Slashdot Mirror
PC1066 RDRAM vs. DDR SDRAM
Brad wrote into send us his "Comparison of PC1066 RDRAM vs DDR SDRAM. Quote - RDRAM is considerably more expensive that DDR SDRAM, and up until now the 100MHz PC800 specification didn't do well in comparison. Just recently 133MHz PC1066 was launched, and is now officially supported by the new Intel P4 and the Intel 850E core logic chipset, but this time promises to bring memory performance to the next level."
but why would anyone want to shell out for an RDRAM/P4 system? You can get an Athlon for much cheaper, and load up on DDR memory. It may not be quite as fast as the Intel system, or play a fancy tune in some commercials, but it'll get the job done for a lot less, in most cases.
http://www.theinquirer.net/24050203.htm
That said, PC1066 has been tested before (can't find the article at Ace's Hardware), and the bandwidth of DRDRAM appears to compensate quite nicely for the P4's generally lousy architecture, as does its increased cache size (now 512k L2).
Why didn't they show us any Quake III comparison benches? We all know that at lower resolutions the processor drives Quake III and that its extremely sensitive to memory bandwith capabilities. Anyway it appears that RDRAM 1066 is a definite improvement over RDRAM 800. Its good to see that Intel is still continually raising the bar.
Also I believe there were some initial benches (better ones) on http://www.tomshardware.com
J
I love idealists not because I am one, but because they make life bearable for pragmatists such as myself.
While the benchmarks he ran show nice bandwidth figures (Negligible, really, in light of how expensive that RDRAM is- if that's all this new memory spec can do, well...) it doesn't tell the whole story. There's bandwidth and then there's latency. In the case of RAMBUS, there's more latency involved with the access of the memory than with DDR SDRAM- latency that may eat some or all the bandwidth gains you see there when you start doing something other than benchmarks. If it's not really much faster (Sorry, it's not when you start looking at the bigger picture), why are you spending 3 or more times for it?
I am not merely a "consumer" or a "taxpayer". I am a Citizen of the State of Texas
Basically, CPU cooling has been hitting us for a good while.
From an article about a bigass Beowulf cluster running Transmeta processors, you have Wu-chun Feng of the Los Alamos Labs stating
Oh my. So - what else can we do to stop this trend? Relatively slow multi-processor machines. If we keep working on multi-threading our applications, we might be able to make a computer with 8 1ghz efficient chips outperform an 8ghz Moore-compatible Intel hype-chip-based system. Really. Multi-processor machines have traditionally been too expensive for the desktop. The software people have not spent a lot of time making sure that the regular end-user applications scale well across several processors.Take something like a web browser. Given a bit of wizardry (obviously, we need to consider concurrency and critical sections), you could have separate images downloaded and processed by separate processors. Your flash ad would run on another processor.
Frankly, I'm wondering what's stopping us from using this approach to increasing performance? Is this like the fact that OEMs equip the low-end PCs with too little RAM so that Joe Shmoe will buy a new one as quickly as possible, since he does not know that spending 100 bucks on more RAM will make his computer last another year or two?
And, really, as long as the focus is on the gigahertz, do the chip makers really concentrate on making their designs as efficient as possible?
Stop the brainwash
The article is about PC1066, a new kind of memory. The memory specific benchmarks do show quite a big performance increase!! (see the last three graphs on this page of the article)
The fact that the other graphs show little or no performance difference I think is quite likely due to the fact that the tests employed have different kinds of bottlenecks due to system limitations -- limitations other than memory bandwidth.
You might get similar results if you tested a new sound card (for example) that had faster hardware acceleration -- sure, the Quake III benchmark would only show a small difference, but another test that made more significant usage of the sound card (a test in Cubase for example) would show a greater performance increase. (Umm, I know it's not a great example, but I'm hope you get what I mean!).
So what you are saying is that the P4 has a flawed architecture, but they overcome that by ramping up the speed to 6GHz (and 1.2MW heat dissipation)? I'm not seeing how this is a plus in any way. Most people consider good design to be able to do more (processing, executions per clock cycle, memory movement) with less (voltage, energy, heat).
Vote monkeys into Congress. They are cheaper and more trustworthy.
No, the P4 has an architecture that was designed for the computers of the future. It's like a small dog with very big paws. It will be impressive when it grows up.
The heat dissipation comes from using the P4 architecture with the larger design rules. As the die sizes shrink, the heat dissipation will go down, and the wisdom behind other elements of the design will become more apparent.
Notice that we are already seeing this effect. The 2.4 GHz P4 performs very well.
Intel is demonstrating a 5 GHz P4 that runs cool with no fan. See, for example, Intel to demo fanless, cool 5 GHz chip. Quote: "Intel has now formally released details of the 3MB cache on chip which it claims will deliver 1.5 to two times [the] performance over the current designs." [My emphasis.]
The utter sadness of Intel's marketing is demonstrated by the fact that this information is being brought to you by a guy [me] whose only connection with the information is that he sells computers to business customers and that he happens to live in the same city as Intel's design team. The guy happened to meet two Intel engineers at parties. If Intel had good marketing, you would already know these things.
The moral of the Intel marketing story is: Don't try to run a high-tech company with low-tech employees in marketing. If I were running Intel's marketing, your little brother and maybe even your mom would be asking you about Intel's great new achievements.
In the beginning there was PC100 SDRAM. Well, actually, that was mid-nineties, but that's about when most Slashkiddies were born, so moving on. Obviously everything is just a marketing label, but this one meant 100 MHz. With SDRAM, each Hz gives you 64 bits, so the bandwidth is 6400 megabits per second.
Thus PC133 and PC166 are 8500 and 10700 Mb/s.
DDR is the same tech as SDRAM, except that it uses a trick to transfer data twice per clock cycle, so you get 128 bits per Hz. Thus PC100 DDR-SDRAM would be 12800 Mb/s. But Marketing decided that was unfair, so they labeled DDR based on twice the clock speed, so we have PC266 and PC333, which of course run at 133 and 166 MHz and give you 17000 and 21000 Mb/s.
RDRAM is based on a new tech that gives you only 16 bits per clock cycle instead of 64 for SDRAM and 128 for DDR-SDRAM. The difference is that you can clock it way up. So there was PC600, PC700 and PC800 RDRAM, again based on MHz, so that gave you 9600, 11200, and 12800 Mb/s bandwidth. Basically you divide the number in four to compare with SDRAM speeds, since you only get 1/4 as many bits per cycle. Actually I believe modern Rambus controllers double this by interleaving two sticks, so now you divide by 2 - PC800 has four times the bandwidth of PC100, but requires a matched pair of sticks.
Then the DDR people decided to start talking direct bandwidth, rather than megahertz. But unlike me, they mean megabytes, rather than megabits, per second. PC1600 is DDR-SDRAM at 100 MHz, since DDR gives you 128 bits or 16 bytes per cycle. PC2100 is DDR at 133 MHz, formerly known as PC266. PC2700 is DDR at 166 MHz, and PC3200 is DDR at 200 MHz.
With interleaving, Rambus gives you 32 bits or 4 bytes per cycle. PC800 has the same bandwidth as PC3200 DDR, and the relatively new PC1066 has more - 4266 megabytes per second.
Bandwidth is a good baseline for comparison, but RDRAM has a higher latency than SDRAM or DDR-SDRAM. That's why DDR, with its lower maximum bandwidth, is still speed-competitive with RDRAM (for a lot less money).
"How can you claim that you are anti-crack, while still writing a window manager?" — Metacity README
Perhaps, but you're probably comparing single-stick to single-stick. With RDRAM you have to buy a matched pair. So the right comparison is 2x128 PC800 ($80) versus 1x256 PC2400 ($51).
Or go on up to 512MB. 2x256 PC800: $148. 1x512 PC2400: $114.
So RDRAM costs an additional 57% for 256MB, or 30% for 512MB. Nice that it's no longer double the cost, but to me that is still a significant markup. Anyone know approximately how much of that is due to
(b) manufacturing cost after accounting for (a), or
(c) patent licenses?
"How can you claim that you are anti-crack, while still writing a window manager?" — Metacity README