Slashdot Mirror
Intel To Pull Plug on RAMBUS, Use SDRAM?
Ratteau writes " Cahner's Group Electronic News, is reporting they have come across documents that Intel "has pulled the plug on plans to use Rambus direct memory in the mainstream PC market". " Given the troubled past with Rambus, this wouldn't surprise me - but it's a big move for Intel.
From the article, this sounds more like a delay in implementation than Intel giving up on Rambus.
It's one small step on a long road that Intel must travel to win my confidence back. Their last year of gaffes has lost nearly all the goodwill they built up with me
SteveI work closely with a DIMM engineer working with PC100 and PC133 memory systems. We've been watching all the lawsuits flying recently. Here's a few links about royalty and patent lawsuits:
e ws/OEG20000324S0022
u es/200004/rambus&page=1
6 1500.htm
7
Rambus asks ITC to bar Hitachi, Sega imports: (3/24/00) http://www.eet.com/story/industry/semiconductor_n
Will Rambus Go Bust?: (4/17/2000) http://www.32bitsonline.com/article.php3?file=iss
Toshiba Signs Patent License Agreement with Rambus
For SDRAM & DDR SDRAM Memory and Controllers: (6/16/00) http://www.rambus.com/general/press_releases/pr_0
Tech Report Analysis of Toshiba agreement: (6/16/00) http://www.tech-report.com/onearticle.x/881
RAMBUS using patent claims to lift RDRAM share: (6/25/00): http://www.ebnews.com/story/OEG20000623S0042
DRAM industry considers antitrust lawsuit vs. RAMBUS: (7/10/00) http://www.techweb.com/wire/story/TWB20000710S000
400 Mhz is a really big jump from the industry standard 100 - 133 Mhz. Minor impedance variations from board to board in production can cause significant phase and wave form changes in such fast signals.
To accurately transmit 400 Mhz square waves it is necessary for the board traces to handle 4 Ghz sine wave signals. That is more of an Analog micro wave transmission problem than it is any kind of digital design problem. Evidently the board designers have had a great deal of trouble doing this.
The bottom line is that Rambus motherboards will need to be produced with tighter tolerance on both trace and board substrate thickness than current motherboards. Result: even more expense for a Rambus system compared to a DDR based system.
Basically, RAMBUS has the theoretical capability to be significantly faster than SDRAM (not DDR, more later). However, the controllers have problems that prevent this. Basically, RDRAM can keep many pages open and many devices active at a time (more than SDRAM), but the i820 doesn't do this. So the chipset is crippling the RDRAM. Also, as soon as multiple devices are put on the bus, the latencies increase, so if too many chips are present things slow down. This is because of the longer wires needed. at 400MHz (not 800 - its DDR) that really matters. Also, RDRAM has been hindered by low yields and hence higher cost. It is now down to about double PC133 (see pricewatch). Also, the chips are more complex. However, the specs say that a good controller ought to be able to outperform PC133. Not by huge amounts, but by enough to matter. i820 is far from a good controller. Something to think about: the EV7 (maybe EV68, I can't remember) is going to use RDRAM. (also on Ace's hardware). However, it is going to increase performance by using 8 channels in parallel. So until there is a good desktop controller, and RDRAM is similar in price, AND the benchmarks say it's better, I'm using DDR SDRAM. But, the technology isn't inherently bad, just having more than it's share of problems.
---
It takes less power to run your CPU than a light bulb. On the desktop, power consumption doesn't matter so much as long as you can dissipate the heat. Transmeta already has a solution for reducing power use in mobile devices. Intel is doing the intelligent thing by occasionally devoting some time and R&D money toward developing lower power CPUs, but focusing on the biggest, best, and brightest.
In response to your core question, IE "who actually needs a processor this fast", the answer is, we all do. As higher-end CPUs get faster, lower-end CPUs get cheaper. As more processing power becomes available, we are able to predict weather more accurately, make more fuel efficient automobiles, and discover more about the cosmos in which we are only a tiny speck. Computers help us do everything, making products cheaper and giving a better way of life to all people (though admittedly there are billions for whom the benefits are slow to trickle down to.)
Next time you buy a car, stop a moment and marvel at the amount of processor time that went into modeling various aerodynamics characteristics, enabling you to get dramatically better gas mileage than the refinements made to the engine alone; Which are all designed and tested on the computer before they ever see metal. This is true of nearly every product with more than five moving parts, and many that have none at all. Be thankful for the race to technology, or get the hell off your computer and go plant some carrots or something.
"You're right," Fisheye says. "I should have set it on 'whip' or 'chop.'"
Um... right, but when they patented these ideas while participating in an industry-wide panel to discuss RAM standards, without notifying anyone else in the panel, and when these patents suspiciously look a lot like the technology that the panel, as a group, developed... To say that they 'contributed significantly' seems a little shakey. More like 'grabbed the technology and ran'. Yeah, real 'solid'.
That only TWO megacorps have given in should tell you something -- especially since the two that gave in are both known to be non-confrontational and to have a vested financial interest in being able to continue to sell RDRAM (the patent for which no one disputes). When you are in the lucrative position of supplying the RAM for the PS/2, you do _not_ want to piss off the company that lets you do this. This is hardly an admission that the Rambus Inc patents have any merit.
Rambus chose their first targets very carefully. They know the rest of the industry will not give in to their ludicrous assertions, and they wanted to give the impression that the whole industry was going to cave in to them. Sadly (for RMBS stockholders), the actions of the other memory companies demonstrates that this is not the case.
God forbid anyone should make money in the computer business? HA! It is to laugh. Seems to me all the major memory makers have been doing just fine financialy without feeling the need to hold a patent on SDRAM.
The enemies of Democracy are
There are applications where throughput is more important than a significant amount of latency.
An example is context-switched highly-parallel processing, where the number of crunches per second on each piece is relatively low and fixed. You pipeline and context-switch, and get multiple virtual processors from one set of gates and a RAM. The higher your bandwidth the more virtual processors you get. This is important in many applications. Sometimes space for the box is more expensive than the box itself. Sometimes the cost per virtual processor is critical.
But some problems don't parallelize well. And even for those that do, parallizing them can be a real bitch. So your desktop or laptop (which tends to be doing only one or a few crunch-intensive things at a time) is organized to stick with a task for a significant time, then hop to another. RAMBUS isn't a match for that.
And it looks like it has some other kisses-of-death even for interleaved context-switched stuff. Oh, well...
Bantam Dominique roosters crow a four-note song. Once you've heard it as "Happy BIRTHday" you can't NOT hear it that way
Unless of course your doing disk access, in which case your far better off performance wise with PC133 and a SCSI Disk subsystem... which interestingly enough, can be done for still less money then 256MB of 800mhz RDRAM costs.
Yeah... you take your RDRAM, I'll take my Ultra160 controller and 10k RPM drive, and we'll see just who beats who as soon as you have to stop to access your good old IDE hard drive.
-- "So they told me that using the download page to download something was not something they anticipated." - Bill Gates
Sorry, but RAMBUS _is_ a bad design for a next-gen memory technology.
Benchmarks demonstrate that as processor speed increases, the performance gap between PC133 and RAMBUS increases (meaning RAMBUS does _worse_).
This makes complete sense to me. As processor speed increases, the number of CPU cycles that are wasted waiting for a critical piece of data to return from RAM increases. Latency becomes more important and more difficult to tolerate with clever architecture tricks. There's a reason new chips have 3 levels of cache, and it ain't bandwidth.
The only way this will change is if the _workloads_ change. RAMBUS does beat PC133 in some cases, mostly that involve streaming data from ram to the cpu from a contiguous buffer. Any small or random accesses are going to hurt with RAMBUS.
But that's why they made DDR. It has the low(compared to RAMBUS) latency, and better bandwidth than RAMBUS. As Pokey would say, HOORAY! ^_^
The enemies of Democracy are
it's because Rambus simply hasn't delivered. they keep saying it will enhance performance later on down the road when chips are faster, but in the benchmarks, the faster the chip, the better SDRAM is at slaughtering Rambus. wait until DDR. that's honest performance gain hopefully unlicensed by anyone. cheap, fast. just like the athlon. tswhat makes computer shopping so much fun.
Rambus is a design for a memory system from Rambus Inc. It is extraordinarily fast on paper. Intel chose their design and decided to support it on a lot of their new products.
The implementation took a long time to get around to getting around. It is now here. Intel bet a LOT on Rambus, because it would give them significant control over a lot of markets. (IE: They own rambus designs)
Rambus is significantly different from the DRAM used commonly today. It requires changes to how stuff is laid out on the motherboard. And it is manufactured differently, to very demanding tolerances.
It is now in production and is competing with DDR-DRAM, which uses existing manufacturing processes, generally works with existing chipsets, and is easy to support. And it doesn't require a fan setup for the memory alone. And runs far cooler. And gives almost as good performance when set up correctly as a RAMBUS setup. And is also capable of being manufactured in quantity, whereas RDRAM is extremely difficult to manufacture. DDRDRAM is also about a fifth of the cost of a RDRAM setup.
You do the math, and read up on it a bit.. I think you will agree that for all intents and purposes (read: mainstream pcs, servers, et al), Rambus is DOA.
You hit the nail on the head, except...
RAMBUS gives almost as good performance as PC133, not DDR gives almost as good performance as RAMBUS, as you said.
Last year I had the priviledge of excersizing my masochism by working on an RDRAM mobo. While reading up on the architecture I said (aloud) "Gee, all this bandwidth is great and all, but the latency is so high won't that kill performance?" The Rambus Inc(tm) PowerPoint slides said no, bandwidth is all that matters, ignore the high latency. But the engineers I worked with were also skeptical. And it turns out we were right to be so, because RAMBUS sucks it up in real-world performance.
Other than that, though, you're right. And it is a sunovabitch to design for. When you have pico-seconds of margin to deal with at the _motherboard trace_ level, you know you're screwed.
The enemies of Democracy are
RDRAM as a technology on the pc platform has only been in production for maybe 9 months. From $1000 for 128 megs to $270 in 9 months is an incredible drop in price.
I dont have a
Intel's Desktop 2001 Roadmap Update indicates direct Rambus DRAM (RDRAM) will only be used in the high-end desk-top market
So it's not like Intel is giving up on RDRAM altogether - just on the lower-end machines. If they were switching because one is better, then they would just drop RDRAM totally.
Being with you, it's just one epiphany after another
The release of their 815i chipset already pointed in this direction (Rambus didn't really like that move :)
Also, the 815i chipset seems to be faster then the 820i chipset (which uses expensive Rambus memory). Now, it looks like they're indeed going to drop it. Look at some of these articles:
Every expression is true, for a given value of 'true'
Good question! I think it's because of:
1) Area. Putting a reasonable amount of RAM (128 MB) onto a CPU core would make the chip huge. Die area is a big concern. That's why until recently the L2-caches were off-chip.
2) Process. DRAM uses a special process that allows for vertical capacitors to achieve maximum density. Using this same process in a chip core would be expensive and wastefull.
3) Effectivness. Putting the memory on chip wouldn't necessarily increase the speed that much. Powering the huge arrays of RAM still takes time. The delays on the traces and going on-off chip aren't the dominating factors.
4) Upgradeablitily. You can't go to Best Buy and get another 64MB when your ram is on-chip.
Eventually, you might see this. More likely to me, though, is just expanding cache hierarchies. You'll have a 128MB L5 cache to go with your 32GB of RAM. ^_^
The enemies of Democracy are
This comes as no major surprise, Intel announced not too long ago that it would include support for :), and the fact that the Tbird Athlon is already, IMHO, a better processor than the Pentium 3, Intel really needs to get rid of some of it's crazy ideas and get back down to business. Comparing the Sledgehammer with the Itainium, it doesnt look all that great for Intel either (of course, Itainium has been hyped forever already, and were just now starting to learn about Hammer, so who knows).
SDRAM with the Pentium 4 (where as the article stated, was originally going to only support Rambus). Considering AMD's push with DDR SDRAM with the Athlon, which is considerably cheaper than Rambus (duh
IIRC, Rambus is still going to be used solely for Itainium--of course in 20 years when Itainium is finally ready to ship (but only to Intel's bedbuddies like Dell for the first 6 months), maybe Rambus will finally be affordable.
It is silly to use Rambus RAM, since SDRAM has lower latency. Rambus RAM has higher bandwidth, but if you need bandwidth you can always interleave memory. (the idea is similar to RAID striping. byte 1 comes from one ram chip, byte 2 comes from the other, so you have the same latency and twice the bandwidth. If you want more bandwidth, use more controllers and more chips. You can't do anything about latency except make each chip faster, though, which is why there's nothing you can do with Rambus to make it have SDRAM latency.)
#define X(x,y) x##y
#define X(x,y) x##y
Peter Cordes ; e-mail: X(peter@cordes ,
Well this isn't a bad move on intels part since they really dont need their high end systems having a $800 price difference from AMDs. Certainly they've wasted a TONNE of money and time over rambus and i'm sure they wouldn't be loosing out to AMD now if they had better directed all that effort.
:) Without the burden on rambus this should give intel a fighthing chance against amd and bring prices down more :)
However as I see it the current big looser is toshiba who I think are one of the few DRAM manufacturers that agreed to pay rambus royalties on DIMMS, in order to continue their license to produce RIMMs.
For those of you that dont remember this, rambus turned round and claimed that they had also patented regular SDRAM as well as their funky rambus. They started putting pressure on companies who already licensed rimm technology to pay up for dimm tech also. Toshiba (i think) were one of the few that complied, scared of loosing lucrative rambus contracts.
Now they are stuck paying royalties to rambus for dimms... and without the big return on rimms they could seriously dent their business.
AMD on the other hand got it right and i'm well pleased