SATA Express/NGFF making RAM less important/obsolete?

[etherealfocus]

So with DDR3-1600 supplying 12.8GB/s per channel and SATA-E scheduled to supply 16GB/s or 32GB with mPCIe, how much will we really need RAM moving forward?

-Going to dual channel gives a ~10% performance increase, implying that current CPUs need a bit more than 12.8GB for optimal performance, but nowhere close to the 25.6GB that dual channel provides. Say 16GB or so is ballpark for optimal performance - obviously that spikes with big IGPs.

-16GB SATA-E beats single channel 12.8GB DDR3 and 32GB mPCIe beats dual channel 25.6GB DDR3. Given the SFF push, I'd expect mPCIe to gather momentum quickly.

-Even if dumping RAM did introduce some bottlenecks particularly in random IOPS, the tradeoff would probably be worth it especially for cheap/low power applications, and not needing to load programs (or Windows) before running them would be huge perceptually. The Starcraft II single player campaign would probably take an hour less to play through without load times.

-SSD companies are already prioritizing performance consistency and random IOPS and are improving quickly. RAM, by contrast, is mostly stagnant. Yeah, DDR4 is coming, but nobody seems super-excited about it. And if history is any guide, it'll be mostly stagnant with only minor clock speed increases. In all the years we've had DDR3, we've only moved our average/peak officially supported speed up one level (1066 average/1333 peak to 1333/1600... or I guess 1866 if you count AMD chips). SSDs are clearly catching up fast.

-Everyone's dealing with tight margins and limited space with the SFF push, and removing RAM would help with both in a pretty significant way.

-There's no reason we couldn't use multiple mPCIe drives to satisfy hefty bandwidth requirements as needed.

-AMD's HSA would seem to gain massive benefits from being able to work directly from nonvolatile storage, maybe with a large L4 cache.

Thoughts? I'm obviously not an engineer, but the prospect is exciting.

 

[Hellhammer]

You're mixing bits with Bytes. SATA Express is up to 16Gbits per second (32Gbits for SF-8639), which is equal to 2GBytes per second. RAM is still an order of magnitude faster and what's even more important is that its latency is much smaller (measured in nanoseconds, whereas SSD latency is measured in microseconds). That's not going to change anytime soon since NAND is architecturally slower than DRAM and is only going to get slower as the lithography gets smaller.

 

[R0H1T]

So with DDR3-1600 supplying 12.8GB/s per channel and SATA-E scheduled to supply 16GB/s or 32GB with mPCIe, how much will we really need RAM moving forward?

NAND operating at such speeds is gonna be crazy expensive besides I think only certain raid configs on PCIe allow such speeds currently. Power & data corruption is also going to be a major issue as compared to DRAM !

-Going to dual channel gives a ~10% performance increase, implying that current CPUs need a bit more than 12.8GB for optimal performance, but nowhere close to the 25.6GB that dual channel provides. Say 16GB or so is ballpark for optimal performance - obviously that spikes with big IGPs.

-16GB SATA-E beats single channel 12.8GB DDR3 and 32GB mPCIe beats dual channel 25.6GB DDR3. Given the SFF push, I'd expect mPCIe to gather momentum quickly.

The competition is DDR4 not current DDR3 & it'll beat SATA-E hands down.

-Even if dumping RAM did introduce some bottlenecks particularly in random IOPS, the tradeoff would probably be worth it especially for cheap/low power applications, and not needing to load programs (or Windows) before running them would be huge perceptually. The Starcraft II single player campaign would probably take an hour less to play through without load times.

-SSD companies are already prioritizing performance consistency and random IOPS and are improving quickly. RAM, by contrast, is mostly stagnant. Yeah, DDR4 is coming, but nobody seems super-excited about it. And if history is any guide, it'll be mostly stagnant with only minor clock speed increases. In all the years we've had DDR3, we've only moved our average/peak officially supported speed up one level (1066 average/1333 peak to 1333/1600... or I guess 1866 if you count AMD chips). SSDs are clearly catching up fast.

By design programs meant to run on a PC or similar computing device only load relevant data & files into memory as & when needed, the rest are swapped in/out of the RAM using a pagefile. Changing this schematic is gonna require a major revamp of virtual memory structure under windows, however not totally sure if linux or any other OS will work this way or not.

-Everyone's dealing with tight margins and limited space with the SFF push, and removing RAM would help with both in a pretty significant way.

-There's no reason we couldn't use multiple mPCIe drives to satisfy hefty bandwidth requirements as needed.

Nope, the expensive NAND replacements are not gonna be viable financially !

-AMD's HSA would seem to gain massive benefits from being able to work directly from nonvolatile storage, maybe with a large L4 cache.

It'll benefit regardless of where the data runs, more so in case of DDR4.

Thoughts? I'm obviously not an engineer, but the prospect is exciting.

Technically I'm a CS engineer but since I'm not active in this field for quite sometime now so take the above points as being more opinionated than objective

edit : Seemingly missed the most important point i.e. speed & latency ~

You're mixing bits with Bytes. SATA Express is up to 16Gbits per second (32Gbits for SF-8639), which is equal to 2GBytes per second. RAM is still an order of magnitude faster and what's even more important is that its latency is much smaller (measured in nanoseconds, whereas SSD latency is measured in microseconds). That's not going to change anytime soon since NAND is architecturally slower than DRAM and is only going to get slower as the lithography gets smaller.

 

[etherealfocus]

Ahh yeah I had a derp there lol. Someone in another forum capitalized the GB/s for SATA-E and I brainfarted right past it.

Too bad... it's a cool idea.

Anyone know what the current status of NAND replacement like phase change memory is?

 

[Hellhammer]

Anyone know what the current status of NAND replacement like phase change memory is?

They are still years away from any kind of mass production. The next "big" thing will be 3D NAND coming in 2014-2015, but that's not going to have any major impact on latency as far as I know (details are still very scarce).

 

[avtar2008]

i won't think RAM is going out of pc components list any time soon, or may be never. Ssd won't be able to replace ram. Normal pc usage performance depends upon random write speed,and generally ssd have only 40/82 MBps of random 4k read/write speed. So, they are not even close to ram.

 

[exdeath]

Bring on the convergence of RAM and HDD to a true multi purpose high speed high density non volatile universal main memory.

We need to go back to the magnetic core memory paradigm where RAM and storage are one and the same and the memory technology that enables this will be a revolution in computing matching that of the transistor and microprocessor. 1:1 CPU:MEM zero wait state computing regardless where the data is stored. We can remove "hourglass" and "progress bar" and "please wait" from computing lexicon.