Next gen RAM choices *spawn

[aaronspink]

So all the people crying for 4GB in PS4 are nuts?

No not really. Part of the issue is that with GDDR5 you are looking at on the order of 2-3 GB of bandwidth per frame and if you only have 2GB of memory... In many games you won't be able to have enough assets to actually use the bandwidth in a useful way.

Wouldn't the 8 GB system be at an advantage over the 2B that now has to go to an HDD or something magnitudes slower to get new data? In essence the 8GB system could "buffer" 3 additional frames, vs 0 for the 2GB. Or am I totally not getting it?

That is pretty much it from my perspective. With marginally higher bandwidth (~50%) but significantly lower capacity (remember the cost trade-off is in the domain of 4x and will increase over time), you become effectively asset limited. You cannot stream in assets at a high enough rate to use them to their potential within a given frame. 4x the memory allows you to have a larger variety of assets loaded at any given time even if you don't use them all on a given frame. This results in things like much nicer overall texturing because you can draw from a much larger pool for any given object.

 

[aaronspink]

Can you please provide a link to this 100GB DDR3 memory? Thanks.

g-spec DDR3 or DDR4 memory can reach upwards of 3+ Gb/s.

 

[aaronspink]

No, he's saying that people thinking 8GB at 1/4 the bandwidth is somehow better are crazy.

But it isn't 8GB of memory at 1/4 the bandwidth. It is 8GB of memory at >50% of the bandwidth vs 2GB of memory at 100% of the bandwidth.

 

[itsmydamnation]

If in a stream like benchmark AMD is only getting ~50% memory bandwidth then they either have a bottleneck at some point between the CPU and the memory controller or have a sub-optimal memory controller.

you actually see this a lot with the bulldozer based mem controller. it has additional scaling with modules but each modules can never get anywhere near max memory throughput. if you compare the max memory throughput of a 2 module bulldozer to a 4 module bulldozer the 4 module can read and write far closer (aggregate) to the max theoretical numbers.

So i guess there is an internal bottle neck there somewhere but it isn't so obvious as performance still scales with memory speed, im guessing outside of synthetic benchmarks your unlikely to need more bandwidth per module?

maybe it has something to do with the latency difference between AMD and intel for direct memory read/write. AMD's controller had a larger focus on throughput and intels high focus on serial performance?

edit: you have to remember there is a big L2 and L3 to try and absorb that latency.

 

[upnorthsox]

g-spec DDR3 or DDR4 memory can reach upwards of 3+ Gb/s.

Do you have a link to this g-spec DDR3? Who has it in production?

DDR4 isn't even in production yet and when it is it may be a non-starter with HMC coming on around the same time. Btw, what density are you talking about? Even an optimistic 4GB density would 32 chips for 8 GB. That's an awful lot of chips to be soldering onto a console board.

 

[AlNom]

Even an optimistic 4Gb density would 32 chips for 8 GB.

16 chips

 

[aaronspink]

Do you have a link to this g-spec DDR3? Who has it in production?

If you want to learn about g-spec DDR3 I would suggest you look it up.

DDR4 isn't even in production yet and when it is it may be a non-starter with HMC coming on around the same time. Btw, what density are you talking about? Even an optimistic 4GB density would 32 chips for 8 GB. That's an awful lot of chips to be soldering onto a console board.

DDR4 is sampling now. HMC isn't a capacity solution but a bandwidth solution. HMC isn't even sampling currently. It is a lab technology without even a specification. DDR4 has a spec and is sampling from multiple vendors to multiple companies and will revenue ship at the end of this year or the beginning of 2013.

It takes 16 4Gb drams to hit 8GB of memory. And it isn't a lot of chips. You can fit that on a single sided DIMM.

 

[AlNom]

If you want to learn about g-spec DDR3 I would suggest you look it up.

Ah... see, the problem is that if you did put g-spec DDR3 into google, you'd come back to your posts here or GDDR3. Could you try to be helpful here

It takes 16 4Gb drams to hit 8GB of memory. And it isn't a lot of chips. You can fit that on a single sided DIMM.

Well, you don't know that they'd use DIMMs in a console...

 

[Squilliam]

Apparently on their site HMC technology specification ought to be finalised this year.

 

[upnorthsox]

But it isn't 8GB of memory at 1/4 the bandwidth. It is 8GB of memory at >50% of the bandwidth vs 2GB of memory at 100% of the bandwidth.

No, the rumor was 8GB of DDR3 vs 2GB of GDDR5 and considering 4GB if densities increase. Not liking that matchup you came up with the mysterious g-spec that no one lists as being in production and somehow gives a 4x bump to highend DDR3 that's currently around 22GB bandwidth or a replacement DDR4 that's not in production.

If we want to compare rumored non-production possibilities then let's discuss 8GB of DDR4 vs 4GB of higher density GDDR5. In that case, I'll take 4GB of GDDR5 with the 2x bandwidth advantage. I am though willing to listen to what you would do with the extra 4GB of memory even if it's at 1/2 the bandwidth.

 

[function]

Wouldn't a 256-bit bus require more layers on the motherboard (much bigger than a GPU mobo too) and provide a hard limit on shrinking the processor over time? If you locked yourself out of future process shrinks that might be a mistake - they were what allowed the 360 to save its hide.

By the time of the Jasper Xbox 360 the GPU was down to 121 mm^2:

http://www.anandtech.com/show/2682/4

What's the smallest GPU with a 256-bit bus? On a console you might need additional high bandwidth off-chip buses for the CPU and Edram too.

 

[Rangers]

Do you have a link to this g-spec DDR3? Who has it in production?

Does that stop anybody from speculating incessantly about HMC/3D stacking/whatever? I'd say DDR4 is MUCH more close to reality than those things...

 

[Rangers]

Wouldn't a 256-bit bus require more layers on the motherboard (much bigger than a GPU mobo too) and provide a hard limit on shrinking the processor over time? If you locked yourself out of future process shrinks that might be a mistake - they were what allowed the 360 to save its hide.

By the time of the Jasper Xbox 360 the GPU was down to 121 mm^2:

http://www.anandtech.com/show/2682/4

What's the smallest GPU with a 256-bit bus? On a console you might need additional high bandwidth off-chip buses for the CPU and Edram too.

Yet a 256 bit bus is on Orbis. Why is it ok for one manufacturer but not another?

Also for the shrinking, if you go to a SOC after some time, you should have plenty of perimeter for 256 bit even after multiple shrinks. Anyways who even knows how many shrinks will be cost effective next gen. 1? 2? They're not exactly shrinking at a breakneck pace this time around.

 

[upnorthsox]

If you want to learn about g-spec DDR3 I would suggest you look it up.



DDR4 is sampling now. HMC isn't a capacity solution but a bandwidth solution. HMC isn't even sampling currently. It is a lab technology without even a specification. DDR4 has a spec and is sampling from multiple vendors to multiple companies and will revenue ship at the end of this year or the beginning of 2013.

It takes 16 4Gb drams to hit 8GB of memory. And it isn't a lot of chips. You can fit that on a single sided DIMM.

I did look it up when you first mentioned it and Google returned me to your post(s). I then gave it due diligence and looked up the major memory manufacturers and none have any mention of it either sampling or in production. So again, who's manufacturing it?

 

[upnorthsox]

Does that stop anybody from speculating incessantly about HMC/3D stacking/whatever? I'd say DDR4 is MUCH more close to reality than those things...

Fair enough, put I'm at least able to provide a link:

http://denalimemoryreport.com/2012/...onsortium-hmcc-first-spec-due-by-end-of-year/

And as I said above, I'm willing to listen to the arguments for 8GB DDR4 vs 4GB GDDR5 so start them up.

 

[Rangers]

As for 100GB/s DDR, DDR3 2800 gets you to ~90 GB/s.

It appears to be pretty expensive right now, but I imagine a company with the resources of a MS can get the ball rolling on mass production between Jun 2012 and June/July 2013. All hypothetical of course, I'd bet on early DDR4 at DDR3-2800 speeds instead.

 

[Rangers]

Fair enough, put I'm at least able to provide a link:

http://denalimemoryreport.com/2012/...onsortium-hmcc-first-spec-due-by-end-of-year/

And as I said above, I'm willing to listen to the arguments for 8GB DDR4 vs 4GB GDDR5 so start them up.

Except it's not 4GB, it's 2GB.

And I have a feeling it's going to stay 2GB. Whch gets back to Spink's cost point. And Alstrong's contention GDDR5 is ~4.5X more expensive than DDR3 based on a costs sheet. Meaning 8GB cost~2GB cost...not 4GB...

Fair enough, put I'm at least able to provide a link:

http://en.wikipedia.org/wiki/DDR4_SDRAM ?

 

[upnorthsox]

Except it's not 4GB, it's 2GB.

And I have a feeling it's going to stay 2GB. Whch gets back to Spink's cost point. And Alstrong's contention GDDR5 is ~4.5X more expensive than DDR3 based on a costs sheet. Meaning 8GB cost~2GB cost...not 4GB...



http://en.wikipedia.org/wiki/DDR4_SDRAM ?

Where does that link discuss g-spec DDR3????

Y'all seem to flip between DDR3 and DDR4 to suit your point.

 

[Rangers]

Where does that link discuss g-spec DDR3????

Y'all seem to flip between DDR3 and DDR4 to suit your point.

I never mentioned "g-spec DDR3"

I have to admit, hoping for 100GB bandwidth on DDR3 seems a little wishful thinking. But it's certainly more realistic than HMC and the like imo.

As a floor, 68GB/s is possible (DDR-2133, very widespread/cheap today)

Since the leaked "doc" mentioned DDR4, that's why I keep falling back to it. If DDR4 is available that matches DDR3-2800 (which exists today), that would work.

I just think low speed DDR4 makes more sense than upper end DDR3. As others have said, DDR4 costs will fall throughout the gen, opposed to DDR3, which would make it worth a bit more expense initially.

 

[function]

Yet a 256 bit bus is on Orbis. Why is it ok for one manufacturer but not another?

Didn't know that about Orbis - where has it been confirmed btw? If you're starting with a large SOC then I guess it makes more sense. Also, if you're starting with a SOC you're probably not exactly bleeding edge so maybe the extra gains from GDDR 5 wouldn't be so valuable.

With the 360 there were several stages between the original components and integrating the CPU and GPU into one. 2005 -> 2010 would have been quite the no-mans land if they hadn't been able to shrink. Not to mention that even on a 256-bit bus DDR2 533 would have cut the main memory BW drastically and left the system with 8 memory chips forever.