Twenty times faster RAM with new Micron technology

[peja81]

Micron has presented its latest progress in the development of RAM products and its Hybrid Memory Cube technology promises up to 20 faster memory modules than current DDR3 products. The HMC technology was presented at Hot Chips last week and Micron showed a sample circuit with capacity to reach speeds up to 128 gigabyte per second.

Micron's HMC modules have already reached speeds 10 times higher than current DDR3 memory modules where DDR3-1600 is capable of data speeds up to 12.8 gigabyte per second.

When Micron is ready with retail products is another story, but the company claims to be able to offer 20 times higher bandwidth as DDR3 and do it with only 10% of the energy needed to power modern modules.

Hybrid Memory Cube stacks and binds together several layers of memory chips in a three dimensional memory circuit. To make the data paths as efficient as possible between the stacked chips it uses TSV connections (Through-silicon via) that lets signals pass right through the silicon. Through its 3D construction Micron claims that the technology will require 90% less space than traditional RDIMM modules.



The high number of TSV connections and the relatively short distances will be the key to the high data speeds, which could remedy the memory bottle necks we have today. Clear examples of this is the new Fusion APUs from AMD where several integrated CPU cores and powerful graphics circuits will have to share a very limited memory bandwidth.

Micron says that the HMC technology will break through the "memory wall" where it talks about how the developmentin this area most often result in just marginal improvements, unlike CPU and GPU markets where the coimputing power continues to improve at a much higher speed.

The question is when we will see the HMC technology in products, but there are not real hints of when this might happen.

http://www.nordichardware.com/news/75-memory/44000-twenty-times-faster-ram-with-new-micron-technology.html

 

[Beagle]

Iskoristiti takav bandwith u kucnim uslovima je tesko sa trenutnim stanjem tehnologije, a da u tu pricu ne ubacujemo uzasnu sporost "non voliatile" memorije kao sto su hard diskovi i SSD (u poredjenju sa ovom memorijom). Od ovoga ce najvise profitirati superracunari i graficke kartice (jer bi se dobio izuzento veliki bandwith sa jednostavnijim memorijskim kontrolerima).

 

[Ace Rimmer]

Pa sad, AMD-u već sada ima neke koristi od brže memorije na desktop APU-ima. Za pretpostaviti je da će to biti izraženije u budućnosti kako bude rasla grafička snaga.

 

[drfedja]

Bice izrazeno i na serverima, u HPC primeni i svuda gde je bandwidth kritican. Da ne pricam o tome sto ce energeetska efikasnost biti veca. Postavlja se pitanje latencije ovakve memorije, koja je druga strana medalje. Bandwidth je znacajan za paralelnu obradu podataka gde nema mnogo petljanja sa grananjem i slucajnih memorijskih skokova, ali ukoliko je latencija velika, to bise lose odrazilo na razlicite vrste izvrsavanja.
Opet, GPU koristi uglavnom dobro paralelizovane podatke, tako da latencija nije kriticna.

 

[Beagle]

Da, latency i mene zanima, ali reko what-the-hell

Kontam da, kad bi latency bio dovoljno nizak, proizvodjaci procesora mogli da zakinu na kesu, a da visak tranzistora uloze u asocijativnost memorije.

 

[drfedja]

Obicno bandwidth ide na ustrb latencije, pa se to posle ispegla sa visim frekvencijama. Ipak, ne znam na kom principu radi ova memorija, pa ne mogu da kazem je li veca latencija ili nije. Obicno sva ta "wow otkrica" na kraju zavrse negde u nekoj "fioci" ili to neko otkupi, kao npr. Z-RAM, koji je kao navodno trebalo da zameni klasican S-RAM jer je duplo vece gustine, pa je moguce praviti veci cache na procesoru. "Mali" problem je sto je latencija Z-RAM-a veca nekoliko puta od S-RAMa, pa nije upotrebljivo za cache, barem ne za L1 ili L2.