An Open Letter to semiconductor companies

[nn_step]

Ever since the IBM standard, PCs have used DRAM because it provided the greatest value per dollar for the temporary storage of values in RAM. This is because the advantage of DRAM is its structural simplicity: only one transistor and a capacitor are required per bit, compared to six transistors in SRAM. This allows DRAM to reach very high densities.

However this structural simplicity brings about its largest weakness, DRAM is volatile memory and because of leakage, requires the cells to have their values periodically refreshed and during which times, the cells can not be accessed. To minimize the effect, a staggered refresh rate is generally used. But each and every time a refresh happens, it consumes additional power and thus reduces the energy efficiency of the system.

Since modern systems now can cheaply have more memory than they currently require (8GB for $60), we should trade what we have excess of for what we want more of. Which in this case is capacity for speed and power efficiency.

Hence I propose that for the next RAM standard, that Intel, AMD, VIA, ARM, etc should move from DRAM to SRAM (Static random-access memory). Which will provide the following benefits at the following cost. SRAM is faster and significantly less power hungry (especially idle) than DRAM. It is therefore used where either bandwidth or low power, or both, are principal considerations. SRAM is also easier to control (interface to) and generally more truly random access than modern types of DRAM. Due to a more complex internal structure, SRAM is less dense than DRAM. Which means per a given dollar, gives you approximately 1/6 to 1/16 the number of bytes that DRAM does.

But assuming that bandwidth and power consumption are not reason enough for the switch, the best argument for SRAM has been and always be latency. SRAM wins every single time, that is why modern CPU cores carry moderately sized SRAM caches instead of possibly much larger DRAM caches. When latency trumps capacity, SRAM wins every time.

So what do you think? Am I wrong to suggest that we should simplify modern memory controllers, speed up RAM by a significant margin, and accept the minor additional cost of switching to SRAM?

To users of the forum: comments, questions, or suggestions?

[TV Addict#2]

should atleast be a good option for the hardware interested of us
I would buy

[Particle]

It would be reasonable if memory controllers could handle modules of either type, modules were physically and electrically compatible with sticks of DRAM, and end users could simply decide which to buy. Otherwise, I think it wouldn't be practical. Leading-edge sticks would be either 1GB or 512 MB, and that's just not enough RAM in 2-4 slots.

[zalbard]

How much of a performance benefit would we get for 6x-16x prices? I don't think it makes sense from price/performance PoV.

[zanzabar]

in terms of numerical performance its like good ddr3 to l3.

the main problem with sram though is that its much higher in transistor count than dram (like 4-8x as much). but the sticks would cost about the same just about 4x less space so u would run 1GB sticks now and 2GB next year based on the transistor count, but it would be awesome since u could store to it and it would make sleep mode and shutdown the same for power consumption, but holding the data for that long and not reinstalling the memory would lead to problems with stability (just look if u dont restart windows every week what happens)

[Ket]

A nice idea... in theory. The specs for DDR4 have been floating around for a while though and if I remember right DDR4 will be the first true 128bit module instead of these old musty 64bit modules which have been WAY overdue a bitdepth upgrade for years. Personally, I'm looking forward to true 128bit modules.

[nn_step]

It would be reasonable if memory controllers could handle modules of either type, modules were physically and electrically compatible with sticks of DRAM, and end users could simply decide which to buy. Otherwise, I think it wouldn't be practical. Leading-edge sticks would be either 1GB or 512 MB, and that's just not enough RAM in 2-4 slots.

Then what exactly in your perspective would be a more reasonable capacity per stick?

How much of a performance benefit would we get for 6x-16x prices? I don't think it makes sense from price/performance PoV.

Assuming that you expect byte for byte identical capacity, 75% or greater reduction in latency. The effect on performance depends on the task involved. For certain easy to cache applications there wouldn't be a benefit. For other hard to cache applications the benefit would exceed 200% increase in single thread performance.

Now remember, the argument is not 8+GB sticks of SRAM, rather 1-2GB sticks of SRAM.

[Serra]

My understanding is that the next notable iteration of RAM will not use the same materials that it does today, whether it be the memrisistor or another technology. I don't think it'll be here for DDR4 - that's already sailing - but DDR5 is a very possible candidate for it.

So yeah... fundamental changes to the materials and designs of the sticks is already underway...