After all of the excitement leading up to it, Advanced Micro Devices (AMD) has finally launched its next generation low power, thin-and-light oriented product called "Kabini" for the 9W to 25W space. Anandtech, the world's leading hardware review site, managed to get its review up in a timely fashion, and as a result, the financial community can get a better sense of just what this product means for AMD and its future.
The "Kabini" Lineup - 9W to 25W TDP
Here's the lineup of parts that AMD is launching today:
So what we see here is at the very high end, we get two quad core APUs: one with a TDP of 25W and one with a TDP of 15W. The dual cores consume from 9W to 15W, with varying graphics/CPU clock speeds.
These will, as AMD points out in its competitive comparison slides, go head-to-head with Intel's Celeron/Pentium/Core i3 lines:
How About "Temash"?
While "Kabini" is designed to try to win back share in the budget laptop space, AMD is really trying to drive down power in order to get into tablets. The "Temash" 4 core is an 8W part (I told you so), the really interesting one is the lowest power 3.9W dual core part. Here's the full lineup:
An Interesting Product Lineup
The main advantage that I see for "Kabini" is that it's a single-chip SoC, rather than a dual chip solution like the low end "Celeron" and "Pentium" parts. This means that the entire platform is lower power (this means battery life), and it further means that OEMs need only buy a single chip, rather than a dual chip from Intel (INTC) (CPU + PCH). If priced well, I could see AMD gaining back some share at the low end of the PC market, and perhaps even enabling fundamentally better systems at those price points.
Intel's Response Is Still A Ways Off
Intel, too, will be moving its low end PC line to the Atom-based "Bay Trail-M" and "Bay Trail-D" platforms, but these will likely arrive a fair bit later to market than AMD's solutions, giving AMD a bit of a window of opportunity to have this segment basically all to itself. Now, based on what Intel presented at its "Silvermont" launch, and based on the numbers Anand gave in his "Kabini" review, we can start to make some educated guesses on how the "Bay Trail" platforms will fare.
Comparing "Temash" To Potential "Baytrail-T"
Anandtech ran "Cinebench" which is a measure of floating point performance, and ran 7-Zip which is a measure of integer performance. In single-threaded operation, the 1.8GHz "Saltwell" (old, crappy Atom) scored 0.17 in Cinebench and 754 in 7-Zip. The 1.5GHz "Jaguar" hit 0.39 in Cinebench and 1323 in 7-Zip. This implies that normalized for clock speed (and assuming that performance scales linearly with clock), the old "Saltwell" Atom offers 47% of Jaguar's performance in integer workloads and 36% of Jaguar's performance in floating point workloads - yikes!
So, with Silvermont offering a ~50% IPC increase (let's assume it's equivalent for integer and floating point - although the Intel architect I spoke with suggested that floating point improvements should be even more pronounced than integer) over "Saltwell," we get to roughly 70% of "Jaguar" in integer workloads and 54% in floating point workloads.
The real rub, though, is that while on a per-clock/per-core basis, the AMD part wins, the 1GHz dual core "Temash" at a 3.9W TDP will actually be going toe-to-toe not with a 1GHz dual core "Silvermont," but more along the lines of a quad core 2.1GHz+ "Silvermont" in a ~3W TDP:
Doing some quick and dirty math, this suggests that if the multithreaded results for quad core "Kabini" at 1.5GHz in the Cinebench and 7-Zip benchmarks came out to 1.5 and 4509, respectively, then we should expect a 1GHz dual core "Temash" to hit approximately 0.5 and 1503, respectively. Assuming that clock-for-clock, core-for-core the "Bay Trail" gives a +50% boost over "Clover Trail," and then assuming we are running at 2.1GHz as suggested above, and in a quad-core configuration, I would expect the multithreaded scores in the benchmarks above to hit 1.19 and 5278 in a similar form factor. Keep in mind that this is mostly a "ballpark" estimate to try to get a feel for the ranges, and not perfect numbers. In short, in thin and light tablets, I expect quad core "Baytrail-T" to have a significant CPU advantage. But what about graphics?
On the graphics side, according to Anand's "Kabini" review, it seems that a 16 execution unit ("core") Ivy Bridge graphics part with graphics clocks that go from 350MHz to 1050MHz, the Kabini 2 Compute unit/128 shader "graphics core next" part offers 70% of the low voltage Intel graphics part (in 3D Mark - I know this isn't the most perfect comparison, but it's a good proxy):
So, extrapolating a bit, we can assume that if the 500MHz 2CU part gets 26019 - and once again assuming perfect scaling with clock speed - then the 2CU 225MHz part in "Temash" 2 core should score roughly 13009. Assuming that the Intel leaks are good, then we are looking at a "cut down" version of the Ivy Bridge graphics at roughly the same clock, but with 4EUs rather than 16, which means we should divide the Intel i5 results above by about 4, to yield 9182. Advantage here is, of course, to AMD by a significant margin.
Die Size - Do The Economics Work?
While all of this talk is about performance, the real key is do the economics work out? AMD's real problem - above all else - has been that its "Bulldozer" and "Steamroller" line of chips were large, watt-sucking, and commanded significantly lower ASPs than chips from Intel with substantially smaller die sizes. The economics of AMD's "big cores" just don't work, and I think the real winner here is going to be AMD's "small cores."
Anyway, according to Anandtech's measurements, the four "Jaguar" cores, plus the 2 GCN compute untis measure to about 107mm^2. In comparison, Intel's "Ivy Bridge" 2 core Ultrabook part (with the good graphics) measures in at about 118mm^2. Note that AMD's "Kabini" includes the PCH on-de while Intel relies on an external controller (that is built on previous generation process technology), so this comparison is slightly unfair. The real comparison will be against "Baytrail" and "Haswell-ULT" due out later this year, as these are the primary competitors to the Kabini parts. But at the end of the day, this is a good, economical die size built on a fairly mature 28nm process at TSMC (TSM).
Conclusion - AMD Has A Good Part Here
AMD has a pretty good part here with "Kabini." I would expect that in low end laptops, AMD has a chance of regaining some share against Intel's low end line. However, I would not get terribly excited about the dual core "Temash" part against Intel, as I expect that "Baytrail-T" will offer a significant CPU performance and battery life advantage (the latter primarily to more sophisticated sleep states - particularly S0ix). "Baytrail-T" should lose in graphics performance, however, which could create another interesting niche for AMD to exploit.
The economics work, the part offers good battery life, and it offers decent performance. It's not going to best Intel's "Core" products on a per-watt basis in ultrathin notebooks, and I suspect the graphics delta gets less favorable for AMD with Intel's "Haswell GT3" in the 25W space, and only slightly worse in the 15W space:
But at the end of the day, AMD has put out a part that will help it regain market share in the higher volume segments of the market, and that's what I believe will drive the company to profitability. It's not a performance race with Intel, it's about making money for the shareholders of the company. This is the right product with which to do that, and I think AMD is finally doing what's best for investors.