it.Quote:
1200 MHz RDRAM at 16 bits bus width (standard 1-channel) would be far slower (less than 1/2 after latency) than 600 MHz DDR SDRAM at 32-bits (standard 1-channel).
Latency has little to due with performance in RAM benchamrks. Just compare RIMM4200s with PC1066. The difference is negligeable even with the differences in latency.
Where are you getting your information? Due you know how complicated and expensive it will be to verify the memory controllers for the DC DDR chipsets? It's insane.Quote:
putting 4 memory controllers on a chip will increase the price of the chip.
I would show you the 64-bit RIMM whitepaper, but the Rambus website is down right now. In any case, 64-bit RIMMs are supposed to be demonstrated this year and will be used on the Sis 659 if it is released.Quote:
64 bit single channel RDRAM will never happen - certainly not at 1.2 GHz. The reason RDRAM can run at such high clock speeds is that they can keep the trace count down and terminate and isolate each one. A 64-bit bus at 1.2 GHz would be nearly impossible with current PCB techniques.
Since neither 16 bit 1200 MHz RDRAM or 600 MHz 32 bit DDR will likely ever see the light of day, benchmarks would be difficult. The reason I say 16 bit 1200 MHz RDRAM would be slower is because it has (16 bits) x (1200 MHz) = 2.4 GB/sec of bandwidth, whereas 32 bit, 600 MHz DDR has (32 bits) x (300 MHz x 2) = 4.8 GB/sec of bandwidth. Like I said, thats single channel. DDR would be used in a dual channel config, RD in a quad channel config.Quote:
Originally posted by elimc
Benchmarks? I would think it would be faster actually, due to the efficient architecture of RDRAM. Granite Bay has more BW and lower latency than RDRAM and still gets beat by it.
Latency has little to due with performance in RAM benchamrks. Just compare RIMM4200s with PC1066. The difference is negligeable even with the differences in latency.
I agree, the DC-DDR verification is insane. That's why I think quad channel RDRAM, with the same number of signal traces, but at 6 times the clock speed will be even harder!Quote:
Where are you getting your information? Due you know how complicated and expensive it will be to verify the memory controllers for the DC DDR chipsets? It's insane.
This is a quad channel RIMM - 4 channels of 16 bits each. Like I said it was.Quote:
I would show you the 64-bit RIMM whitepaper, but the Rambus website is down right now. In any case, 64-bit RIMMs are supposed to be demonstrated this year and will be used on the Sis 659 if it is released.
A quad channel RIMM would provide 9.6GBs. That is plenty fast to destroy DDR SDRAM which would have to be dual channel in order to compete.Quote:
The reason I say 16 bit 1200 MHz RDRAM would be slower is because it has (16 bits) x (1200 MHz) = 2.4 GB/sec of bandwidth, whereas 32 bit, 600 MHz DDR has (32 bits) x (300 MHz x 2) = 4.8 GB/sec of bandwidth. Like I said, thats single channel. DDR would be used in a dual channel config, RD in a quad channel config.
RDRAM verification is tricky, but not insane. The timing specs are very tight and well laid out. DDR SDRAM verification, however, is very tough. Actually, I think that DDR SDRAM has uses four times the number of signal traces that RDRAM uses, and the signaling is much poorer. This has lead to DDR SDRAM boards that are constantly late. Also, you may be forgetting that RDRAM is double pumped, so a 1200MHz RDRAM module would only be twoce as fast as a 600MHz DDR SDRAM module.Quote:
I agree, the DC-DDR verification is insane. That's why I think quad channel RDRAM, with the same number of signal traces, but at 6 times the clock speed will be even harder!
Right. Four channels on one RIMM will give you 9.6GBs per stick.Quote:
This is a quad channel RIMM - 4 channels of 16 bits each. Like I said it was.
[(64 bits) x (300MHz x 2)] / 8 = 4.8GBsQuote:
(32 bits) x (300 MHz x 2) = 4.8