1. Producing Bid Evaluation Formulae An underlying economic model, and the resulting nature of justifiable evaluations for node procurement bids

2. What Are We Trying to Do? More “bang for the buck” in acquisitions –Better match of evaluation formula to our actual needs –Inclusion of considerations beyond just CPU power per dollar –Allowing flexibility to take advantage of cost-effectiveness opportunities Sound economic footing to justify our evaluation formulae

3. Procurement Preparation “Recipe” We can’t just come up with one formula –Each round of procurement is different This task force should come up with –A list of factors that are relevant to the evaluation formula for any nodes acquisition –A set of values for some of those factors which are universal –A mechanism for taking those factors and creating an evaluation formula tailored to the specific acquisition

4. Overall Concepts Keep the Formula Simple –delay costs about.6 % per week –Hard to make this up by slight accuracy improvement CPU units –Choose accurate benchmark(s) but this model is orthogonal to details Wattage –A shorthand for conditioned power needs, cooling, space Computational Luminosity –Anticipated and actual

5. Overall Concepts Value Curve –How much is the availability of a given amount of CPU power at some specified time worth, in terms of physics output? Value is driven by physics output –One can assign a dollar value based on the lab expenses Value increases with CPU power Value is not flat because of: –Uncertainty in actual amount of CPU that would be needed –Diminishing returns on analyses that open up with more CPU availability

6. The Value vs. CPU Power Curve An adjunct document discusses this curve The curve can be expressed as value of obtaining a given expected computational luminosity, and parameterized by the anticipated computational luminosity Effective computing power C is at least as large as actual computational luminosity Effective computing power C is deficient with respect to computational luminosity

7. Value curve looks like A.C.L Optimum Acquisition Level

8. Properties of Value Curve Scales with ratio of acquired to anticipated computational luminosity –Just one key parameter – ratio of sigma to anticipated computational luminosity Negative second derivative –Meaning that overshoots are not as valuable as undershoots are harmful Slope and curvature help determine numbers going into the bid formula

9. Overall Concepts (continued) Moore’s Law –Different evolutions for different … Technologies Aspects (CPU power, Wattage,...) Physics Timeliness Factor –Physics results today are worth more than equivalent results a year later Node Useful Lifetime (assumptions) Imperfectly Fungible Assets –Considerations that can’t freely be traded against price

10. General nature of model The formula will assign a “net value” to any given bid –Net value will depend on CPU power acquired – benchmark based Assets required to support the nodes –E.g., Wattage Corrections for desirable or non-preferred features –E.g., nature of racks and assembly Present and future value curves –Some key details are discussed later The winning bid will be the one with the greatest “net value” per unit cost –Allows for comparison of a bid of $5M for k CPU’s with a bid of $4.98M for similar (non-identical) units

11. How the bid will look “Net Effective Cost” : add up –Bid amount –Support Asset Values (e.g., $400 per Watt required) –Cost additions or (negative) preferences (e.g., $86 per CPU if we must assemble) “Net Value” : –Benchmarked CPU units * N –Possibly modified by bonus/penalty factors (e.g. reliability history bonus) –Translated through value curves –Reduced by expected value of decrease in future values due to consumption of non-fungibles

12. The Bid Formula We will be careful not to double-count the same effect – –The power consumption will likely appear in two places; the model will ensure that the proper cost is reflected (Dimensionless) Figure of merit is

13. Non-fungible assets (And their impact on the model) The prototypical Non-fungible is Power/Cooling/Space –We group these together conditioned power dominates over space costs cooling tracks with power –We can say that it costs $450K to provide a building with conditioned power and cooling for 900 W – hence power needs cost $500/ W –But that would be incorrect modeling because we are not (completely) free to get 10% more power than was planned, nor to “give back” 10% of the power capability we will have built

14. Non-fungible assets The proper model is to include the effect of consuming these assets on the expected amount of computing we will be able to acquire in future bids This is particularly important in comparing two technologies, one of which is superior in CPU/$, the other in CPU/W. –The point is that by buying the power-hungry choice today, we may preclude buying the power-hungry option next year or down the line.

15. Impact of Non-fungible asset needs When calculating the impact of power requirements, we need to consider –Value curve (this is where  is needed) –Physics timeliness factor –Profile of intended conditioned power availability And options, if any, to adjust that profile –Rough expected profile of computational luminosity or CPU expenditures –Moore’s law estimates for the competing technology choices

16. Talk is prepared up to this point Remainder will be added Thursday Morning!

17. Factors that could be part of a bid evaluation

18. Information that will apply to each bid formula Needed Maybe needed

19. Information for overall decision Needed Maybe needed

20. What next?