Samsung & Hynix develop 8Gbit 3.2GHz LPDDR4

[pjbliverpool]

Looks like it's finally starting to arrive:

http://techreport.com/news/25830/samsung-develops-8gb-lpddr4-memory-chip
http://techreport.com/news/25838/sk-hynix-has-an-8gb-lpddr4-chip-too

Both chips seem to be running at the 'starter' speed of 3200Mhz delivering 51.2GB/s in a typical dual channel configuration.

I'm more interested on the apparent peak speed that should be arriving a little later of 4266Mhz delivering 64GB/s on a dual channel interface or a massive 128GB/s on a quad interface like those of the E series Intel CPU's.

I wonder how long it will be before Intel CPU's officially support this speed of memory. I'm guessing a while. The first DDR4 supporting Intel CPU sounds like it'll be either Broadwell or Haswell-E, both of which I expect to support only 3200Mhz. Then again, it would be in the interest of Broadwell to support the faster memory (if it's available by then) to support the IGP.

I'm also hoping for a doubling of the Crystalwell bus in either Broadwell of Skylake. That combined with dual channel DDR4266 would deliver a memory subsystem very similar to that of the Xbox One - which would be very cool for what is essentially a mainstream level CPU with an integrated GPU!

 

[Alexko]

Looks like it's finally starting to arrive:

http://techreport.com/news/25830/samsung-develops-8gb-lpddr4-memory-chip
http://techreport.com/news/25838/sk-hynix-has-an-8gb-lpddr4-chip-too

Both chips seem to be running at the 'starter' speed of 3200Mhz delivering 51.2GB/s in a typical dual channel configuration.

I'm more interested on the apparent peak speed that should be arriving a little later of 4266Mhz delivering 64GB/s on a dual channel interface or a massive 128GB/s on a quad interface like those of the E series Intel CPU's.

I wonder how long it will be before Intel CPU's officially support this speed of memory. I'm guessing a while. The first DDR4 supporting Intel CPU sounds like it'll be either Broadwell or Haswell-E, both of which I expect to support only 3200Mhz. Then again, it would be in the interest of Broadwell to support the faster memory (if it's available by then) to support the IGP.

I'm also hoping for a doubling of the Crystalwell bus in either Broadwell of Skylake. That combined with dual channel DDR4266 would deliver a memory subsystem very similar to that of the Xbox One - which would be very cool for what is essentially a mainstream level CPU with an integrated GPU!

I fear that might be overly optimistic.

 

[Wynix]

I think it would be better to make a new thread; This one is over 3 years old.

 

[pjbliverpool]

I fear that might be overly optimistic.

That's disappointingly slow (although granted 64GB/s is still a lot of bandwidth dedicated to a CPU), hopefully Broadwell will support faster given it's requirement to feed the igp and lack of a quad channel interface. Then again, can't you pair these processors up with faster memory anyway and take advantage of that speed? Or is the CPU limited to whatever speed it supports regardless of the actual memory speed like the old FSB?

 

[Wynix]

That's disappointingly slow (although granted 64GB/s is still a lot of bandwidth dedicated to a CPU), hopefully Broadwell will support faster given it's requirement to feed the igp and lack of a quad channel interface. Then again, can't you pair these processors up with faster memory anyway and take advantage of that speed? Or is the CPU limited to whatever speed it supports regardless of the actual memory speed like the old FSB?

I think that we will get to 4200Mhz very quickly; Though even that amount of bandwidth is not enough to feed future APU's.

 

[pjbliverpool]

I think that we will get to 4200Mhz very quickly; Though even that amount of bandwidth is not enough to feed future APU's.

True, but that speed memory plus Cystalwell should do extremely well, especially of they can double the bus width on Cystalwell, it seems pretty slow to considering it's an on die cache but I don't know the physical configuration to judge how likely a doubling of the width is.

 

[Grall]

It's not actually on-die, it's on the same substrate...but that's probably what you meant.

Does DDR4 support re-training the data bus to a different transfer speed, like GDDR5? That could be a nice power saving feature, especially in mobile platforms. Maybe it's deemed 'too advanced' for a mainstream type memory.

 

[Alexko]

That's disappointingly slow (although granted 64GB/s is still a lot of bandwidth dedicated to a CPU), hopefully Broadwell will support faster given it's requirement to feed the igp and lack of a quad channel interface. Then again, can't you pair these processors up with faster memory anyway and take advantage of that speed? Or is the CPU limited to whatever speed it supports regardless of the actual memory speed like the old FSB?

I don't think Haswell-E needs faster memory. As far as I'm aware, CPU workloads are very rarely memory-bound on Haswell, and Haswell-E has twice as many cores, twice as many channels, but probably lower core clock speeds, faster RAM (2133 vs. 1600 is still an improvement) and more than twice as much L3 (20MB vs 8MB). I doubt DDR4-3200 would have brought much of anything to a Haswell-E setup, and I suspect Intel would rather save power with "slow" memory. They can probably get away with 1.1V or less at these speeds.

Broadwell could indeed be a different story, but that depends on Intel's plans for Crystalwell.

 

[pjbliverpool]

Yep agreed on all points. Haswell-E is still going to sport the highest memory bandwidth of any consumer CPU so I was probably a bit premature to complain

 

[Grall]

The E-series has A LOT more I/O than consumer socket 11xx chips though, and the biggest ivybridge-E chip actually has 15 cores on-chip (with three fused off right now), so memory B/W makes more sense from that point of view. After all, enterprise customers (which is the target of the E series) are way heavier on the I/O compared to desktop users.

Maybe there'll be an incredibly expensive fully enabled version later on, even just the 12-core version has a $3500 premium as a build-to-order option from Apple (and from what I can tell, that's actually not a ripoff price.)

 

[Deleted member 13524]

Isn't LPDDR4 different than desktop/laptop DDR4?

What am I missing?

 

[pjbliverpool]

Low Power so I assume it's aimed at phone's/tables/laptops. But it presumably won't be long until it's followed up with desktop memory. 8Gb chips should also allow for 16GB on a single stick if I'm not mistaken so potentially looking at 64GB in a quad channel configuration with Haswell-E running at 64GB/s. Symmetry is nice.

 

[Deleted member 13524]

But isn't DDR4 supposed to appear much sooner than LPDDR4?
Google tells me the LPDDR4 spec isn't even finalized yet.

 

[Exophase]

Both chips seem to be running at the 'starter' speed of 3200Mhz delivering 51.2GB/s in a typical dual channel configuration.

The article doesn't say the channels are 64-bit. I count 308 pins on the Samsung package, vs 216 pins on a picture of a dual-channel LPDDR3 module. That increase in pin-count per channel (46) is similar to what you see in DDR4 DIMMs (48) and SODIMMs (52).

This is assuming either chip shot was truly representative.

I do wonder how long it'll be before we start seeing PoP options on big x86 cores.

 

[Lightman]

Is it with DDR4 that we will get 128bit DIMM's?
I remember this was planned for the future but haven't seen any mention of it lately. 128bit DIMM's would simplify a lot of things while saving us from OEM's selling PC's and laptops with only 64bit memory channel enabled.
Obvious downside is increased complexity of PCB's for both DIMM and motherboard but in the long term it shouldn't matter as we already have 256bit memory interfaces for Intel EX processors and AMD MCM Opterons.

 

[Grall]

I do wonder how long it'll be before we start seeing PoP options on big x86 cores.

While it would mean potentially seriously tiny computers, it would be hard to fit meaningfully large amounts - for a desktop system - of memory on a CPU package. Especially since you'd lose the ability for memory expansion this way. How much area would you need for 4-8GB of 128-bit DRAM? More than a standard CPU OLGA substrate can provide I'll wager.

 

[Alexko]

While it would mean potentially seriously tiny computers, it would be hard to fit meaningfully large amounts - for a desktop system - of memory on a CPU package. Especially since you'd lose the ability for memory expansion this way. How much area would you need for 4-8GB of 128-bit DRAM? More than a standard CPU OLGA substrate can provide I'll wager.

I don't know. I mean, if you can stack 16 8Gb chips on top of one another, you get 16GB of RAM, and that should still be more than thin enough for most notebooks, not to mention desktops.

 

[Grall]

Problem with stacking vertically is that you'll cook the package - thermal transfer through multiple dies would be super low, not manageable even in ULV mobile CPUs putting out >10W. You'd crash or kill the CPU this way. The dies would have to be spread out to avoid thermal death in any serious computing platform.

Even in cell phone SoCs do we have thermal throttling these days and these chips are >1W typically.