Shortly after the launch of my piece, "AMD's Kaveri Looks Like A Dud", I was very fortunate to find that all of the slides pertaining to the NDA session for the Kaveri part were leaked. While the leaked benchmarks from Baidu seemed to be mostly on-target, AMD's (NYSE:AMD) own marketing materials help me flesh out my thesis on "Kaveri" in quite a bit more detail.
First, The Performance Question
In my prior article, I made the following claim,
In fact, given that the Intel Iris Pro graphics is already right about where the A10-6800K is (except with much more CPU performance and lower power consumption), things are going to look exceptionally bad for AMD's high end PC and notebook APUs once Broadwell (with an entirely new GPU architecture) on 14 nanometer hits.
While the "unofficial" benchmarks pointed me to the idea that AMD's latest "Kaveri" will have graphics performance right around Intel's (NASDAQ:INTC) latest Iris Pro graphics, AMD's official numbers actually paint a slightly less optimistic picture (that is, if 3D Mark Fire Strike is to be believed):
In 3DMark Fire Strike, the highest performing 95W TDP part scores 1,541, while the 65W pulls in 1,304 and the 45W part scores 1,287. On the Intel side of things, the 47W Iris Pro 5200 scores 1,543 and the 55W version scores 1,567. This largely confirms my thesis that as far as graphics performance goes, Intel's "Iris Pro" GPU is right on par with AMD's latest-and-greatest today.
Okay, But AMD's Chip Is Cheaper!
So, there are two interesting things about AMD's official slide-deck:
- We get a die-size
- We get pricing
So, here's the official skinny on how large AMD's Kaveri chip is,
It's 245mm^2 and sports a whopping 2.41 billion transistors on Global Foundries' 28nmSHP process (gotta admit - this looks like a neat chip). This is absolutely gigantic! Now, this is all well and good until we see what AMD is trying to sell it for:
So the very highest end version will go for $173 (with a free copy of Battlefield 4 thrown in, too), then the lower end 7700K and 7600 will go for $152 and $119, respectively. Notice, though, how AMD is positioning these chips - they're much cheaper than their Intel counterparts. What's interesting here, though, is that AMD is positioning chips with a die size of 245mm^2 against parts that have a die size of about 177mm^2 (per Anandtech).
Now, don't get me wrong, AMD's parts will likely offer vastly superior graphics performance (but inferior CPU performance), but the point here is that this is a real economic problem. AMD is:
- Selling chips that are 38% larger for as much as 29% cheaper
- Since Intel owns its own wafer fabrication plants (and does not need to pay an external foundry's 30%+ gross margins), Intel's cost structure is much superior to AMD's.
So, you see the economic problem here, right? AMD shoved a ton of GPU performance onto its chip and is selling the part for a discount to its much smaller and cheaper to make rivals. Oh, and the company making the competing parts has a much better cost structure.
Nothing fundamentally changes from my initial analysis. AMD is selling a more expensive die (with much better GPU performance than its competitors but much weaker CPU performance) for less even though it has a less favorable cost structure. This is pretty bad for margins, but the upside of this is that AMD will not need to suffer take-or-pay charges as long as they can move these parts. Intel will probably need to respond with some price cuts of its own thanks to this "price war".
But the real issue here? People buying $150+ CPUs with the intention to play games will buy a discrete graphics card from either AMD or NVIDIA (NASDAQ:NVDA) and will likely seek to maximize CPU performance as a result of this. This is why AMD's approach of shoving a ton of GPU performance onto a "gamer" oriented CPU suffers the ironic fate of being undesirable outside of a few niches. For desktop gaming-oriented CPUs, a discrete GPU and performance-maximized CPU is the way to go for many gamers.
Integrated graphics, however, is great for tablets, smartphones, and notebooks, and if AMD's Kaveri has excellent power management and performance in low power envelopes when the notebook versions of these parts roll out, then this could be a differentiator. Until then, Kaveri's strong integrated graphics performance will probably be lost on its target audience unfortunately enough.
But it seems from the leaked slides that only the desktop versions are launching, putting Kaveri for notebooks in either Q2 or Q3 where it won't have the spotlight to itself for long before Broadwell shows up - and that'll be an interesting battle to watch.
Additional disclosure: No position in AMD.