It was MemCon this week. Three keynotes, three parallel tracks each with four presentations. An exhibition. In one day you can soak up the state of the memory world by osmosis. Christine Young live-blogged the keynotes by Chris Rowen of Cadence's Tensilica IP Group, Pablo Temprano of Samsung (who gave us all a lesson in how to use TInder), and Craig Hampel of Rambus (who wasn't going anywhere near Tinder since it was his 30th wedding anniversary). You can find her archived blog feeds here . The memory market is healthy right now. DRAM peaked in 1995 at the height of the PC boom at $40B. The market then fell away and it took until 2010, 15 years later, before the total memory market (including flash, etc.) was back to $40B. But since then with SSD, smartphones and a still memory-hungry PC industry, the memory market this year will be $75B. The market is very complex with different technologies requiring a huge spectrum of IP, VIP and other tools. With memory dominating almost any SoC design, it is perhaps the most important area of all in IP. Pablo pointed out lots of areas where memory growth has exploded: at rock concerts in the 1980s and 1990s, people waving a lighter used no memory at all and now people wave their smartphones with the screen in flashlight mode, consuming gigabytes of both RAM and flash. More seriously, phones have gone from 1GB of DRAM and 1GB of flash, to 3GB of DRAM and 16GB of flash or even as much as 128GB. One big driver of memory throughout the hierarchy from cloud datacenters to smartphones and IoT is video. With streaming video in the home (Netflix, Apple TV, etc.) becoming common and YouTube on smartphones, it is estimated that 80-90% of Internet traffic will be video. Chris stated it even more bluntly: to a first approximation, video/imaging is the only thing that matters. The biggest trend that many people talked about was the split between memory and storage. In this definition, "memory" is on-chip SRAM, and on-chip (or at least in-package) and off-chip DRAM (DDR). Storage is flash, SDD (flash-based "disk" drives) and HDD (actual rotating media disk drives). Memory is basically managed by the hardware of the system; the software has no control over the movement of data in and out of the caches. But storage is under software control. Memory is volatile and storage is non-volatile, but in fact even if flash is added to the memory hierarchy it is still effectively volatile since it holds too much stale data and the live data is in the caches and the write-back buffer and is lost if the power is lost. There is work going on (driven by Intel) to add true support for non-volatile memory in this gap, presumably because Intel wants to fill it, not with flash, but with 3D-Xpoint. The trend is for memory and storage to gradually merge. Several speakers pointed out the wall between the two, but the gap is closing. One reason is that if you add flash to the storage hierarchy then you don't need so much DRAM, and flash is 16 times cheaper than DRAM. Apparently in server farms it is well-known that you "don't need so much memory if you add an SSD." Intel/Micron's 3D-Xpoint memory looks like it will slot in here, too, but there isn't really enough information out there yet (power, price) to know. Plus nobody knows how it will yield, and Intel and Micron are not talking; Samsung is rumored to get only 50% yield on their latest vertical flash. One universal rule of memory technology is that memories only take off for the general market once they get cheaper than DRAM. Flash was used for camera and music storage, but only really became as big as it is today once it got cheaper than DRAM (it certainly helped that the smartphone industry took off, but maybe it wouldn't have done so if flash remained too expensive). Jack Handy had a nice characterization of flash memory. If you look at it the wrong way, then it is really expensive storage (about 10 times the price of hard disk). But if you look at it the right way, then it is really cheap DRAM, but with horrible characteristics. If you can find the way around the horrible characteristics (such as wear leveling) then it slots into the hierarchy of price/performance as in the "plaid" (actually in Scotland we call it tartan) diagram above, covering the gap in the middle. Since you then don't need so much expensive DRAM (and, in the limit, as many servers and software licenses), then this is very economically attractive. Jack's analysis was that most applications only needed 1-2GB of DRAM if they had a reasonable amount of flash. But the memory needs to be organized by software so that most of the writes go to the DRAM so that the flash will last for a long time without wearing out. So to me, the big "watch this space" trend is the closing of the gap in the memory hierarchy and the Berlin wall between memory and storage coming down.![]()
![]()
↧