The Bitstream Blog :: Posts

MemCon 08 Pt. 2

More from Session 1

OS Optimization: This year the first Solid State Drives or SSDs appeared in end user applications. For many years, DRAM–based SSDs have been in use in enterprise computing where the insane cost and Soviet-era form factor was outweighed by their almost zero latency, incredible access times and sustained data rates. The new generation of NAND–based SSDs are non–volatile, use less power, are orders of magnitude faster, on average, than rotating media and, are much more mechanically and electrically reliable than HDDs as well. As an example, RAM giant Samsung has an SDD available in either PATA or SATA attach versions, with capacities from 32 to 128 GB. SanDisk, STEC and Avant (see below) are other vendors providing drop–in replacements for SATA disks.

Current generation OSs are optimized for rotating media as their “tier zero” or primary, performance–oriented storage. This does not take into account the completely different power profile of a NAND SSD. Unlike DRAM SSDs, you don’t have to cache the contents when switching to “sleep” or low power mode. Also, rotating media requires a large power burst to initially spin up the drive and relatively large continuous application of power to keep the device operational. Quiescing the drive means that another burst of current will be required to spin up the drive. In contrast, SSDs or Solid State Drives have relatively miniscule power requirements to operate the device and essentially no power to quiesce or sleep the device.

Also, RAM is read or written with high granularity but is erased, in contrast to HDDs, with very low granularity. So, an OS should deal with task prioritization and workload management that is currently not handled.

on to Session 2: Steven Woo, Rambus

Graphic subsystems are still the most demanding subsystem for memory applications, closely followed by game consoles. This is especially true as the number of compute cores in GPU now exceeds 800. GPU memory bandwidth now exceeds 128 GB/s, which is quite a bit higher than the memory bandwidth that CPUs or CE (consumer electronics)appliances require. This makes it tough for memory vendors to design one product for different application markets.

Rambus Terabyte Tech Initiative: From XDR at 3.2+ GHz to 1 TB/s…How? Among other things, multi–threaded, fully differential architecture, including data and command/address links achieve 16 Gbps speeds. Also, EQ; Equalization provides greatly reduced ISI or Inter–Symbol Interference. Eye patterns without EQ are almost non–existant while, with EQ, the eye pattern can be textbook–perfect.

Session 3: Panel

Amongst other topics, they mentioned “ZRAM,” “MRAM” and Numonyx’s “PRAM

Session 4: Don Barnetson, SanDisk

Currently, only top shelf laptops, like the Lenovo X300 and MacBook Air, employ an SSD as their primary storage. SanDisk predicts that SSDs will, by 2011, be part of the $900 to $1500 notebook market, while small capacity SSDs are already a part of the “netbook” or ultra–low cost, sub–notebook market. In Ye Olden Days, netbooks would have been considered “thin clients.”

He discussed the concept of LDE (Longterm Data Endurance), or XRZPD, a single number “gas gauge” measurement of SSD drive endurance that indicates the amount of data, say 20, 40 or 100 TB, that can be reliably written during a product’s lifetime.

Session 4: Amber Huffman, Intel

NAND is the only commodity memory type that has no standard interface. Huffman talked about the ONFi or Open NAND Flash Interface standard, which attempts to address this conundrum. They have put forth uniform electrical and protocol interfaces as well as a universal mechanism for device self–reporting so the host can “ask” what the device is capable of. They have also published a connector spec for standardized NAND in mobile products and DIMMs, in place of those flash DIMMs you have now. ONFi 2 is backwards compatible with trad DRAM and it’s also scalable, bringing the future a little closer while reducing time to market and development costs. ONFi is one example of how vendors improve the specs of their finished products while keeping the end user cost constant.

Session 5: Nathan Orb, Microsoft

He discussed NVMHCI, the Non-Volatile Memory Host Controller Interface, a proposed host controller method where, unlike the current interfaces for removable media and ATA, the host and device controller are merged…

Session 6: Paul Goodwin, Avant Technology

Goodwin discussed SSD features and form factors, mentioning the low power (< 1W vs 2-9 W), high reliability, ruggedness, wider operating range for temperature and altitude (yup, atmospheric pressure is spec’d for HDDs) 10x performance over HDDs. He also mentioned that CE applications are a perfect application for SSDs since they are very quiet, an important factor in the living room.

Avant has a brand new SSD for servers and desktops that uses RAID inside the package to provide capacities up to 640 GB, all in a 2.5″ form factor. This provide the standard RAID modes for increased fault tolerance (mirroring), increased performance (striping) or both since the device includes up to 5 “drives” inside…Holy mackerel! They even have a hybrid mode that provides a combination of separate RAID 0 and RAID 1 within the total device capacity. One portion provides high reliability and the other portion provides blazing performance. Trouble is, the 640 GB product has got to cost mucho $$. [I couldn’t get a price on Avant’s drive but, to provide some perspective, a top of the line, 146 GB “Zeus oips” SSD from STEC retails for US$21,505 while a SanDisk 80 GB UATA will cost you about US$6100. Did I mention SSDs ain’t cheap?]

Session 7: Panel

CFast was mentioned, a transitional standard for CompactFlash with SATA interface

Intel presented an interesting experiment whereby, attached to a Sun server, a pool of old school 7k RPM rotating disks are replaced with lower cost 4k RPM drives in conjunction with a SSD for frequent data requests and another SDD for high performance file system journaling cache. The price for both approaches are about equal while the hybrid approach offers a 2x improvement in capacity and 4x performance gains.

Vendor Booths:

In addition to memory vendors like Qimonda and Samsung, design tool vendors such as Cadence and test & measurement vendors like Agilent, the MultiMediaCard Association (MMCA) had some poop on two emerging standards, e–MMC as I mentioned earlier, and miCARD. e–MMC or Embedded MMC, is an industry standard jointly developed by JEDEC <> and the MMCA for embedded flash memory applications. e–MMC leverages the existing MMC standard for controllers and flash memory. It addresses the need to divorce individual vendor’s functional differences from the host, which simplifies design and results in shorter design cycles, briefer qualification times and quicker time–to–market. It operates at either of the current powering standard, 3.3 or 1.8 volts, and scales up to 52 MBps and will find a place in a broad range of CE market applications.

miCARD, the Multiple Interface Card, is another standard developed by the MMCA. It combines two common interfaces, USB 2 Type A like you’d find on your mouse or keyboard, and the MMC standard you’d find in CE products such as cell phones and PDAs. With the help of a very small “passive mechanical adapter,” the e–MMC card plus a sort of slipcase that fans out the contacts and provides the correct physical profile, the e–MMC can be plugged directly into a USB receptacle. miCARD supports up to 480 Mbps with full compatibility with USB 2.0 while also providing full MMC v4.1 compatibility. Trouble is, at 12 mm by 21 mm by 1.95 mm, I find the package size to be stupidly small for day to day use.

All in all, a very useful conference for me. the presentations were mostly well done, the food was decent and I learned quite a bit. Hope you gleaned something useful from this post…l8r!


One Response to “MemCon 08 Pt. 2”

  1. […] – MemCon08 Pt. 2 saved by thealmostmusic2008-09-15 – Samsung & Microsoft to tune up Vista for solid state […]