replacing DRAM - Intel is roughing Samsung's and SK-Hynix' feathers with eDRAM

[spartaman64]

Oh, i agree with you. I call BS on that 16gb, but according to their block diagram that we've seen previously, i doubt it will be able to be used as unified system memory. If it could, it would be sweet, but there are just so much going on in the x86 desktop environment, and the fine control over timings would be non-existent on these different memory solutions.

 

No matter what, the enthusiasts will always want normal memory, just for the sake of being able to tinker with it. There might come a day where these embedded memory solutions replace physical DIMM modules, but that day is not now, or in the near future.

its not just for gpu both the cpu and gpu can use the hbm memory and that is probably a server processor

[spartaman64]

the question is how expensive is it because if they charge too much it will have the opposite effect and damage intel cpu sales and i would still like to be able to upgrade the memory

[MageTank]

its not just for gpu both the cpu and gpu can use the hbm memory and that is probably a server processor

Can you point me to where it says the CPU can use the HBM memory like its normal DDR3/4 memory? Not that i am saying you are wrong, i just would like to see that. I am still waiting to see if Zen is going to be a compelling option over skylake, and the more i know, the easier it will be to make a decision.

[spartaman64]

Can you point me to where it says the CPU can use the HBM memory like its normal DDR3/4 memory? Not that i am saying you are wrong, i just would like to see that. I am still waiting to see if Zen is going to be a compelling option over skylake, and the more i know, the easier it will be to make a decision.

well amd has been pushing HSA for a while and the problem is that the ram speed is too slow

[Deletive]

AMD: Intel, your IGPs are shit!

Intel: O'really?!

Hynix: what's that? HMC? nop! .. HBM

Intel: O'really!?

nVidia: Larabee? nop! can't have that

Intel: you're next!

 

Really? Because their graph about Zen completely disagrees with you. 

 

 

Notice the 4 DDR4 controllers on it. They are even claiming 16GB of HBM memory on the die (while i doubt it, i could still be wrong). They are not planning on doing away with normal memory.

 

This is mainly for the server market talking about the Server market, we are pennies to that market. When you're fitting 384gb-2tb of ram in a single system for Virtual machines/ ram disks and what not. Now if a server rack with blade servers,(like a few m1000e enclorues) each needing less ECC/Registered ram there is less demand for memory.

But getting rid of Ram in general at this point is not at all practical.  

[Touché! Good Sir]

Damn! Never mess with Intel I guess....

[patrickjp93]

d®amm it

That's not really correct. Intel confirmed desktop GT4e SKUs last year, but it may be Christmas or later before they arrive.

[patrickjp93]

Damn! Never mess with Intel I guess....

I'm surprised people still need to come to terms with that. IBM learned that the hard way falling from grace as the computing king back in the day, and Intel started with practically no capital but put MOS out of business, Texas Instruments into niche land with calculators and weather instruments, and IBM out of servers and personal computers, to remain in mainframes for a decade before making an even half-decent comeback.

[patrickjp93]

Can you point me to where it says the CPU can use the HBM memory like its normal DDR3/4 memory? Not that i am saying you are wrong, i just would like to see that. I am still waiting to see if Zen is going to be a compelling option over skylake, and the more i know, the easier it will be to make a decision.

The HBM will be a shared LLC between the CPU and iGPU. The CPU will never directly touch it. The L2/L3 cache will make the interactions transparent.

[patrickjp93]

This won't impact Samsung or SK Hynix in the slightest. eDRAM will do nothing to the current production of memory, and likely never will. Whatever they can fit on the die, you will always be able to fit more on a DIMM module. Capacity demands are still growing too.

Capacity demand isn't really growing in the consumer segment 8-16GB is the sweet spot unless you do a lot of excel work or CAD or modeling.

Also be aware of the number of chips on a DIMM and how we just got into chip stacking, something not in the DDR4 standard. If Intel gets 8GB on-die, that's plenty for a lot of people.

[MageTank]

Capacity demand isn't really growing in the consumer segment 8-16GB is the sweet spot unless you do a lot of excel work or CAD or modeling.

Also be aware of the number of chips on a DIMM and how we just got into chip stacking, something not in the DDR4 standard. If Intel gets 8GB on-die, that's plenty for a lot of people.

That would be quite the leap in such a short amount of time, considering they are just now at 128mb. While i don't doubt it will reach that point eventually, i can't see it happening anytime soon.

 

Also, welcome back. I sincerely missed you.

[Touché! Good Sir]

I'm surprised people still need to come to terms with that. IBM learned that the hard way falling from grace as the computing king back in the day, and Intel started with practically no capital but put MOS out of business, Texas Instruments into niche land with calculators and weather instruments, and IBM out of servers and personal computers, to remain in mainframes for a decade before making an even half-decent comeback.

I'm a newbie....... Sorry!

[patrickjp93]

you guys who are thinking about this in term of DDR tech are completly wrong. When this matures, and intel completly removes DDR controllers from the CPU, they will be in a 3D process era, and they will just stack the DRAM into however much you need, on top of the CPU. so there will be an X core CPU with a certain frequency and different amounts of ram on it. May be XPoint 2/3 by then

 

Hynix is screwed. Samsung will need to try and get an x86 licence to keep their chip business competitive. Mark my words

Nah, Samsung joined the HMC consortium, so between that and HBM, it can work alongside Intel by producing interposers or TSVs or both. I don't see that as being an issue. That said, Hynix is fried before 2030 easy.

[patrickjp93]

That would be quite the leap in such a short amount of time, considering they are just now at 128mb. While i don't doubt it will reach that point eventually, i can't see it happening anytime soon.

 

Also, welcome back. I sincerely missed you.

1 flat chip which is itself a 4-pipe memory system, made on 22nm FF. When it moves to 14nmFF as we'll see with Skylake and we start stacking, and widening the die, it'll grow pretty quickly.

 

Further, I'm not really back. I'm on for maybe 20 minutes a day just to see what the news is. Class work (reading, homework, presentations at the senior and graduate levels), setting up interviews for internships/jobs, managing the video game design club, attending choir rehearsal 8 hours a week across 2 choirs with 34 pieces, prepping for the career fair in 2 weeks, working on my senior capstone project (building software for Keybank), and then trying to eat 3 squares a day takes a lot of time out of my week.