Rambus RAM... What's wrong with it?

[WINDSHEAR]

I've seen very little of RDRAM, but I can't help but wonder why it isn't better, since it is so much faster? 1066Mhz sounds pretty fast to me... is it because it's so expensive? I know very little about this... :?

 

[RambusTech]

Nothing is wrong with it.

Intel arss-fu*&ed us.

 

[nobly]

Yeah for starters its expensive... Really expensive! Relatively speaking, i suppose Its funny that my RDRAM was cheaper when i bought it 6 yrs ago than it is now...

Another thing is that it has high latency for random access. It has a lower latency if the memory addressed is readdressed, etc. That's why it was seen as better than DDR266/333 (what it originally went up against w/ RD800), especially for audio/video editing.

But with the advent of DDR400 and when DDR400 finally went dual-channel, it kind of brought DDR really close in performance to RD. Ask anyone which they would choose (a $50/128MB for RD or a $30/128MB for DDR), they choose DDR. Plus licensing stuff blah blah blah everyone started hating Rambus, etc etc.

Another thing was that it was almost impossible to overclock because of the strict timings that were needed. RDRAM was very very specific. Each wire on the chip had to be so long. Plus for RDRAM800 it only did 16bit interfaces, which meant we were back to the 'install in pairs' thing. RD1066 was the answer to DDR400 but by then everyone wanted to save $$$ and get more memory.

Funny thing is that the PS3 will use XDR - Rambus again. Should be interesting to see how that'll stack up.

That's how I saw it went down. I might be a little off on the timings and such, so hopefully someone will correct me.

RDRAM's high price is from the licensing costs...
Intel gave up on RDRAM because no one wanted to pay that much.

 

[theaxemaster]

I used to use it (had 512 of RD800). I had both a P3 and a P4 board that I used it on and it was very fast. It really is good stuff. When I switched to DDR400 due to the inability to find new, featureful boards that had it, my system slowed down noticably. Granted I didn't spend as much on the RD800 as I did on the DDR, which would probably explain the slowdown.

I believe it was originally intended as server memory that some tried to push into the desktop market. I think in the server market the install in pairs bit wasn't as big a deal as it is in the desktop market. Yes it was expensive relative to DDR at the time, although to buy DDR now with similar performance, the price difference would probably be negligible.

But hey, I sold the stuff for the same price I paid for it, which is probably the first time in memory history that that was possible

 

[crizazykid2]

It is OLD.

 

[the_guru]

I've seen very little of RDRAM, but I can't help but wonder why it isn't better, since it is so much faster? 1066Mhz sounds pretty fast to me... is it because it's so expensive? I know very little about this... :?

First it's expensive to manufacture. Second MHz is not always what counts. Latencies is equally important. That's why AMD 64 with 400MHz memory is better than Pentium 4 with 800MHz memory.

 

[nobly]

yeah when I built my P4 RDRAM was the only way to go. And it outperformed DDR266/333 for the most part. Plus P4's didn't support DDR then.
I paid $80 for 2x128MB 6 years ago. I upgraded last year to get another 2x128 for $90.

Obviously we can't compare anymore because RDRAM is pretty old. And DDR400 and more can make some 1T 2T timings.

My computer 1.8 P4 512MB RDRAM800 can edit audio files faster than my Athlon XP 2500 (1.8Ghz) 512MB DDR333. I would have gotten DDR400 but that processor didn't support it.

All in all, it wasn't a bad bit of technology, but since then, DDR400 has made a pretty good case for getting close to that performance. Nothing wrong w/ RDRAM, its just been bypassed like all technologies.

But if you start talking about like RDRAM1200, then yeah, I think DDR400 would be worse comparatively. But it would probably be like 3x cheaper

 

[mpjesse]

RDRAM is still used in PS2's! XDR will also be used in the PS3.

RDRAM is techincally (from a bandwidth and technology standpoint) better than the original DDR. However, because Rambus is a technology IP company they have to charge royalties to anyone who manufactures RAM based on RDRAM patents (they have to do this to make money). Rambus does not actually make memory chips- they make the technology. The memory industry is already cut throat as it is and no one really wants to pay a 5% royalty.

There's a lot of history between Rambus, Infineon, etc that I won't go into. Rambus has left a sour taste in a lot of memory manufacturers mouths. The reasons are complex- but both sides have legitimate gripes.

Watch for Rambus to make a come back. AMD signed a $75 million dollar licensing deal w/ Rambus a couple months ago that <might> lead to XDR tech in AMD chips. Rambus also holds a few patents in the PCI-E arena that AMD has licensed.

1066Mhz sounds pretty fast to me...

Oh yeah, it's fast. However, the original and subsquent versions of RDRAM only had data paths 16 bits wide which allowed for a very high clock speed. On the flip side, DDR has a 128 bit (in dual channel) wide data path. Since the data path in DDR is huge the clock speed has to be lower (or at least, it used to be).

For the record (as some pointed out in a different Rambus thread), Rambus invented dual edge clocking (aka, DDR). JEDEC sort of stole it... Again- I won't go into it 'cause it'll create another rambus flame thread. Just know that RDRAM and XDR are actually good technologies... the business plan of Rambus however wasn't so good.

-mpjesse

 

[mpjesse]

RDRAM's high price is from the licensing costs...

That's actually not totally accurate. The high price is due to the basic supply and demand of RDRAM. Not a lot of RDRAM is made (nor was it ever). Hence... the supply is small and the costs are high.

Infineon and others liked to cite the licensing costs as an excuse for the high cost of RDRAM (this was more a P.R. move to get everyone pissed off at RDRAM). That turned out to be complete B.S. That was very evident back in 2002 when memory prices for SDRAM and DDR bottomed out to all time lows.

;-)

-mpjesse

 

[theaxemaster]

mmmm....16 GB/s theoretical throughput XDR, tasty!

 

[RambusTech]

The latency was only an issue with the 600-800Mhz modules. Once you got the 32-bit 1066Mhz modules, the latency was effectively halved.

The latency was even FURTHER reduced by the 64-bit quad channel RDRAM that was just around the corner when intel decided they could no longer keep up with OUR memory speeds and dumped us.

Yes, it was expensive, but DDR still hasn't reached the speed of RDRAM, and it's been almost 5 years.

 

[darkstar782]

The clock speed was fast, but it had a much narrower datapath as has already been mentioned.

If you push 16 bits per clock down at 800MHz, then the data rate is the same as pushing 64 bits per clock down at 200MHz. Therefore 266MHz DDR had a higher bandwidth than 800MHz RDRAM.

Add to this the latency issues that the 800MHz stuff had, and the cost, and you see why people switched.

The 32bit Dual Channel modules were better (and ran at 1066MHz as previously mentioned, but by then DDR had dual channel too, so its 128bits vs 32bits, and Dual Channel DDR 400MHz still had a higher throughput.

The narrower data pathway did however make motherboard design simpler.... and as we have seen there is a celing on massively Paralell technologies (PCI givs way to PCI-E, PATA to SATA, Parallel to USB/FireWire, due to crosstalk.)

 

[dvdpiddy]

cause of the crap they put the industry through thats why i hate rambus!

 

[captaincharisma]

it failed because it was expensive and only can be installed in pairs to work

 

[dvdpiddy]

that too but the licensing really pissed me off

 

[RambusTech]

The clock speed was fast, but it had a much narrower datapath as has already been mentioned.

If you push 16 bits per clock down at 800MHz, then the data rate is the same as pushing 64 bits per clock down at 200MHz. Therefore 266MHz DDR had a higher bandwidth than 800MHz RDRAM.

Add to this the latency issues that the 800MHz stuff had, and the cost, and you see why people switched.

The 32bit Dual Channel modules were better (and ran at 1066MHz as previously mentioned, but by then DDR had dual channel too, so its 128bits vs 32bits, and Dual Channel DDR 400MHz still had a higher throughput.

The narrower data pathway did however make motherboard design simpler.... and as we have seen there is a celing on massively Paralell technologies (PCI givs way to PCI-E, PATA to SATA, Parallel to USB/FireWire, due to crosstalk.)

Not that I'm nitpicking, but your understanding of the technical aspects of "bits per clock" is a little flawed, or you're using the wrong term to express what you're trying to say. I know what you mean, of course, but it's the wrong term. Consequently your final judgement that 266Mhz DDR has higher bandwidth than 800Mhz RDRAM is also incorrect. The 800Mhz 16-Bit modules were installed in pairs, and there were two channels. (one channel for each pair.) In the 32-bit modules, this pairing was done on-module, and there were still usually two channels per board. This is where it gets really confusing because in reality it was a 4 module setup, being expressed as 2 channels on the board. Ahhhh.. I degress.

Lastly, the 32bit modules ran at not only 1066Mhz, but 1200Mhz also. To which I have several pieces of myself. Samsung also has a few pair of 1333Mhz modules that were basically sitting on the shelf waiting for a memory controller (and a main BUS that could handle the throughput.) Plans for such a controller were in the works from SiS, but were scrapped due to the lack of market support after Intel's departure.

As for bandwidth, You see... the problem wasn't with RDRAM, it was with the available bandwidth between the CPU and northbridge that was holding back the CAPABLE throughput of the RDRAM modules. Unfortunately, Intel couldn't keep up (or didn't want to) and dropped us to save face rather than look like they couldn't innovate.

Quad-channel 1600Mhz 64-bit modules (which were TAPED OUT in 2003) would be worthless with even the latest Intel CPU's. The memory was just too fast for them.

However, new light has been sparked into Rambus, and I look forward to the day that XDR shines on the AMD platform. It will be interesting to see how a computer performs when there is almost no main system BUS limitation to restrict memory performance. It all depends on the success of the PS3 however.

 

[mpjesse]

Yes, it was expensive, but DDR still hasn't reached the speed of RDRAM, and it's been almost 5 years.

Huh? DDR400 has the same theoretical bandwidth as RDRAM. 3.2GB/s. I wont even go into DDR400+ speeds as they are not JEDEC approved.

-mpjesse

 

[RambusTech]

Huh? DDR400 has the same theoretical bandwidth as RDRAM. 3.2GB/s. I wont even go into DDR400+ speeds as they are not JEDEC approved.

You're forgetting that the bandwidth of the RDRAM wasn't limited by the bandwidth of the components themselves on the module, but by the passthrough on the northbridge.

 

[mpjesse]

Ooook... but what good is that if you can't use the RAM w/o a northbridge processor?

I know what ur saying... but in real life DDR400 and RDRAM PC800 have the same amount of bandwidth.

I suspect PC1066 has more than 3.2GB/s of bandwidth?

-mpjesse

 

[RambusTech]

I know what ur saying... but in real life DDR400 and RDRAM PC800 have the same amount of bandwidth.

I suspect PC1066 has more than 3.2GB/s of bandwidth?

The problem was that the FSB simply could NOT communicate fast enough with system components to push the RDRAM at the speed that it was capable of. Intel gave us a roadmap that was incredibly overambiguous and when they couldn't deliver on their technology outlook, they dumped us so that they wouldn't look like they got caught with their pants down. Changing from 16-bit to quad channel 64-bit was as simple as swapping out the memory controller and the technology was ON THE SHELF, WORKING.

We of course, made it all too easy for them to abandon us by simply doing what any reasonable intellectual property rights holder would do - We pursued litigation against members of the JDEC forum for infringement on some of our patents. It's a double edged sword, either you protect your rights to your creations, or you let other companies take advantage of the situation. I'd like to see Motorola come out with an MP3 player shaped like a small white rectangle, and name it the "M-Pod Wano". US courts would literally shit themselves shutting that one down. Semiconductor Engineering is a different ballgame - Nobody understands it without a degree in the field. Hell, I doubt there's more than 2 people reading this that even know what a RAS to CAS Precharge delay REALLY is, 99.999% of america would be stumped by even basic terms like tRAS, tCAC, etc... To try explaining how you rotate a phase of electricity so that you can transmit 8 seperate wavelengths down one conduit to a judge is a pathetic attempt at best.

Another point is that we simply DO NOT have the capital to manufacture our own components in any mass quantities. Thus - We rely heavily on other companies to produce, manufacture and distribute the modules/components for us. These organizations are by nature also going to be a part of the same JDEC forum that we are currently in litigation with, there just isn't any way around it.

Edit - Also, comparing DDR 400 and 800Mhz RDRAM is silly. The first DDR 400 board didn't come out until September of 2002. By that time, 800Mhz RDRAM had been around for almost 3 years.

 

[dvdpiddy]

hey rambus tech do you think that if rdram had 32 bit at start it would've succeded?

 

[RambusTech]

hey rambus tech do you think that if rdram had 32 bit at start it would've succeded?

Well, as I said, the problem wasn't really with RDRAM itself. It was that it was too fast for the platform and throwing raw power at an already saturated BUS would have been useless.