David Kanter (dkanter.delete@this.realworldtech.com) on August 8, 2014 11:51 pm wrote:
[snip]
> For a 3-issue core, there is no way that using the ARM ISA will give more than a 5-10% advantage over
> Intel (all things being equal). However, things are not equal. Intel has larger design teams, Intel
> has better process technology, Intel has higher volumes and can bin more aggressively, etc.

I was going to comment that the binning advantage would not apply to median performance, but then I realized that even if one was comparing median performance more binning would still provide a price/performance advantage since binning increases revenue. In addition to the benefit of higher prices and higher volume (from being able to sell less useful parts), the option of a better product can not only encourage sales (a preference for buying from a vendor that gives some choice now and implicitly promises such for future products) but encourage higher-value sales (a customer choosing the second or third best among ten options will tend to provide more revenue than one choosing the middle option among three). The marketing effect of being the best in some figure of merit is also significant.

Using median performance seems to have some justification for a certain type of comparison. While best against best¹ is meaningful in some contexts (e.g., when modest benefits provide great value), most users do not — and cannot — choose the highest performance per watt or per price bin.

On the other hand, binning could reduce the niche benefit that ARM vendors are more likely to have. Intel's mainline x86 seems to have generally targeted personal computer workloads with even server workloads seemingly given less attention in the core,² so ARM vendors have some opportunity for more specialized targeting. However, some binning strategy might prevent ARM vendors from getting a large foothold.³

The effects of off-CPU system factors further complicate analysis. E.g., the more CPU energy use is lowered, the more the energy use of other system components becomes a limiting factor (the extension of Amdahl's law to partial improvements generally).

With respect to Intel's larger design teams, I was under the impression that part of this was Intel strategy of bringing in a larger proportion of less experienced engineers. The benefit of larger teams is also not linear since communication effects must be considered.

Furthermore, the design difficulty does not increase linearly with performance. Matching a 5% ISA performance benefit would be expected to increase design difficulty by more than 5%. I also suspect that target specialization could provide a larger benefit than ISA. Intel has started providing more product diversity with the introduction of Atom and Many Integrated Cores, but some specialization opportunities may remain.⁴

Included in the "etc." and related to process technology (given that Intel is an integrated design and manufacturing company), I was also under the impression that Intel has better than average development tools (including validation tools).

---

¹ An informed rational chooser would choose the best SKU for a likely complex collection of concerns. Calculating total cost of operation, including lost opportunity from specialization or lack of performance headroom, is probably difficult even within a single organization, and no single figure of merit can be applied to all users.

²Of course, the uncore — core count, on and off-chip interconnect, higher level cache capacity, memory controllers, I/O integration, etc. — is more important for server workloads and less specialized workload targeting avoids glass jaws which can unexpectedly occur even in workloads very similar to a specifically targeted workload.

³ I do wonder if x86 system vendors, with Intel's cooperation, could use access to low volume bins at median-bin pricing as a loyalty reward. This would allow high-value processors to be sold at median prices without significant pressure on pricing and volume for lower-value processors, at least outside of the privileged customers. This kind of strategy might reduce the opportunities for ARM vendors to get significant adoption.

⁴ One subject that does not seem to have been addressed well is how much the various server ARM vendors will be competing with each other. If their products are sufficiently differentiated, there might be more of a death-by-a-thousand-papercuts effect for Intel, but if they are largely competing among themselves, then the effect would be more like dividing a land inheritance among many children where no child gets enough land to survive. (I suspect the latter will be much more the reality.) Vendors that have even a modest guaranteed user base (in-house or closely aligned organization) would seem to have an advantage as would vendors with applicable special intellectual content and vendors with established distribution channels. David Kanter, with his economics background, could probably do an unusually good article on this subject.