JEDEC reveals DDR4 roadmaps and specs

[Manicdan]

the table is pretty clear, i was more looking at the BIGGER chart (caps required due to image size, lol)
since alot of what is being talked about is when standards are established, i figured thats what people should be keeping the context about unless otherwise stated.

as far as accuracy goes, it looks a little to simplistic, did ddr3 really come out retail in end of 07?

[Particle]

DDR3 launched in 2007 but didn't really become mainstream until maybe 2009. It has only really been ubiquitous as of this year.

[WaterFlex]

Wow, thanks for this. DDR4 seems to be the Future

[Darakian]

Am I reading this right? Every DIMM is to have it's own channel?

[Particle]

Kind of. Hard to say what sort of link exists between these "switch fabric" logic blocks and the CPU. Chances are it's smaller than four full DIMM links would be or else there would be little reason to go with the breakout design. DIMMs are already going to be running at what, 1-2GHz? (anyone know if DDR4 is going to use DDR signaling on both a regular and phase shifted clock signal or just regular DDR?) A high frequency link with fewer I/O lines might save you some I/O lines. Dunno if it'd be worth it.

[nn_step]

unfortunately the current JEDEC spec for DDR4 lacks the several controversial features that would have make it worthwhile much earlier.

[Darakian]

Ya, I'm not sure on the switch fabric either. Seeing the word switch, I'd like to say that it'll work something like a network switch. Hopefully with a larger back-end :p. Off of the desktop platform and off the switch fabrics it looks like 4 independent channels. Quad channel ram? :3

[Oranuro]

And how is migrating to these faster memory chips beneficial to anyone? Sure benches show the minuscule improvements over each generation, but what real gains are we looking at here besides maybe the lower power draw of DDR4? Front side bus is gone, now we are looking at the QPI's and HTLINKS/HYPERTRANSPORTS and what have you. Microprocessors have been stagnant in megahertz gains for the past 6 to 10 years, outside of overclocking, neither IBM, Intel nor AMD have been able to mass produce 4 gigahertz microprocessors to the public. I know there's the argument that there is a need for faster memory for memory intensive applications, but memory running faster than the logical processor(s) is surely gonna become some sorta bottleneck to the overall platform I'm guessing.

I dunno, with the roadmaps laid out for Intel and AMD for the next five years, is DDR4 even needed?

[kiwi]

I dunno, with the roadmaps laid out for Intel and AMD for the next five years, is DDR4 even needed?

Nope, we need faster hard drives

[Manabu]

I dunno, with the roadmaps laid out for Intel and AMD for the next five years, is DDR4 even needed?

Yes, very much.

Faster DDR3 and DDR4 are essential for AMD Fusion, and maybe for intel's IGP too (if they become serious about graphics speed). Dual-channel DDR3 PC-1600, that should be common place next year, has the same bandwidth than the Radeon HD5750 with it's 128bit 800mhz GDDR5. But this will have to be shared with the CPU, so even the first generation Fusion may be limited by memory speed, and ask for faster bins to have it's potential adequatedly in use. The third generation will surely want DDR4, else it will not even be possible to double the GPU power. And I doubt triple channel memory will grow beyond enthusiats crowd any time soon, if ever.

That besides having to share 512mb+ of ram for IGP, and Windows 7 64bits, whose will make people want 4~8GB in their systems. Memory makers will be busy in comming years.

[Particle]

I think the more important question to ask would be, "Do you really want to stop making memory faster?" To answer in the affirmative would be absurd.

Faster memory is always a good thing. Lower power is also great. Max performance gains may be small, but that's the world we live in anymore. A few percent here and a few percent there add up to offer what performance gains we have. As for what benefits, anything with a large dataset will benefit. Laptops and other battery powered computing devices are going to like this as well.

[hecktic]

DDR3 2133 @ 1.25v is possible but only on a case by case basis by underclocking a higher spec bin set. And there are tons of good sets out there for DDR3 2500

[Serra]

Sweet, they're going to a switch fabric system. Presuming this works for this system like it does for others that's pretty positive, definitely an "in" thing to do right now. With larger systems I've seen some fabrics that are fully line-rate, and some that aren't, so it's hard to say how that will work. It might even be that you could buy some that will be full-speed and cost more and others that will be 80% of max speed but much lower in price (and still deliver more than enough bandwidth for most people).

To carry it over from larger systems for those who aren't aware, one nice thing about a switch fabric is that you can present it to the CPU by just saying "I'm memory, toss stuff at me without trying to perform any calculations on how to do it best", and typically the fabric will be designed to handle any manipulation/best path calculations/whatever else it may offer in hardware (eg. at full speed). So although the bus itself is like an extra step in the way, it actually saves time in the end by letting the CPU do a bit less work and only delaying information as long as it takes to go through a slightly longer circuit to its ultimate destination.

[Particle]

It's funny how we keep going back and forth between one thing and another. Just like we keep alternating back and forth between serial and parallel data links, we keep moving back and forth between full integration and moving functions out onto discrete dies/chips.

[nn_step]

And how is migrating to these faster memory chips beneficial to anyone? Sure benches show the minuscule improvements over each generation, but what real gains are we looking at here besides maybe the lower power draw of DDR4? Front side bus is gone, now we are looking at the QPI's and HTLINKS/HYPERTRANSPORTS and what have you. Microprocessors have been stagnant in megahertz gains for the past 6 to 10 years, outside of overclocking, neither IBM, Intel nor AMD have been able to mass produce 4 gigahertz microprocessors to the public. I know there's the argument that there is a need for faster memory for memory intensive applications, but memory running faster than the logical processor(s) is surely gonna become some sorta bottleneck to the overall platform I'm guessing.

I dunno, with the roadmaps laid out for Intel and AMD for the next five years, is DDR4 even needed?

Depends for what exactly you are talking about.

In terms of latency sensitive applications. NO

In terms of bandwidth intensive applications. YES

Nope, we need faster hard drives

One of many things that must be improved to improve overall system performance

Yes, very much.

Faster DDR3 and DDR4 are essential for AMD Fusion, and maybe for intel's IGP too (if they become serious about graphics speed). Dual-channel DDR3 PC-1600, that should be common place next year, has the same bandwidth than the Radeon HD5750 with it's 128bit 800mhz GDDR5. But this will have to be shared with the CPU, so even the first generation Fusion may be limited by memory speed, and ask for faster bins to have it's potential adequatedly in use. The third generation will surely want DDR4, else it will not even be possible to double the GPU power. And I doubt triple channel memory will grow beyond enthusiats crowd any time soon, if ever.

That besides having to share 512mb+ of ram for IGP, and Windows 7 64bits, whose will make people want 4~8GB in their systems. Memory makers will be busy in comming years.

Side channel GDDR4 is also an option but I doubt we will see that for any real serious performance products.

I think the more important question to ask would be, "Do you really want to stop making memory faster?" To answer in the affirmative would be absurd.

Faster memory is always a good thing. Lower power is also great. Max performance gains may be small, but that's the world we live in anymore. A few percent here and a few percent there add up to offer what performance gains we have. As for what benefits, anything with a large dataset will benefit. Laptops and other battery powered computing devices are going to like this as well.

In terms of latency, we are literally at the wall given capacities that are involved. [You need smaller memory to make memory lower latency]

In terms of bandwidth, we can continue to see slow steady progress forward for another decade. Beyond that no one knows.

[Glow9]

Anyone else find this to be shady end of the tech business. I swear the screw around with prices so much I'm never jumping into whatever is new with these guys.

[saaya]

interesting graphs...
but they are more of a general overview i think... they dont look very accurate, then again i dont think they were ever meant to be very accurate

this is already possible with ddr3 afaik...

In terms of latency, we are literally at the wall given capacities that are involved. [You need smaller memory to make memory lower latency]

In terms of bandwidth, we can continue to see slow steady progress forward for another decade. Beyond that no one knows.

true... diont you think theres some room for improvement through caching branching and prefetching though? still thats going to happen either in the imc or memory switch chip...

[Particle]

I like to remind people that the end of the world always appears to be just around the corner. Let us not be so short sighted as to think there is a serious chance of progress stopping just because we can't see the way past a point today. Things will continue to get faster, smaller, cheaper, even if we don't see how today. There's sure to be an ultimate end to density, but we aren't getting there in a decade or even two decades.

[Sn0wm@n]

particle is speaking the truth

[saaya]

I like to remind people that the end of the world always appears to be just around the corner. Let us not be so short sighted as to think there is a serious chance of progress stopping just because we can't see the way past a point today. Things will continue to get faster, smaller, cheaper, even if we don't see how today. There's sure to be an ultimate end to density, but we aren't getting there in a decade or even two decades.

i actually disagree...
there's always a point of diminishing returns, and if you ask me, late ddr2 already reached it... the demand for memory performance grows, sure, but at a veeery slow pace, and for most consumer aps its barely growing at all...
single channel ddr3 is just as fast as tripple channel ddr3 in 90% of todays apps... need i say more?
as if youd even remotely notice a difference between ddr3 1333 and ddr3 2000 without using a benchmark and compare numbers...
why do you think the ddr3 adoption took this long? cause it just didnt make sense to upgrade to ddr3, cause it cost more and nobody would notice the difference between a ddr2 and ddr3 pc...
dont you think if thered have been a notable performance boost people would have wanted ddr3, and oems and shops would have equipped their pcs with ddr3 to have a competitive advantage over other systems?

just look at whats important to have a good pc experience nowadays...
1.) display
2.) vga
3.) ssd
4.) cpu
5.) memory capacity
6.) memory speed

and then theres internet connection which id rate no2 actually, some might actually rate it no1...

anyways, memory speed makes pretty much the smallest impact of system performance these days... and even if ddr7 came out tomorrow, 99% of the people wouldnt even care...

[Manicdan]

and if you ask me, late ddr2 already reached it...

+1, 800mhz ddr2 running CAS 4 or 5 is fast enough for almost every common task or game. as long as you have the minimum memory required, you do good enough.

[Particle]

I cannot believe how short sighted some of you continue to be. Without us today pursing tomorrow's technology, next week's is never going to come. It just doesn't make any sense to stop unless you're really suggesting that memory is fast enough for the rest of time. That would be the only reason to not advance. Nobody is going to make that claim, so you really have no choice in what action is appropriate. Keep advancing as is economically and technologically feasible or simply die outright--those are the only two options.

Here is another thing to think about. Why do you think CPUs have cache? There is no mystery here, and it is no matter of opinion. Cache is present because system memory is so darn slow compared to the CPU's logic. System memory is in turn present because hard drives are so slow. Hard drives are present because of the ultimate need for capacity. This is a pretty big deal when as much as half of a CPU die is dedicated to SRAM cells. If system memory were as fast as cache instead of where it is now, your CPU wouldn't need cache at all and that extra 50% of the transistor budget might be absent (cheaper CPUs) or used for logic (faster CPUs). It would be better still if it were as fast as the CPU's registers. It's not possible to make system memory that fast of course with the two most obvious problems being physical distance versus clock speed and capacity due to the lower density of faster memory technologies. The closer you can get system memory to that point, however, the better off the design is.

It may not be interesting to you, but it's not a bad thing. I don't get where the determination to dog this technology so hard is coming from.

[Oranuro]

What Saaya said is where I was going with my post, SDRAM is just SDRAM. No matter what the iteration is it's just memory. You do not get the impression of a performance boost when moving from DDR2 to DDR3, and the power savings you may not even recuperate within 10 years. The move from "SDRAM" to DDR resulted in boosts in performance in most day-to-day applications, I mean you could really see the difference depending on what you where doing, and it is because Both AMD and Intel platforms could use it, both platforms really thrived off of the change.

Fast forward to today, three generations later and the performance margins are much smaller than what DDR brought to the table. Will the new iteration bring lower prices of SDRAM to the market?(even high density which is much needed), will it force current microprocessor and chipset producers to change their architectures to really utilize and unleash the proposed power of DDR4? As many of you have already said, DDR3 can already meet these specs or better.

[saaya]

I cannot believe how short sighted some of you continue to be. Without us today pursing tomorrow's technology, next week's is never going to come. It just doesn't make any sense to stop unless you're really suggesting that memory is fast enough for the rest of time. That would be the only reason to not advance. Nobody is going to make that claim, so you really have no choice in what action is appropriate. Keep advancing as is economically and technologically feasible or simply die outright--those are the only two options.

Here is another thing to think about. Why do you think CPUs have cache? There is no mystery here, and it is no matter of opinion. Cache is present because system memory is so darn slow compared to the CPU's logic. System memory is in turn present because hard drives are so slow. Hard drives are present because of the ultimate need for capacity. This is a pretty big deal when as much as half of a CPU die is dedicated to SRAM cells. If system memory were as fast as cache instead of where it is now, your CPU wouldn't need cache at all and that extra 50% of the transistor budget might be absent (cheaper CPUs) or used for logic (faster CPUs). It would be better still if it were as fast as the CPU's registers. It's not possible to make system memory that fast of course with the two most obvious problems being physical distance versus clock speed and capacity due to the lower density of faster memory technologies. The closer you can get system memory to that point, however, the better off the design is.

It may not be interesting to you, but it's not a bad thing. I don't get where the determination to dog this technology so hard is coming from.

nobody is saying we should stop ddr4 research
once we need it, it will be ready, im sure... actually its much more likely that itll be ready BEFORE we need it

what im saying is that:
1.) itll be quite a while before it comes out
2.) i wouldnt expect any notable performance boosts from ddr4 once it does...

[Manicdan]

why has memory become a minor perf boost? back in DDR1 days it was pretty noticeable, have new instructions relied more on cache than memory chips?