Intel CES 2014...what the?

[witeken]

Eh, I mean at 51:00 (till ~52:00).

 

[WaitingForNehalem]

Oh, they're showing off augmented reality. You can't see it with your eyes.

 

[norseamd]

Intel actually wanted us to leave x86 for IA64 at that time. Its a shame we didnt go the IA64 route on the long term. But thats a bill we have to pay later on.

i am interested in the other non x86 archetectures. it seems some of them could offer much better performance than x86. why would IA64 have been better at least in the long run. what about the 8 core 64 thread 5 ghz cpu that ibm makes. could you put that in a desktop.

how would xdr2 compare to ddr4 or ddr3

 

[ShintaiDK]

i am interested in the other non x86 archetectures. it seems some of them could offer much better performance than x86. why would IA64 have been better at least in the long run. what about the 8 core 64 thread 5 ghz cpu that ibm makes. could you put that in a desktop.

how would xdr2 compare to ddr4 or ddr3

IBM PowerPC is not a solution either.

IA64 moves the entire sechduling part to the compiler. And it gives you the ability for the same reason to make a very wide and smaller core with higher resource utilization.

There is a reason why there was a hype for AMD to use XDR2 with Tahiti.

Now we are stuck with DDR4(1 DIMM per channel only) and GDDR5 until stacked memory arrives in...2017+

 

[norseamd]

IBM PowerPC is not a solution either.

so how do power pc x86 and ia64 compare with each other

There is a reason why there was a hype for AMD to use XDR2 with Tahiti.

is this a great product ruined by a manufacturer that wants to use propietary standards and take over the whole market

 

[ShintaiDK]

so how do power pc x86 and ia64 compare with each other

Maybe more specific? Anything special you think on? Specially IA64 is the most different one to the other 2. Also the newest of the 3.

is this a great product ruined by a manufacturer that wants to use propietary standards and take over the whole market

It was ruined by the DRAM cartel yes. Meaning Samsung, Hynix, Micron, Infinion and Elpida. They got away with a billion$ in fines and some execs jailed. The consumer...just lost as always.

 

[norseamd]

Maybe more specific? Anything special you think on? Specially IA64 is the most different one to the other 2. Also the newest of the 3.

just overall architecture wise. know very little about them.

Meaning Samsung, Hynix, Micron, Infinion and Elpida.

meaning rambus. there were lawsuits. lots of lawsuits.

 

[ShintaiDK]

just overall architecture wise. know very little about them.

IA64 is EPIC design, PowerPC is RISC design, x86 is CISC/RISC design.

meaning rambus. there were lawsuits. lots of lawsuits.

RAMBUS won most of the lawsuits. But the damage was done. We moved to DDR and the slower tech for good. Specially GDDR5 gonna be a very long and painful experience. And DDR4 looks like the next horror story. Late 2015 for mainstream usage it seems. And only 4GB and 8GB DIMMs to start with.
http://techreport.com/news/25893/micron-provides-ddr4-details-for-servers-high-end-desktops

PS3 uses XDR btw.

 

[norseamd]

RAMBUS won most of the lawsuits.

but they sued the whole industry fairly recently.

And only 4GB and 8GB DIMMs to start with.

32 gb should be enough for at least several more years. 8 gb is all that is needed right now but i think that programs could use 12 gb or even 16gb in the next few years if they are designed to take advantage of it.

PS3 uses XDR btw.

read that. remember reading that before and wondering how much faster it was.

 

[norseamd]

what about these memory types

http://en.wikipedia.org/wiki/Magnetoresistive_random-access_memory

http://en.wikipedia.org/wiki/Millipede_memory

http://en.wikipedia.org/wiki/Racetrack_memory

 

[ShintaiDK]

32 gb should be enough for at least several more years.

Remember, 1 DIMM per channel unlike DDR3. So 16GB with 8GB modules would be the highest on the desktop/laptop for quite a while. If you exclude LGA2011 that could have twice as much on the desktop.

You need additional switch chips to enable more than 1 DIMM per channel. Something only servers will see. And maybe a few LGA2011 desktop boards.

 

[norseamd]

why not make quad channel. does it need a new cpu or socket.

 

[ShintaiDK]

why not make quad channel. does it need a new cpu or socket.

Bigger socket, more pins, more traces, more expensive board etc etc.

 

[TerryMathews]

why not make quad channel. does it need a new cpu or socket.

They could. They won't. Gotta keep those premium features segmented for extra $$.

 

[ShintaiDK]

They could. They won't. Gotta keep those premium features segmented for extra $$.

So why is AMD keeping it dualchannel? For the premium? Or to keep cost down?

 

[norseamd]

what about the other memory types

 

[ShintaiDK]

what about the other memory types

Either not suited, not finalized, not ready for mass production or simply too expensive.

 

[norseamd]

which could be ready in 5 to 10 years. would cube memory be better

 

[ShintaiDK]

which could be ready in 5 to 10 years. would cube memory be better

Stacked memory is already decided as the future. And in the future you will buy CPUs and dGPUs(if they still exist) with a prefit amount of memory you cant change.

 

[norseamd]

And in the future you will buy CPUs and dGPUs(if they still exist) with a prefit amount of memory you cant change.

why? because of architectural constrainsts or to be cheap control consumers buying habits and to make them buy a whole new computer when they want or need more ram

 

[ShintaiDK]

why? because of architectural constrainsts or to be cheap control consumers buying habits and to make them buy a whole new computer when they want or need more ram

Because thats how stacked memory like Cube and HBM works.

Example of a CPU package with HBM that AMD signed up for.

Its also a cost saving issue. Less components means cheaper and more reliable. And the bonus with HBM/Cube/etc is you remove the bandwidth limitation. Think main memory at 250-500GB/sec instead of 25GB/sec.

 

[norseamd]

does it allow for different memory types. this just refers to how the chip is assembled right. where does universal memory fit into this.

how about mram. looks like it is fairly well developed. it seems they just have to start producing it and do some more tweaking for several years and then it might work. fab capacity right. gareented products get the first bid on the newest fabs right.

 

[norseamd]

hmc is before fully stacked chips right

 

[ShintaiDK]

does it allow for different memory types. this just refers to how the chip is assembled right. where does universal memory fit into this.

how about mram. looks like it is fairly well developed. it seems they just have to start producing it and do some more tweaking for several years and then it might work. fab capacity right. gareented products get the first bid on the newest fabs right.

MRAM is fine, if you want very little memory. I think the biggest MRAM chip is still 8MBytes.

Cube and HBM is DDR.

 

[ShintaiDK]

hmc is before fully stacked chips right

HMC=Cube=Micron/Samsung.
HBM=Hynix.