Robson: Intel's one-two punch
Posted: 2006-03-14 04:43pm
Most of us have seen the Conroe benchmarks from IDF at this point and you'd have to be a fool not to be impressed. Intel has been at a performance disadvantage since the Prescott core, but now they are set to win back the performance crown and then some.
But it looks like they aren't stopping with the new Core architecture:
What is interesting about this though is that Intel looks to be readying this for Conroe, which would add to its already commanding performance leads, not to mention being the first serious attempt to drastically reduce load times.
But it looks like they aren't stopping with the new Core architecture:
The performance gains look impressive, and this is a much better solution than hybrid hard drives since you could share a single cache memory among several drives and it would be user defined storage.Extreme Tech wrote:Accessing a disk drive is hundreds of times slower than accessing main system memory. Flash memory is slower than the DRAM used for system memory, but it's still far speedier than pulling data from rotating magnetic media. If you've ever waited for a large game level to load, you'll know what we mean. There you sit, with the hard drive light flickering, staring at a progress bar on the screen. For this, you've paid $50?
But it's not just a matter of loading applications faster. One of the major sources of battery drain in a note interfaces have helped, as hard drives moved from ATA66 to ATA100, then to SATA 150 and today, SATA 300. Areal densities on the hard drive platters have also contributed to faster data transfers. Despite all this, overall drive performance hasn't kept pace with the tremendous advances in CPU capabilities. In reality, the impact of faster disk transfer rates has a marginal effect on performance for typical drive usage, because the majority of transfer sizes are quite small.
Grimsrud considers mechanical positioning latency—the lag associated with moving the drive head to specific tracks on the platter—the main determinant of overall drive performance. In that department, hard disks have only seen a 2.5% yearly improvement in the past decade, amounting to performance scaling of 1.3×. For example, the transition from 5400rpm to 7200rpm standard drives reduced the latency from 5.6ms to 4.2ms.. While we've seen a few 10,000RPM hard drives, it's unlikely we'll see dramatic increases in rotational speeds.
In the same time, normalized CPU performance has scaled by a factor of 30. So hard drives really haven't kept up.
Advances in mechanical latency reduction are no competition for Moore's Law. How does that affect system performance? It's simple: The CPU has to wait. When the CPU needs data, it first searches the CPU cache, and then looks to main memory. If it's not there, then the data needs to be loaded from the hard drive, which takes multiple seconds—an eternity in processor time.
To combat this performance disparity, Intel's approach is to insert another layer into the hierarchy to bridge the gap. They've codenamed this platform acceleration technology "Robson."
How Robson Works
Intel proposes that the flash memory cache should be located on the motherboard. It's unclear whether it should be permanently installed or could be another type of memory socket. A Robson nonvolatile cache would require a driver to be loaded, however. The Robson cache connects to the I/O controller via PCI Express. Part of the technology is an intelligent prefetcher. The prefetcher anticipates which disk blocks will be needed from the hard drive and stores them in NV memory ahead of time. This data can persist even across boots. When requested by the operating system, the data is accessed using the low latency of solid state memory rather than the much higher-latency hard drive.
For write commands, data is buffered in NV memory and later written to the hard disk to minimize unnecessary disk accesses. Cache configurations range from 128MB up to 4GB in size. And it's compatible—Robson works with any SATA drive across multiple operating systems, including Windows Vista.
Robson itself burns a scant 0.1W on average, a minor draw easily compensated by its net system power savings. In addition to application loading that's two to three times faster and quicker boot and resume, Robson prolongs battery life on laptops by reducing hard disk activity. Particularly beneficial in mobile applications, Robson-enhanced disk drives spend more time spun down—not only saving power but also rendering them more resistant to damage from mechanical shock.
Synthetic benchmarks are all well and good, but Intel also demonstrated how Robson might perform in a real-world scenario. During the Digital Home Keynote, two gamers loaded up Battlefield 2. Both were systems equipped with Pentium 955 Extreme Edition CPUs and ATI CrossFire X1900s. One system implemented Robson, the other was stock.
The system with Robson booted Battlefield 2 and loaded the level nearly 30 seconds ahead of the standard system. While it's not quite the competitive advantage for gamers that the demo suggested, imagine working with large Photoshop files, 3ds max or other applications with large memory footprints that swap a lot of data to disk. Or imagine substantially faster bootup into the operating system. All of these seem pretty attractive.
Robson is potentially a more general solution than building hybrid hard drives. The hard drive business is highly cost driven, but you could add a socket to a motherboard that can be expanded at the user's option. It's unclear whether or not Intel will be licensing Robson widely, only making it available to Intel OEMs or restricting it to Intel-branded motherboards.
Intel suggested that systems or boards with Robson technology would be available at roughly the same time as their next-generation CPU, Conroe. If so, then it's likely that Conroe systems with Robson technology added in will appear even faster than the simple addition of a new, faster processor. So users will get faster processors, lower power usage, and much speedier response times. That's good for all of us.
What is interesting about this though is that Intel looks to be readying this for Conroe, which would add to its already commanding performance leads, not to mention being the first serious attempt to drastically reduce load times.