By: Paul A. Clayton (paaronclayton.delete@this.gmail.com), August 11, 2014 1:37 am
Room: Moderated Discussions
David Kanter (dkanter.delete@this.realworldtech.com) on August 10, 2014 9:12 pm wrote:
[snip]
> And I'm quite willing to bet that no ARM server design in the next 2-3 years will provide 80% of the performance
> of the highest bin Xeon. If they are lucky, they might get to the low-end territory. They probably also
> won't have as much memory capacity and generally be inferior across a number of dimensions.
Is comparing highest bin Xeon even with the highest bin of a particular ARM implementation fair? The top bin Xeon might be more than twice as expensive (acquisition cost) as the ARM (both because it has better performance per watt and because it supports x86 software). While processor acquisition cost may not be especially significant in terms of total cost of operation, I am guessing it is not negligible. Comparing median bins seems (to me) more appropriate (as I mentioned before), especially as that allows for a less constrained fraction of sales.
(While physics does not prevent Intel from producing twice as many Xeon chips as will be sold in order to shift the binning upward, it seems likely that, with Intel's volumes (even for Xeon), that economics would prevent such. If there was an unexpected excess of fab capacity—perhaps not entirely unreasonable given fab capacity must be planned long before sales volume is known—it may be cheaper to throw out a month of production to provide more of a high-value product and perhaps even an additional higher bin, especially if that also hurts competitors more than it hurts Intel. However that kind of wild speculation seems a bit weird.)
In addition, comparing the highest performance x86 product with the highest performance ARM product might not be appropriate if TCO is being considered. Supposedly, minimum, typical, and peak power are significant factors in TCO. The difference between a 130W (TDP) x86 and a 90W ARM might be enough give the ARM better TCO.
On a strategic level, viable competition has some value. If the cost of supporting a new ISA in addition to x86 is modest (even if only for a fraction of an organization's servers), then providing a check to Intel's power might have enough value to justify migration costs and dual maintenance costs (if the TCO was otherwise equal). While there is a tragedy of the commons issue, I suspect that some organizations are willing to make modest sacrifices for a common good. The weakness of AMD and the hype around ARM may facilitate such choices (i.e., the danger of an Intel monopoly is greater and the possibility of broader competition more credible [credibility can be increased by hype]).
[snip]
> And I'm quite willing to bet that no ARM server design in the next 2-3 years will provide 80% of the performance
> of the highest bin Xeon. If they are lucky, they might get to the low-end territory. They probably also
> won't have as much memory capacity and generally be inferior across a number of dimensions.
Is comparing highest bin Xeon even with the highest bin of a particular ARM implementation fair? The top bin Xeon might be more than twice as expensive (acquisition cost) as the ARM (both because it has better performance per watt and because it supports x86 software). While processor acquisition cost may not be especially significant in terms of total cost of operation, I am guessing it is not negligible. Comparing median bins seems (to me) more appropriate (as I mentioned before), especially as that allows for a less constrained fraction of sales.
(While physics does not prevent Intel from producing twice as many Xeon chips as will be sold in order to shift the binning upward, it seems likely that, with Intel's volumes (even for Xeon), that economics would prevent such. If there was an unexpected excess of fab capacity—perhaps not entirely unreasonable given fab capacity must be planned long before sales volume is known—it may be cheaper to throw out a month of production to provide more of a high-value product and perhaps even an additional higher bin, especially if that also hurts competitors more than it hurts Intel. However that kind of wild speculation seems a bit weird.)
In addition, comparing the highest performance x86 product with the highest performance ARM product might not be appropriate if TCO is being considered. Supposedly, minimum, typical, and peak power are significant factors in TCO. The difference between a 130W (TDP) x86 and a 90W ARM might be enough give the ARM better TCO.
On a strategic level, viable competition has some value. If the cost of supporting a new ISA in addition to x86 is modest (even if only for a fraction of an organization's servers), then providing a check to Intel's power might have enough value to justify migration costs and dual maintenance costs (if the TCO was otherwise equal). While there is a tragedy of the commons issue, I suspect that some organizations are willing to make modest sacrifices for a common good. The weakness of AMD and the hype around ARM may facilitate such choices (i.e., the danger of an Intel monopoly is greater and the possibility of broader competition more credible [credibility can be increased by hype]).