[AlStrong]

It's somewhat vague, but this sounds like it ranges from having DRAMs on-package to any of the various stacked options. You might even include DRAMs soldered onto the motherboard in terms of user inaccessibility, but Intel is probably just thinking about on-package solutions for now.

RSX already has on-package GDDR3 though.

I imagine yields with TSVs would be downright awful for that in the next couple years too, and the power dissipation would certainly be an issue. Whatever Intel chooses, they'll likely only be looking at relatively low power/high density DDR3 - mobile/tablet/HTPC.

[Brad Grenz]

Frankly, it's a bit weird that it's taken so long for the old paradigm of these clunky memory sticks you plug in to a bunch of slots on a motherboard to go away. It would also give Intel a better way to differentiate their product line (IE, by RAM amount) than the arbitrarily disabling CPU features.

[Shifty Geezer]

Originally Posted by Brad Grenz

Frankly, it's a bit weird that it's taken so long for the old paradigm of these clunky memory sticks you plug in to a bunch of slots on a motherboard to go away.

How else do you provide upgradeable RAM? It's a valid feature in a PC. If PC had been that way for years with fixed RAM, and then someone came along offering upgradeable RAM just by plugging in a package, people would hail that as a great improvement. RAM on chip for faster BW and lower latency makes sense, but wanting to get rid of the RAM slots just because it's outdated doesn't.

Regards consoles, they've have fixed, soldered RAM, and this news doesn't really seem to affect what we know. The SemiAccurate article doesn't explain what exactly Intel are wanting to do, and RAM stacked on die is unfeasible at this point. Next gen is going to be stuck with DRAM chips soldered on the PCB somewhere

[tunafish]

Originally Posted by Brad Grenz

Frankly, it's a bit weird that it's taken so long for the old paradigm of these clunky memory sticks you plug in to a bunch of slots on a motherboard to go away.

It's a product of the technologies used. A lot of people seem to think that RAM is slow because it's far away. This is incorrect. RAM is slow because sense amping a DRAM cell is slow. It's far away because when it's so slow to start with, tagging on +20% transport latency does not hurt you enough to make the advantages of upgradeability and commodizability of the DIMM modules to go away.

So, DIMM modules as they are will go away when we move from DRAM to a better kind of memory, with less access latency. There are a lot of new kinds of memory tech "waiting in the wings", ready to take over if they ever get close enough to plain DRAM in cost per bit. However, DRAM manufacturers have managed to cost-reduce fast enough that it hasn't happened yet.

Personally, I think that we are actually pretty close to revolution on this front. The new memory techs still aren't cheaper than (or all that close to) DRAM per bit, but as the need for more memory hasn't kept up with the advances in production, it doesn't matter as much as it used to. RAM used to be a fifth of the price of a computer, now you can get 8GB for <$40, and that's more than good enough for most real uses. So even if the price can't keep up, what if in 2-3 years you can get 4-8GB of really fast ram for ~$40 instead of getting 16-32GB for ~$20? Assuming the rest of the system is updated to make use of it, that would probably help real, user-noticeable speed more than all the advances in CPUs in the past 5 years or so...

( If someone is interested, the layman explanation of why DRAM is slow:

Generally, the biggest reason that circuit elements get faster with shrinks is that as geometries get smaller, there's less distance to cover, there's less conductor to load and in other ways electricity just has to do less for the same effect.

DRAM works by having a really small capacitor that you stuff electrons into to wait until you read from them. On a read, you release these electrons to your sense amps, amplify the signal to something useful, pass it along, and re-store it in the cap.

The problem is that as you move to smaller geometries, the amount of electrons you can store in the capacitor goes down, to the point that reliably sensing the signal from them is really very hard. You need to have long chains of ever more powerful amplifiers to turn those initial few hundreds (soon few tens!) of electrons into a real current that can be sent across a wire, or even passed to the nearest normal transistor. )

[tunafish]

Originally Posted by Shifty Geezer

Regards consoles, they've have fixed, soldered RAM, and this news doesn't really seem to affect what we know. The SemiAccurate article doesn't explain what exactly Intel are wanting to do, and RAM stacked on die is unfeasible at this point. Next gen is going to be stuck with DRAM chips soldered on the PCB somewhere

I think that if a manufacturer would be willing to delay to late 2014, silicon interposers might be just on the border of the realm of the possible. It would be a really risky play, but it might pay back big if they otherwise would be restricted to a narrow memory bus.

Another possibility is that the promise of stacked ram in the future might compel them to accept much wider busses that would doom their chips to not shrink unless they can stack the ram. As I've heard it, the reason that the DDR4 spec is so badly delayed is that one of the big players wants to have a standard for a stacked interface with similar performance characteristics as the traditional one. AMD, Microsoft and Sony are all members...

[Shifty Geezer]

Originally Posted by tunafish

I think that if a manufacturer would be willing to delay to late 2014, silicon interposers might be just on the border of the realm of the possible. It would be a really risky play, but it might pay back big if they otherwise would be restricted to a narrow memory bus.

Would the performance benefits be worth anything in a market of cross-platform games where the rival has a much larger install base? If waiting would allow a price advantage, or a massive performance advantage that made a tangible difference to gamers and so be the platform of choice, then it'd make sense. Otherwise it's more important to get a decent product out there and selling, and compete on other more important fronts.

[tunafish]

Originally Posted by Shifty Geezer

Would the performance benefits be worth anything in a market of cross-platform games where the rival has a much larger install base? If waiting would allow a price advantage, or a massive performance advantage that made a tangible difference to gamers and so be the platform of choice, then it'd make sense. Otherwise it's more important to get a decent product out there and selling, and compete on other more important fronts.

More bandwidth is imho the single most effective way to differentiate in such a situation. It's very hard to make a game that utilizes additional cpu resources or the like on another platform and still runs on another, but more bandwidth allows for things like less compressed textures, better filtering or better aa, cheap things (for the developers) that gives noticeable IQ improvements.

As for cost, I'd expect the memory itself to be cheaper, but the bonding process (and the yield loss caused by it) is essentially a mystery at this point.

Of course, whether those improvements are actually worth the wait is a whole another question. Also, as I mentioned, it adds considerable risk. Truly new things in this industry have a tendency of being fucked up a few times before they stick. If it's not out by then, you've now got a crappy console and/or you're years late to the party.

So I'm not really expecting it will happen. It would be an extremely risky move to the point of gambling. I don't think large companies like that kinds of odds. I was just commenting on you using the word unfeasible -- I think it's borderline. Probably doesn't make any sense, but not necessarily unfeasible. Or put in other words, I think Cell was more crazy.

[joker454]

Originally Posted by Shifty Geezer

How else do you provide upgradeable RAM?

How often to people upgrade ram in their pc's nowadays anyways? Seems like most people now buy what they need, usually either 4gb or 8gb, and that's it.

[tongue_of_colicab]

I used to work part time at a pc store until the end of last year and there is still a decent amount of people buying ram. Compared to new systems it's not that many but the same goes for gpu's and especially cpu's.

I'm not opposed to the idea of including ram on the cpu just like including a decent gpu on the cpu isn't a bad idea at all but I wouldn't want to get rid of expansion slots. Mainly because I have no doubts intel will do everything in it's power to screw everybody over. Same as with phones. 100 euro difference between 16gb and 32gb models while the actual costs or only a couple of euro's.

[AlStrong]

Originally Posted by joker454

How often to people upgrade ram in their pc's nowadays anyways? Seems like most people now buy what they need, usually either 4gb or 8gb, and that's it.

Originally Posted by tongue_of_colicab

I used to work part time at a pc store until the end of last year and there is still a decent amount of people buying ram. Compared to new systems it's not that many but the same goes for gpu's and especially cpu's.

hm... Would be curious to know about RAM failure too. I mean, what do you do if the on-package/die RAM goes bad? Is it mostly a problem because people are unable to handle them properly? Regardless, the quality control* would have to be stepped up a notch I would think. There can still be bad RAM off-the-shelf and perhaps other external (electrical) issues.

*the integrated DRAM would probably be designed with a lot more redundancy for example - probably not much of an issue.

[Shifty Geezer]

Originally Posted by joker454

How often to people upgrade ram in their pc's nowadays anyways? Seems like most people now buy what they need, usually either 4gb or 8gb, and that's it.

Enough to keep Kingston and Crucial in business. It may be way less common now, but up until a few years ago at least it was still a common enough requirement. So unless Brad meant that the past two years should have seen the DIMM disappear, it's perfectly sensible that we still have them around.

[joker454]

Originally Posted by Shifty Geezer

Enough to keep Kingston and Crucial in business. It may be way less common now, but up until a few years ago at least it was still a common enough requirement. So unless Brad meant that the past two years should have seen the DIMM disappear, it's perfectly sensible that we still have them around.

Well you have to buy ram now to build new machines so yeah, not surprising that they are still around. For example I bought lots of ram last year but not to upgrade existing machines, it was to assemble new ones. But if the cpu itself had ram then I wouldn't have had to bother. I figure if most new hardware sales are laptops, tablets and phones then the need for upgradeable ram is falling to the wayside, leaving mostly system builders to buy it to make new machines. I'm fine getting rid of it really, system building has been getting simpler and simpler every year and this would be just one less thing to deal with. No more adding in an AST Six Pack Plus and a whole host of other cards. Right now its just add cpu, ram, and video card, with the latter two apparently on the endangered list.

[Shifty Geezer]

What if you want 16 GBs RAM for your PC and the CPUs only have 4 GBs inbuilt? Or in 5 years time, Intel is offering 16 GBs but you could do with 64 GBs because of 1080p60 3D video editing or whatever future techs would benefit? For limited function devices, certainly tablets and netbooks, in built RAM makes sense (not looking at performance advantages). But then these devices just solder it on the board anyhow. In a PC, if there's no performance benefit than there's no real reason not to have the flexibility of parts, so there's no reason to throw out the DIMM. Dell haven't started soldering their memory on the mobo. It's presumably more straightforward to have sockets and fit the required amount at build time.

[upnorthsox]

Originally Posted by Shifty Geezer

What if you want 16 GBs RAM for your PC and the CPUs only have 4 GBs inbuilt? Or in 5 years time, Intel is offering 16 GBs but you could do with 64 GBs because of 1080p60 3D video editing or whatever future techs would benefit? For limited function devices, certainly tablets and netbooks, in built RAM makes sense (not looking at performance advantages). But then these devices just solder it on the board anyhow. In a PC, if there's no performance benefit than there's no real reason not to have the flexibility of parts, so there's no reason to throw out the DIMM. Dell haven't started soldering their memory on the mobo. It's presumably more straightforward to have sockets and fit the required amount at build time.

then the competition sells a 16GB version and gets your money. I don't get the fear that you'll somehow be underpowered in your memory options, the markets will accomodate. The question is, would you rather have 4x the bandwidth, much lower power reqs, and a uma archetecture with the gpu or neither but expandable capacity?

Btw, with DDR4 you no longer are able to string more dimms on the channels. You'll have to replace to upgrade.

[Brad Grenz]

Yeah, that was my point. Memory performance is one of the biggest things holding back CPUs and it doesn't seem like upgradability alone is a reason to cling to the old paradigm instead of much higher performance and reliability. The rise of SSDs have also deemphasized the need for gobs of RAM as well.

[MrFox]

Sony made the Vita processor with a wideIO memory stacked on the chip. The Vita started production BEFORE the wideIO standard was even finished. That means they had no problem ordering custom memory from Samsung. So to a certain extent it means customized memory packages might not be such a crazy idea for consoles either. Whether early forms of HMC or JEDEC WideIO-2.

http://www.i-micronews.com/news/Sony...Vita,9334.html

The IO layout is not the same as the JEDEC wide I/O standard issued earlier this year probably because this program was underway before the std was issued.

It's pretty clear to me that they'll use the proposed JEDEC 1066MHz wideIO-2 standard on the PS4, before anybody else, whether the standard changes or not.
It'll probably be used for PCs and servers (and I assume GFX cards too) so the long term cost reduction of wideIO-2 chips are a given.

http://www.xbitlabs.com/news/memory/...d_Servers.html

While neither the HMC nor the wide I/O technologies are free of caveats, the trend towards multi-stacked memory devices with wide memory busses is generally clear. Such devices - given that they cannot provide truly high memory capacities - will likely be used in various notebooks, ultrabooks, micro-servers, storage servers and other devices of the kind that benefit from high-speeds, but do not need extreme memory capacities.

[MrFox]

I'm not sure it's possible, but yield issues could be solved the hard way by using one more wideIO chip than necessary, a nice crossbar switch in the interposer, and some fuses available to disable one of the chip. Yield would go way up.

Power issues don't look too bad, there's only one power layer, it'd be the side that is soldered on the heat spreader.

[Shifty Geezer]

Originally Posted by upnorthsox

then the competition sells a 16GB version and gets your money. I don't get the fear that.

What fear??? Brad didn't understand why we still have DIMMs. Answer is they are useful and haven't yet been replaced (eg. we could have had soldered on memory). Going forwards, if there's a better tech that provides the performance and memory, then great. That depends on what's possible though. I don't know the die size of DRAM and whether large enough quantities will fit on a CPU die.

No-noe's afraid of that. No-one's expressed any sort of emotion, or broken into cold sweats. As long as people have a desire for a certain quantity of RAM, if inbuilt RAM can't provide that, then they'll want another option. If the performance advantages are there (as I've already expressed), then the size of the RAM will probably become less important.

[Shifty Geezer]

Originally Posted by Brad Grenz

Yeah, that was my point. Memory performance is one of the biggest things holding back CPUs and it doesn't seem like upgradability alone is a reason to cling to the old paradigm instead of much higher performance and reliability. The rise of SSDs have also deemphasized the need for gobs of RAM as well.

I thought that, but when I asked in the PC forum about building a video editing suite, an SSD wasn't a good substitute for great gobs of RAM as I believed.

This is all way OT though. What the PC does isn't going to impact the consoles much. They can already go with whatever memory system they choose, and AFAIK no console has ever used DIMMs. RAM has been surface mount, included on a chip package, and layered even in the case of Vita. Intel's plans are immaterial to MS and Sony unless they actually have a tech on offer for next-gen that they may choose, but I expect Intel's prices to be too high to be worth considering.

[Laa-Yosh]

Originally Posted by joker454

How often to people upgrade ram in their pc's nowadays anyways? Seems like most people now buy what they need, usually either 4gb or 8gb, and that's it.

My work system has been upgraded at least two times in the past year.

[Laa-Yosh]

Content creation will need more and more memory, and no SSDs can help that. There are just a few apps that haven't made the switch to 64-bit by now (Zbrush is one of them though) and next-gen games will quickly eat up 8-16GB, not to mention movie VFX and such. Not that familiar with video editing but I'd say it can benefit from more RAM as well...

[sebbbi]

I do not see Crystalwell (or any other CPU embedded memory technology) to be a threat to DIMM modules (large RAM memory on motherboard) in near future. SSDs were warmly welcomed because there was a gap in the memory hierarchy. The speed difference between RAM<->HDD was too high compared to difference from L1<->L2, L2<->L3 and L3<->RAM. Modern SSD caching solutions are pretty good, and offer an extra step in the memory hierarchy. However if you remove RAM from the hierarchy, and basically replace it with L4 cache (even 1 GB of embedded memory is not enough to replace RAM), there will be an ever larger gap to fill, even if we assume that all systems have SSD caches. L4<->SSD will just be a too large gap to handle efficiently.

A huge L4 cache would be a great thing to have, especially if it is larger than the memory access footprint of a single frame in a game. A game running on Ivy Bridge or Trinity APU can theoretically newer access more than 366 MB of memory per frame (22 GB/s shared bandwidth for GPU/CPU, 60 fps). And this figure assumes that no memory is touched multiple times and all cache lines are fully utilized (and there's no overdraw in rendering either). In practice the upper limit of accessible memory is under 200 MB per frame (more analysis here: http://forum.beyond3d.com/showthread.php?t=62108).

If we had a 512 MB L4 cache, we could assume that:
- All data we access multiple times during a frame costs no extra memory BW, and is low latency (coming from L4)
-> Overdraw (objects behind each other, particles, alpha layers, etc) costs no memory BW, depth buffering costs no memory BW. Other temporary buffers cost no memory BW, shadowmap sampling costs no memory BW (assuming the map is rendered earlier this frame or used last frame), shadowmap rendering (updating) cost no BW either (if the same map has been already used last frame). Latency in accessing this data is also reduced (so we have less texture cache stalls, etc).
- The current working set from last frame is always in the L4 cache, and never needs to be reloaded from memory.
-> We only need to access RAM for data that is new.
--> At 60 fps only a few percent of data changes from frame to frame (this is a requirement for smooth animation). Old data that is updated (object is rotated/moved or shadowmap contents refreshed, etc) is not considered new data, as it lies in the same memory region (and thus is already in L4 cache). The same is true for all temporary work buffers that are used every frame.

Do I need to tell that I would absolutely love a system like this . Unfortunately Haswell didn't bring it yet, and I don't believe we will see this anytime soon. As memory bandwidth of CPUs/GPUs keep rising all the time, the required L4 cache size to hold a full frame keeps rising as well. If technology keeps advancing as usual, in a few years we would likely need at least 2 GB of L4 to keep a full frame in cache at once. 2 GB L4 wouldn't be an overkill solution for systems that have 32+ GB of main memory. But the L4 wouldn't be enough alone (as all cache misses would need to be queried from SSD/HDD... that would be really bad indeed).