It seems that Intel (INTC) has finally caught the attention of ARM Holdings (ARMH), and it seems that ARM has begun to sound incredibly desperate when it comes to the performance/watt of its mobile processors. I believe that ARM is now making unsubstantiated, claims regarding the competitive positioning of its products, and I further believe that such actions simply fuel the bear thesis as ARM begins to act and sound a lot like Advanced Micro Devices (AMD) did following the Intel "Conroe" launch.
Remember The Silvermont Launch?
Intel disclosed its next generation low power mobile architecture known as "Silvermont" back on May 6, in which it - in no uncertain terms - claimed that its own internal measurements found that "Silvermont" not only provided better performance/watt at lower power levels but indeed at higher power levels than ARM's Cortex A15 + Cortex A7 in big.LITTLE configuration:
Further, Intel gave public comparisons of the best it expected ARM and its partners to have out during the time that Silvermont-based products hit the shelves (this means comparisons against 4x Cortex A15 Tegra 4, 4x Krait 400 Qualcomm Snapdragon 800, and a hypothetical 4x Swift Apple SoC) and noted that at a given power level, its own parts were vastly superior:
Now, Cortex A15 parts are available today, and Intel has - in its labs - "Silvermont" based silicon, so it has the most accurate view of the competitive comparison. ARM, on the other hand, after seeing these claims from Intel, decided to publish this interesting slide at Computex:
So, after seeing the Intel slide, ARM decided to simply "reverse the results," give a vague metric of performance, and then make a rather bold claim that ARM's already ultra-power-hungry Cortex A15 outperforms Intel's "Silvermont" all while consuming less power. Of course, it bases this on "Intel assertions" while Intel's comparisons are based on in-house testing (since once again, there are a number of A15 implementations available). Now, here's where it gets dicey for ARM...
Bogus Claims About "Saltwell"
See the "Saltwell" measurements? This is Intel's 32nm Atom, and ARM is claiming that A15 has a massive power/performance lead over that chip. Unfortunately for ARM, these results do not tally up in real life. For example, in the popular "AnTuTu" benchmark for Android, dual core "Saltwell" Atom parts actually outperform quad core "Krait" and even "Cortex A15" within the given power envelopes. So here is an AnTuTu run with the Exynos 5 Octa in the Galaxy S IV (int'l edition):
In terms of CPU, the integer score is 5156 and floating point is 5832 for a combined CPU score of 10988. This represents four A15 cores running at up to 1.6GHz in a "phablet" form factor. Now, let's look at the dual core/quad threaded "Saltwell" running the same benchmark in a similarly power constrained envelope:
Intel's CPU combined score is 9399 in a similarly power constrained envelope, suggesting that two of Intel's 5-year old cores built on its old 32nm process can stack up very well against 4 of ARM's latest and greatest A15 designs built on Samsung's 28nm HKMG process. While in a power unconstrained environment, the quad A15s would probably do a whole lot better, the reality of the situation is that when we talk about smartphone/tablet chips, we care about smartphone/tablet form factor results.
If the old Atom is this competitive on 32nm, I find ARM's claims about "Silvermont" which should be substantially faster than the old Atom on a per-core basis, and should come in quad-core configurations for small tablets/phablets, very unlikely to be true in power constrained environments, and I would still even question these claims in power unconstrained environments as Intel probably has some very serious frequency headroom.
ARM is sounding a lot like AMD when it lost its lead to Intel back in 2006, inflated stock price included.
My problem with ARM was, up until now, that it simply traded at too optimistic of a multiple. But now it is starting, for the first time since it became the dominant smartphone force, a FUD campaign against Intel - signaling that for the first time, the firm is running scared. If ARM were using "Intel's assertions," then the given performance/power curves would be swapped in position with ARM's graph, since Intel gave its very own version of the same graph. But unlike Intel - which has certainly tested A15 silicon on 28nm and can make real comparisons - ARM is simply taking advantage of the fact that Silvermont has not yet been publicly benchmarked to make whatever claims that it wants. If these turn out to be bogus, then ARM is not liable and can simply say that it misinterpreted Intel's claims, or that it was merely referring to one particular program/test.
Actions speak louder than words, and real world tests trump fantasy-land projections. Today, in a power constrained environment, 32nm "Saltwell" does quite well against quad 28nm A15s, and at 22nm with a totally new micro-architecture, "Silvermont" will almost certainly drastically outperform ARM's offerings in almost every relevant area. Time will tell, but I think following Computex, we will have a very good picture of just how "Silvermont" stacks up performance wise as it would be very unlike Intel to show off "Baytrail" tablets and not let anybody benchmark them, particularly in light of this recent FUD.
Additional disclosure: I am long INTC