elimc said:
Of course, I personally believe that today's processors could easily soak up the BW from a RIMM9600. In fact, I wouldn't be suprised if the 2.8GHz P4 could eat up the BW from a 40GBs RAM stick. The BW needs at 10GHz will be much more important than the needs we have today and people will be willing to pay more for RAM.
Heh. I'm all for more bandwidth. IMO the PC-platform got stuck at PC133x64-bits for far too long. On the other hand, there are several areas that don't benefit from better main memory subsystems. And most benchmarks tend to do a lot of work on small data-sets, rather than increasing the problem size, which is what I often see as causing a need for increased processing capability. Thus bandwidth doesn't carry the marketing clout it could have, if we only had looked at problems that really DO benefit from higher performance, rather than estimating office or internet use performance (* cough *).
Simply an example which my meager language skills failed to make very clear.
Hard to say much about their stability at this point in time since the nForce is the only current example.
The laws of physics dictate that using a four layer motherboard with cheap components will result in poor stability. You can't have something for nothing. There are always tradeoffs. If the current high end 64-bit six layer DDR SDRAM motherboards have stability problems, I don't see how the 128-bit wide mobos on four layers using cheap components are going to do much better. Maybe the channels will have to be at right angles to each other. But that is just another tradeoff which results in decreased room for periphials on the mobo.
We'll see. I'd tend to agree in general, but the proof of the pudding is in the eating, and if these manufacturers ship kit that works, I'll accept that they have gotten good enough. I have some faith in Intels validation programmes. Actual implementation could well matter quite a bit, so I would be conservative in my MB brand choice, if I was in that market. But that may just be my suspicious nature.
Saem said:
DRDRAM's offered bandwidth isn't chewn up, yet. However, DDR's is not only chewn up but falls short and then some. This can easily be seen in benches where you can see 2.6-2.8GHz P4s blowing pas 3.0GHz ones. Tom did a fair number of OC tests.
Even if we pair ANY P4 performance class MPU with PC 2700 and a 64 bit interface we'd still be in a fair amount of trouble when it comes to being starved for bandwidth, to the point that the effects of latency would take a back seat.
This is both application dependent obviously, and dependent on just how large differences in latency you are talking about. There is little doubt that the on-chip memory controller of the Hammers will be a huge benefit. Preliminary data say that latency will be roughly half that of the Athlons, and that is enough to have a major impact. Of course, Real Men will buy Opterons with 128-bit busses.
With all this said and done. I don't believe Rambus has claims to the IP they've "grandfathered" and claim to own. SLDRAM, seems like it would have been a better choice.
I'm not touching that nest of vipers with a ten-foot pole. I stood at the sidelines enjoying the fight when these technologies slugged it out, and they seemed to offer similar potential. They have both lost in the marketplace now of course, but good ideas tend to hang around and be implemented at a later date.
Personally, I hold little hope for DDR II, since that entire multiplexing the bus and reducing memory array clocks to get more bandwidth doesn't sit well considering the fetch size will be a wee bit big. I'm sure people will use it, but it'll suck donkey balls for the most part.
Are we feeling grumpy today?
I'm not with you on the fetch size. I have the option of changing the burst length on this motherboard, and my experiments have shown only minor differences for actual applications. I'd be more concerned about lack of critical word first bursting, but while that should have an impact, it is relatively limited compared to the (projected) clock benefits.
What I am looking forward to is more intelligent memory. I think there has been too much centralization of "intelligence" in machines and I'm of the firm belief that this is the wrong way to go.
Check out Session Four on this years MicroProcessor Forum.
"Session Four: Extreme Processors
Max Baron, Senior Editor, Microprocessor Report; Principal Analyst, In-Stat/MDR Active Memory Device Delivers Massive Parallelism
Graham Kirsch, Chief Architect, Active Memory Program, Micron Technology, Inc.
Micron will present its Active Memory device architecture featuring closely interfaced on-chip processors and DRAM memory. Micron's Active Memory architecture achieves high-performance by taking advantage of on-chip distributed private memories that can deliver fast access times and high bandwidth."
Interesting, no?
It is regrettable that the current model for PC-evolution doesn't suit radical changes well. Optimums that require major redesign or rethinking of either hardware or software, (particularly together!) are unlikely ever to be reached, likewise if backwards compatibility is compromised.
Entropy
