Xdr Dram
The optimum memory for PPC 970 systems is here, XDR DRAM, aka Rambus' Yellowstone. Offering effective datarates up to 3.2 GHz, 25.6 GByte/s, with a roadmap up to 6.4 GHz. Sustained datarates can be from 6.4 GBytes/s to 12.8 GByte/s.
Too bad Apple is following the big boys here and going with DDR memory...
Too bad Apple is following the big boys here and going with DDR memory...
Comments
Who would want to deal with a company or technology where one group thinks they have control over it and if they don't think you are using it the way they want you to, they sue. Not me, been involved with groups like that.
Originally posted by Bigc
Don't think Rambus made a lot of friends after their law suits over DDR technology.
The reality isn't what it appears to be. If by magic we learn all that happened, we'll learn that all involved have dirt on their hands, and we may be surprised who were the saints and who were the sinners in the imbroglio.
Who would want to deal with a company or technology where one group thinks they have control over it and if they don't think you are using it the way they want you to, they sue. Not me, been involved with groups like that.
If Apple wants the world's fastest desktop computer, then maybe they should use the world's fastest memory.
Originally posted by THT
The optimum memory for PPC 970 systems is here, XDR DRAM, aka Rambus' Yellowstone. Offering effective datarates up to 3.2 GHz, 25.6 GByte/s, with a roadmap up to 6.4 GHz. Sustained datarates can be from 6.4 GBytes/s to 12.8 GByte/s.
Too bad Apple is following the big boys here and going with DDR memory...
Isn't this kind of mem-tech slated for mass-production in 2005?
If so, I think Apple "easily" can change from using DDR to XDR, and then supplement whatever memory-eating beast (G6) they're going to use then. I'd bet alot on that the time between now and 2005 will involve a major topology-change for processors/memory/motherboard, wouldn't you?
Originally posted by Zapchud
Isn't this kind of mem-tech slated for mass-production in 2005?
Yes. Apple's volume is small such that I hope they can ship 1 GHz to 2 GHz parts in 2004. For the dual G5 Power Mac, the two G5 machines can suck up 10+ GB/s of bandwidth though, and will only get worse in 2004 when PPC 970s between 2 to 3 GHz will ship.
If so, I think Apple "easily" can change from using DDR to XDR, and then supplement whatever memory-eating beast (G6) they're going to use then. I'd bet alot on that the time between now and 2005 will involve a major topology-change for processors/memory/motherboard, wouldn't you?
No, not that much in topology. Only AMD will have a NUMA style memory topology. It looks like Apple and Intel are going with shared memory systems, which is fine for dual systems, maybe even for quad systems.
A concern: Latency? How is it? I fail to find any info on it (I might be blind, though), but even with this kind of massive of bandwidth, there will be non-bandwidth-bound ops that are more bound by latency. This is, as you know, a very bad thing about current RAMBUS tech.
Another thing, I noticed that XDR uses Octal Data Rate. This would probably not hurt too much when streaming large datasets, but wouldn't it make XDR somewhat less effective compared to the max theoretical specs, in comparision to DDR now?
But don't get me wrong, I'm not "against" XDR, I'm just trying to get the correct perception of the tech, which seems pretty godly, judging from what RAMBUS itself advertises this as
Originally posted by Zapchud
A concern: Latency? How is it?
The only info I have found is: "1.25/2.0/2.5/3.33 ns request packets". I don't know if that is good or bad. I'm not sure if DRDRAM is the same as XDR DRAM so they may have different issues, how to equate that latency figure to a system. I'm sure it's just for one XDR DRAM and going from the XDR controller to the DRAM itself.
I fail to find any info on it (I might be blind, though), but even with this kind of massive of bandwidth, there will be non-bandwidth-bound ops that are more bound by latency. This is, as you know, a very bad thing about current RAMBUS tech.
"Very bad" is overstating it. I don't even think it can be called bad, maybe a deficiency, and it was mostly used to demonize Rambus. So, the propoganda made it out to be worse than it appeared. In the end P4 systems with Rambus were the fastest systems available at the time. The latest generation Alpha machine uses Rambus. The Sony PS2 Emotion Engine uses Rambus, and in all likelihood, the PS3 will use XDR DRAM.
We'll see if third party Rambus chipsets (SIS) for the P4 will be able to compete with dual DDR systems in the near future. I bet SIS's quad channel PC1066/1200 system will be faster than dual DDR400 systems by 10% overall. Even though it'll be bottlenecked by the 800 MHz P4 bus, it'll have better sustained bandwidths then dual DDR400 will.
Another thing, I noticed that XDR uses Octal Data Rate. This would probably not hurt too much when streaming large datasets, but wouldn't it make XDR somewhat less effective compared to the max theoretical specs, in comparision to DDR now?
No, I wouldn't think so. Rambus always has had higher sustained bandwidth compared to the theoretical than DDR SDRAM did. Ie, it had better bus utilization. So, I don't think this will be different. The issues with Rambus and XDR DRAM can be mitigated pretty well. The P4 did this very well, and conincidentally, the PPC 970 have some of the same features.
The PPC 970 has 128 byte cachelines and good hardware prefetching, so initial latency won't hurt as much as it would with the G4's 32 byte cacheline and poorer hardware prefetch mechanisms.
Originally posted by THT
The reality isn't what it appears to be. If by magic we learn all that happened, we'll learn that all involved have dirt on their hands, and we may be surprised who were the saints and who were the sinners in the imbroglio.
RamBus ought to go down for their actions, but that's not to say others are saints. The business world gets pretty nasty, especially where it comes to standards and patents.
Re: G5 memory latency. There is already a lot of latency in Apple's new system, XDR would make it considerably work. On the other hand the 970 is designed for this kind of high latency memory system, and I'm sure by the time XDR arrives the latest PowerPCs will be even more capable. Bigger caches, more OoOE, SMT, etc.
Originally posted by THT
The Sony PS2 Emotion Engine uses Rambus
Yes, and it suffers for it very badly.
You're right, however, modern processor designs are getting better at dealing with the long latencies. SMT / hyperthreading, in particular, will be a big improvement.
Originally posted by THT
The only info I have found is: "1.25/2.0/2.5/3.33 ns request packets". I don't know if that is good or bad. I'm not sure if DRDRAM is the same as XDR DRAM so they may have different issues, how to equate that latency figure to a system. I'm sure it's just for one XDR DRAM and going from the XDR controller to the DRAM itself.
Latencies for the good'ol RAMBUS' were something like ~30 cycles, if memory serves me right. that's nearly 30ns on 1066MHz RAM (but that's not the real clock of it, half, so I guess that makes it 60?). Okay, I'm somewhat confused here, but take the seemingly "worst case" of those, 3.33ns, and it'll be ~10 cycles when running the RAM at 3.2GHz.
Well, any way I see it, RAMBUS seems to have resolved their latency-issues, and unless I'm ridiculously mistaken in my "calculations" (or if memory is playing me a trick), latency will be a non-issue with XDR RAM! Excellent
Originally posted by THT
Too bad Apple is following the big boys here and going with DDR memory...
Yeah, they should have delayed the G5 until 2005 and increased the price so that they could use better RAM. Not.
Who would want to deal with a company or technology where one group thinks they have control over it and if they don't think you are using it the way they want you to, they sue. Not me, been involved with groups like that
That sounds like a company I know of that I really like. Oh yea...
Seriously though, it's better if Apple goes with industry standards on this one. G5s are already totally out of price range for average PC buyers. And unless their prices come down or they release a Cube or MASSIVELY speed up iMacs then G5 towers will remain out of reach for average buyers which is a big mistake but hey Apple likes making big mistakes. Adding expensive proprietary tech isn't going to lower prices. The G5s are already so fast that price is more important at this point. And there is faster DDR out there that Apple isn't using yet but will upgrade to later.
Originally posted by wmf
Yeah, they should have delayed the G5 until 2005 and increased the price so that they could use better RAM. Not.
Why delay when they could just have easily went with quad channel Rambus instead dual DDR channels. Quad channel PC800, PC1066 and PC1200 RDRAM would give 6.4 GB/s, 8.5 GB/s and 9.6 GB/s memory bandwidths. Going quad channel also will alleviate a lot of the latency "issues" and probably give twice the sustained bandwidth that dual DDR400 would.
XDR DRAM could be used next year when memory becomes even more of bottleneck when 3 GHz PPC 970 buses are at 1.5 GHz and would need 12 GB/s bandwidth per processor.
Originally posted by THT
Why delay when they could just have easily went with quad channel Rambus instead dual DDR channels.
I doubt they could have "easily" gone with Rambus. First of all, Apple's experience is with DDR and that is what their existing supply chain carries (at least in the DDR333 variety which the lowend G5 uses). Second there is the licensing fees for Rambus protocols. And third, the current market in Rambus memroy is not very appealing since the PC has steered away from it. It is in Apple's best interests to use the same commodity RAM as the majority of PCs. If the PC adopts XDR in a significant way then Apple could jump on board, but the technology issues here are really secondary to the supply issues.
Originally posted by Programmer
I doubt they could have "easily" gone with Rambus.
Probably. Engineering-wise, I'm not sure implementing dual channel DDR400 is easier than quad channel RDRDAM, but I can see your point.
It is in Apple's best interests to use the same commodity RAM as the majority of PCs. If the PC adopts XDR in a significant way then Apple could jump on board, but the technology issues here are really secondary to the supply issues.
Thus, I said Apple is going with the big boys. But technology-wise, RDRAM and XDR RAM provide the necessary bandwidth for PPC 970 chips now and in the future, moreso than DDR SDRAM. It can give Apple an additional 10% performance boost as 970 FSB bandwidths outstrip DDR SDRAM bandwidths, nevermind that it already does in the dual 2 GHz system.