Introduction
The tech world has kept an interested eye on VIA's Nano since before the turn of the year, but the level of interest in the new processor has grown significantly in recent months, thanks in part to Intel's focus on the ultra-low-power/low-cost market. Over the past six months, VIA has found itself pushed from the perpetual twilight of an also-also-ran into a position of genuine competitive interest. The company is finally ready to sample Nano for performance testing, and I've had the opportunity to put the chip through its paces.
Atom vs. Nano: not a perfect match
In order to test VIA's new chip, I've benchmarked it against Intel's Atom. There's a lot of curiosity out there about how the two low-power processors stack up against each other, and this article attempts to satisfy that curiosity, but it's important to note that this is not an apples-to-apples comparison. According to Intel executive VP Sean Maloney, Atom is "built for low power and designed specifically for a new wave of Mobile Internet Devices and simple, low-cost PC's." As for Nano, VIA's whitepaper (PDF) states: "It [Nano] will initially power a range of ‘slim ‘n’ light’ notebooks." and "will also appear in ultra mobile mini-note devices and small form factor, green desktop systems for home and office use." In this case, we're benchmarking a Nano reference system at the upper end of VIA's product range. The L2100 CPU at the heart of the system is a single-core 1.8GHz processor, with a TDP of 25W.
| Chip | Design Type | Process | Frequency MHz | SMT | FSB MHz | L2 cache | TDP (W) |
| Intel Atom 230 | In Order | 45nm | 1600 | Yes | 533 | 512K | 4 |
| VIA Nano L2100 | Out-of-order | 65nm | 1800 | No | 800 | 1024K | 25 |
While there is a certain degree of overlap between the two processors, it's limited to the relative upper end of Atom's target market and the relative lower end of Nano's. This might not seem so evident at the moment, given the limited number of Atom configurations Intel is currently selling on the DIY market (one), but the two products are focused in two different directions. There are other factors that cloud the comparison, including an early reference platform from VIA and a horribly mismatched processor+chipset combination from Intel, but I've done what I can to tease those differences out and present the two products from a variety of angles.
Performance summary
There are a number of different facets to consider when evaluating Atom vs. Nano, and that's a good thing for Intel. Were this simply a question of which CPU was faster, Nano would win, and by no small margin. Our benchmark results demonstrate that VIA's wunderkind is more than capable of competing in its target market; Nano beat the tar out of Atom in the majority of the tests we ran. The chip might have extended its lead further on a different platform; several tests indicated that the integrated S3 GPU was limiting total performance. Results are directly accessible from the links below. Anyone interested in the questionable effects of benchmark "optimization" should find the PCMark 2005 results of particular interest, while the DVD/HD content playback tests are one spot where the Atom + 945GC chipset pull well ahead of Nano's integrated GPU.
The entire point of these platforms, however, is that they don't focus on raw performance to the exclusion of all else. Power efficiency is at least as important as raw speed these days, but how VIA and Intel rank in this area depends entirely on how we choose to measure performance-per-watt (ppw). If we only consider processor TDP, Atom wins by a landslide. It may lose most benchmarks in absolute terms, but it always remains competitive enough to easily win any power efficiency comparison. So, VIA wins absolute performance but Intel wins power efficiency, right? Wrong. Superman has Kryptonite, Rogue can't touch people, and Atom, for all its super-low TDP, has been effectively hamstrung by the 945GC chipset. With a TDP of 22W, Intel's chipset draws nearly six times more power than the processor itself, a fact that's driven home when you realize that the tall heatsink + fan combination on the retail D945GCLF board is actually cooling the northbridge, rather than the CPU.
The power-hungry nature of its platform destroys any current chance Atom had of establishing itself as a truly low-power alternative. Total system power draw is still quite low by desktop standards, but the D945GCLF's maximum load power of 59W is only about nine percent lower than VIA's reference motherboard. That narrow discrepancy isn't enough to offset VIA's sizeable performance advantage in many benchmarks, and the Nano ends with a higher overall, platform-level performance-per-watt ratio than Atom in many of our benchmark tests.
The bottom line
Nano is an excellent step forward for VIA. It's by far the most compelling CPU the company has ever launched, and could potentially carve a spot for itself in its target market segments. VIA's mini-ITX reference platform is similarly impressive; the board's PCIe x16 slot opens the door for a variety of potential applications that the Atom reference platform can only dream about. Intel's D945GCLF may run just $75 for a 1.6GHz HyperThreaded Atom processor, but it's painfully obvious that the board was designed with an eye towards guarding Celeron sales, and the lack of expansion capabilities hurts Atom's overall attractiveness.
VIA's CN896 chipset may be a better overall fit for Nano, but the integrated Chrome9 HC graphics solution leaves much to be desired. While it proves marginally faster than Intel's GMA950 in some tests, it slumps badly when asked to decode much of anything. The built-in PCIe x16 slot significantly addresses this issue, but the availability of an expansion slot, in and of itself, does not compensate for lousy integrated video—even on a netbook-class solution.
There are too many long-term questions across too many areas to deem this a complete slam-dunk for VIA. Reviewer samples and reference platforms are great for publicity, but VIA has yet to demonstrate that it can ship Nano boards and chipsets in volume. The company has promised that mini-ITX Nano boards will be available in the retail channel by the end of the third quarter, so we should know in a few months if the company can make good on its promise of availability. The fact that Nano is a drop-in replacement for C7 could make the chip attractive to manufacturers with C7-based devices, but again, neither VIA nor its partners have announced plans in this area. A few words from HP confirming Nano as a basis for an upcoming refresh of the 2133 would do wonders for both Nano sales and VIA's reputation.
The largest potential barrier to Nano's long-term success, of course, is Intel. Santa Clara has made no secret of the fact that it believes MIDs, netbooks, and nettops are the future of the industry, and that it intends to offer an Atom that could fit inside any of these devices. Right now that might sound laughable, but Intel isn't kidding. The company's current retail Atom 230-based board might not be what you'd call compelling, but that doesn't mean that future products won't be. VIA's long-term success will be directly proportional to the number of design wins the company can gather for itself in an area Intel has announced it intends to dominate long term.
Testbed configuration
The following components were identical between both testbeds.
- Acer AL2216W 22" 1680x1050 LCD.
- Seagate Barracude 7200.9 250GB HDD
- 2GB OCZ DDR2-1066 @ DDR533
- Windows XP w/SP3 installed
- Enermax 250W PSU
Our VIA reference board consisted of a VIA Nano combined with the CN896 northbridge. More information on the northbridge can be found over at the company's website, but the board we tested is a full implementation, with two DIMM slots and a single PCIe x16 slot. As for Intel's Atom, the company currently offers just one SKU on the retail market. Details on the Atom 230-based D945GCLF are available here.
It doesn't take Nancy Drew's insight to see that these are two very different products. The Nano board is full-featured for a mini-ITX product, with two DIMM slots and a PCIe x16 slot, while the Atom-based D945GCLF can scarcely be expanded at all. Intel makes a limp-wristed token gesture to expansion with a single PCI slot, but the highly integrated nature of the board, combined with the extremely limited performance of PCI, leaves me wondering why the company bothered. As I'll explore, Atom's limited expansion does impact the board's attractiveness, even when judged solely by the standard of its target low-end market. Both systems were tested with 2GB of RAM, and both are limited to a single, 64-bit memory channel.
As I mentioned earlier, the large heatsink+fan on top of the Atom board is cooling the chipset, not the CPU. The fan itself isn't audible unless you literally put your ear up to it. Neither the Atom nor the Nano testbed produced any noise above a simple operating hum unless a CD was spinning up in one of the drives.The largish fan on VIA's reference board is an integrated solution that covers both northbridge and CPU, with a separate heatsink for the southbridge.
I debated between Windows XP and Windows Vista, but eventually settled on XP as the better fit for these type of systems. The LCDs I used both support 1680x1050 resolutions, but Worldbench 5 runs all of its tests in 1024x768 by default. This seemed a reasonable resolution for the systems anyway, so I kept it for all further performance testing. The one exception was when I tested HD content playback—I pushed both monitors up to 1680x1050 when measuring DVD, HD DVD, and MPEG-2 performance.
As for the benchmarks themselves, I chose a suite of older tests that could measure system performance without requiring Vista or requiring more than 2GB of total RAM. Tests like Cinebench and PCMark aren't strictly representative of likely workloads, at least not in the netbook market, but they provide additional subsystem performance information. Ultimately, I think we'll see a new series of MID and netbook-specific tests emerge as this market matures.
One final note before we turn to benchmarks. Although I haven't broken the results out here, I ran a number of tests with HyperThreading disabled, in order to judge how much of a performance boost Atom gained from the feature. Back in the Netburst days, Intel targeted a ~20 percent performance increase with HT enabled over HT disabled in an appropriately threaded test, but that was years ago, and a very different architecture. Atom is an in-order architecture, and as such, is potentially vulnerable to stalling out. HT could potentially help the processor keep its pipeline full by seeding the execution units with additional uops.
Based on what I saw, HyperThreading plays a major role in keeping Atom data crunching. Exact results varied from application to application, but enabling HT often boosted performance by around 50 percent. Almost all of the Atom SKUs Intel has announced will offer HyperThreading; it seems to be an essential part of the secret sauce that keeps Atom (somewhat) competitive.
The other advantage of HT is the way it smoothes system responsiveness. The Nano might be faster than the Atom, but the chip suffers from the same small stutters and tiny pauses that impacts all single-core processors when the OS is busy doing other things. Thanks to HyperThreading, Atom feels smoother, and the system remains more responsive while engaged in other tasks. This effect is something of an optical illusion, since a system that finishes a task more quickly can move on more rapidly to do something else, but it does give Intel's new core a perceived advantage.
Worldbench 5
Worldbench 5
Before there was Worldbench 6.0 Beta 2 for Vista, there was Worldbench 5 for XP, and it's this older version of the test we're examining today. I've highlighted certain tests—the Roxio Videowave test refused to run on either system—so there's no composite WB score listed for either platform.
VIA sweeps the first series of tests, save for Nero Express 6, where Atom pulls ahead. This is almost certainly a platform issue; Nero tends to be sensitive to southbridge performance and Intel has often held an edge over its competitors in these types of benchmarks. Note that the gap between Atom and Nano can vary widely; this will be a recurring factor in other tests.
Another set of tests and another set of wins for Nano, though Atom does come within 14 percent of VIA's new core in the OfficeXP benchmark. Of all the applications we've examined thus far, this (and possibly Nero) are the two most likely to be encountered by actual Atom or Nano users, and the two cores have compared relatively well thus far, though Atom does slip back quite a bit in WME9 encoding, despite that application's support for SMT.
Our final set of Worldbench results are another clean sweep for Nano, though we do see the impact of Atom's HT support in several tests. VIA wins both the WME9 and Mozilla 1.4 tests, as well as the multi-tasking test that combines the two, but it takes the chip 1.77x longer to complete the combined task than it did to finish the Mozilla 1.4 benchmark. Atom takes almost twice as long to finish its Mozilla test, but tossing a WME job on top of the browser only extends the test length by about 33 percent.
No comments:
Post a Comment