Original Link: http://www.anandtech.com/show/6981/the-kabini-deal-can-amd-improve-the-quality-of-mainstream-pcs-with-its-latest-apu
The Kabini Deal: Can AMD Improve the Quality of Mainstream PCs with Its Latest APU?by Anand Lal Shimpi on May 24, 2013 1:45 PM EST
There are two non-negotiables in building a PC these days: the cost of Intel silicon and the cost of the Windows license. You can play with everything else but Intel and Microsoft are going to get their share. Those two relatively fixed costs in the PC bill of materials can do one of two things: encourage OEMs to skimp on component cost elsewhere, or drive the entire ecosystem to supply higher quality components at lower prices. If you’ve been following the PC industry for the past decade, I think we’ve seen more of the former and less of the latter.
Apple occupying the high-end of the notebook PC space has forced many OEMs to reconsider their approach, but that’s a more recent change. What AMD seems to offer is an easier path. AMD will take less of the BoM, allowing OEMs to invest those savings elsewhere - a move Intel will never make. Given how much pressure the PC OEMs have been under for the past few years, AMD’s bargain is more appealing now than it has ever been.
With Llano and Trinity, AMD’s story was about giving up CPU performance for GPU performance. With Kabini, the deal is more palatable. You only give up CPU performance compared to higher priced parts (you gain performance compared to Atom), and you get much lower power silicon that can run in thinner/lighter notebooks. Typically at the price points Kabini is targeting (sub-$400 notebooks), you don’t get pretty form factors with amazing battery life. AMD hopes to change that.
While AMD hasn’t disclosed OEM pricing on Kabini (similarly, Intel doesn’t list OEM pricing on its mobile Pentium SKUs), it’s safe to assume that AMD will sell Kabini for less than Intel will sell its competing SKUs. If Kabini’s die size is indeed around 107mm^2, that puts it in the same range as a dual-core Ivy Bridge. AMD can likely undercut Intel a bit and live off of lower margins, but there’s one more component to think about: Ivy Bridge needs its PCH (Platform Controller Hub), Kabini does not. As a more fully integrated SoC, Kabini’s IO duties are handled by an on-die Fusion Controller Hub. Intel typically charges low double digits for its entry level chipsets, which is money AMD either rolls into the cost of Kabini or uses as a way of delivering a lower total cost to OEMs.
Traditionally, OEMs would take these cost savings and pass them along to the end user. I get the impression that AMD’s hope with Kabini is for OEMs to instead take the cost savings and redeploy them elsewhere in the system. Perhaps putting it towards a small amount of NAND on-board for a better user experience, or maybe towards a better LCD.
As we found in yesterday’s article, Kabini does a great job against Atom and Brazos. However, even with double digit increases in performance, Kabini is still a little core and no match for the bigger Ivy Bridge parts. Much to our disappointment, we pretty much never get sent low end hardware for review - so to make yesterday’s NDA we had to stick with 17W Ivy Bridge and extrapolate performance from there. In the past day I grabbed an ASUS X501A system, a 15-inch entry-level machine priced in the low $300s. More importantly, it features a 35W Ivy Bridge based Pentium CPU: the dual-core 2020M.
The Pentium 2020M’s base clock speed is still relatively high at 2.4GHz, but there’s no turbo. In the low level CPU performance analysis yesterday I used a Surface Pro with a 17W 1.7GHz Core i5, but max turbo on that part can hit 2.6GHz. I actually don’t expect there to be huge CPU performance differences between the results from yesterday and what I have here but I wanted to be sure.
The bigger difference is actually on the GPU side. While a mobile Core i5 comes with Intel’s HD 4000 graphics, the Pentium 2020M gets a vanilla Intel HD GPU. At 22nm, Intel’s HD graphics is a 6 EU part with lower performance than the previous generation Intel HD 3000.
With that said, let’s see how Kabini (AMD’s A4-5000 in particular) compares to an Ivy Bridge based Pentium 2020M in CPU performance, GPU performance and finally in power.
CPU Performance & Power
On the CPU front, the difference between the Core i5-3317U and the Pentium 2020M is actually quite small. The former has a nice 2.6GHz max turbo while the latter has a higher TDP and thus a higher base clock as well. The result is that most single threaded performance results are very close between the two. If we look at Kraken, SunSpider, PCMark, Cinebench (1-thread) or 7-Zip (1-thread) - the two perform quite similarly. As a result, Kabini doesn’t really gain any ground here. In my own use, I can feel a performance difference between the 2020M and the A4-5000 in tasks like installing/launching applications, as well as bigger CPU bound activities.
|PCMark 7||Cinebench 11.5 (Single Threaded)||Cinebench 11.5 (Multithreaded)||7-Zip Benchmark (Single Threaded)||7-Zip Benchmark (Multithreaded)|
|AMD A4-5000 (1.5GHz Jaguar x 4)||2425||0.39||1.5||1323||4509|
|Intel Pentium 2020M (2.4GHz IVB x 2)||4214||1.00||1.96||2856||5434|
|Intel Core i5-3317U (1.7GHz IVB x 2)||4318||1.07||2.39||2816||6598|
A big issue here is Kabini, at least in its launched versions, lacks any turbo core support. The 15W A4-5000 runs even single threaded tasks as if all four cores were active and eating into that TDP budget. The fastest Jaguar implementation seems to be 2GHz, but even if the A4-5000 could turbo up to that level I feel like I’d still want a bit more. There’s obviously room on the table for a Kabini refresh, even at 28nm.
For light web browsing and general use workloads Kabini, like many modern platforms, can really be good enough.
It is impossible to have a performance discussion without looking at power consumption when it comes to mobile devices. This is where Kabini makes up a lot of ground. The Pentium 2020M is a 35W part (Intel does offer slower 17W parts but I unfortunately don’t have a system that uses one of those), compared to the A4-5000’s 15W TDP. I measured total platform power of both notebooks without a battery and with the display disabled (and using the same SSD in its lowest power state). While isolating SoC power would be ideal, this does give us a general idea of platform power consumption:
|Platform Power Consumption|
|Idle||Cinebench 11.5 (1-thread)||Cinebench 11.5 (multithreaded)||7-Zip (1-thread)||7-Zip (multithreaded)|
|AMD A4-5000 (1.5GHz Jaguar x 4)||4.75W||7.91W||11.5W||7.9W||11.3W|
|Intel Pentium 2020M (2.4GHz IVB x 2)||8.14W||17.9W||22.4W||17.6W||21.7W|
The difference is pretty big. Kabini will either last longer on the same size battery, or be able to fit into a smaller chassis altogether. I also suspect the 15W TDP is perhaps a bit conservative, total platform power consumption with all CPU cores firing never exceeded 12W (meaning SoC power consumption is far lower, likely sub-10W).
It’s also worth pointing out that there’s clearly a lot of thermal headroom when only a single CPU core is active. Design limitations would probably keep a single core from ramping up too high, but there’s clearly room for improvement here.
The 17W Pentium/Celeron parts are architecturally very similar to the 2020M I’m featuring here, they just run at 75% of the clock speed. If we assume perfect scaling, Intel would appear to still hold substantial single-threaded CPU performance advantage even if the comparison was to a lower clocked Pentium. Interestingly enough, the multithreaded advantage would pretty much disappear though. These 35W Pentiums seem a lot more common in retail (likely because of the spec shopping that’s presumed at these lower price points).
Compared to Atom (or Brazos), Kabini does extremely well though. Similar to Brazos, AMD is looking for Kabini to do battle slightly above its weight class. In its press materials AMD specifically calls out Pentium and Core i3 as potential targets for the A4/A6-class Kabini APUs. Part of this is AMD looking at the CPU and GPU as a whole though, so let’s move on to the graphics comparison.
Looking at 3DMark’s Ice Storm test, the comparison between Intel’s 22nm HD Graphics in the Pentium 2020M and the Radeon HD 8830 in the A4-5000 is extremely close. In fact, across almost all of the 3DMark benchmarks we see the two perform very similarly. The lone exception being 3DMark 11 where the A4-5000 maintains a significant lead and even approaches Trinity in terms of performance (making it feel more like a fluke than the norm).
Turning to GFXBenchmark (formerly GL/DXBenchmark), we see performance tilt in favor of Kabini once again. The T-Rex HD test is extremely shader intensive. There’s about a 20% gap in raw shader performance between the 2-CU GCN implementation in Kabini and the 6 EU Gen7 graphics core in the Pentium 2020M, which maps almost perfectly to the performance delta we see in T-Rex HD. Now we see where the Pentium/Core i3 comparison comes from.
|3DMark Ice Storm||3DMark Cloud Gate||3DMark Fire Strike||3DMark 11||3DMark 06||GFXBenchmark T-Rex HD|
|AMD A4-5000 (Radeon HD 8330)||23196||2159||310||580||3803||37 fps|
|Intel Pentium 2020M (HD Graphics)||23135||2168||285||401||3542||30 fps|
All of this is fine if we’re looking at theoretical GPU benchmarks but what about actual games? In our Kabini review Jarred found the A4-5000 to be incapable of playing modern titles at reasonable frame rates, but what about titles from a few years ago? To find out, I dusted off Oblivion (with the Shivering Isles expansion) and threw it on my Kabini, Brazos and IVB Pentium systems.
I configured all three systems the same way: 1366 x 768, with medium graphics quality presets. I even used our old Oblivion SI benchmark from 2007. The results seemed to mirror what we saw in 3DMark:
|Oblivion - 1366 x 768 Medium||Diablo III - 1366 x 768 Low||Oblivion - Power Consumption|
|AMD E-350 (Radeon HD 6310)||20.1 fps||21.9 fps|
|AMD A4-5000 (Radeon HD 8330)||26.1 fps||25.8 fps||15.2W|
|Intel Pentium 2020M (HD Graphics)||27.7 fps||20.3 fps||31.4W|
Kabini is about 30% faster than Brazos in GPU performance, and almost identical to the Pentium 2020M. Intel has a 6% performance advantage here, but I’m wondering if that’s from the CPU and not the GPU (Oblivion tends to hit both pretty hard). At lower quality settings (and/or resolution) you can definitely get Kabini above 30 fps, but even here I’d say it’s playable. More importantly, it’s performance competitive with Intel’s HD graphics.
I was also curious to see how Diablo III ran on Kabini so I fired up an early save and ran through the Cemetery of the Forsaken recording average frame rate. On a more modern title, both Kabini and Brazos actually hold a performance advantage over the Pentium 2020M.
As far as power goes, Kabini delivers relatively similar performance at roughly half the power of the Pentium 2020M.
With any of these integrated GPUs, the gaming experience even on previous generation high-end titles isn’t going to be a walk in the park.
AMD’s family of cat-cores began their lives as higher performing alternatives to Intel’s Atom. Ambition (and a desire for higher ASPs) drove them to compete with the lower end of Intel’s Pentium lineup. The Jaguar based Kabini APU continues the trend and even aims higher up the product stack, with the highest spec Kabini APU aiming for Intel’s Core i3.
When it comes to GPU performance, Kabini is absolutely there. Depending on the benchmark we either saw parity, a slight disadvantage or large advantage for Kabini’s integrated Radeon HD 8830 compared to Intel’s 22nm HD Graphics.
On the CPU front, we already established that Kabini runs laps around Brazos and Atom. The point of today’s article however, was to find out if comparing to a $320 notebook based on a Pentium 2020M would change things. The answer ends up being, not really. Kabini is very efficient for what it is, but the Ivy Bridge cores simply have better single threaded performance. Even a lower clocked/lower TDP Pentium part would maintain a healthy single threaded CPU performance advantage.
Whether that CPU performance advantage matters really depends on your usage model. Anyone who’d be happy with a Cortex A15 based Chromebook for example would likely be just fine with Kabini under the hood of their next notebook.
Where Kabini excels however is in its power consumption. Better battery life than a Brazos based notebook is a given, but in our earlier comparison we even noted better battery life than a 17W Ultrabook. I suspect AMD is indeed being a little conservative with its 15W TDP rating for the A4-1500.
That brings us to the deal that AMD is offering OEMs. A lower cost, lower power APU that’s better than Atom/Brazos, comparable to Pentium in GPU performance but behind it in single threaded CPU performance. Historically PC OEMs have taken the cost savings AMD offered them and delivered a lower priced system, my hope with Kabini is that we’ll finally get something different.
Kabini alone offers a power advantage, which itself may be enough, but if an OEM were to take Kabini's cost savings and put the money towards a better LCD or better storage that could significantly alter the balance of things. I agree with what Jarred said in his conclusion yesterday: "Give me a reasonable Ultrabook-style chassis (or maybe a dockable tablet) with Kabini and a decent quality 1080p touchscreen and do it at the right price and there are plenty of people that will jump at the offer."
The traditional approach would be for an OEM to offer a Kabini system at a lower price than a Pentium based system. The Kabini system might offer better battery life, while the Pentium machine delivers better single-threaded CPU performance. Should that same OEM offer both systems at the same price, but give the Kabini an appreciably better display, I think many users would be willing to overlook the difference in CPU performance entirely.
As it stands, I like Kabini. It’s a good part that could make for the foundation of a nice, affordable ultraportable system. The best selling notebook on Amazon remains the Cortex A15 based Samsung Chromebook. The second best selling? A Core i3 based ASUS X202E. Consumers are clearly willing to sacrifice CPU performance if the price is right and the rest of the machine/experience is well designed. I see no reason that a Kabini based platform couldn’t be up there as well. It’s just up to an enlightened OEM to do the right thing with what AMD has given them.