MODIFICATION
D -- Google Cloud Services
- Notice Date
- 9/8/2017
- Notice Type
- Modification/Amendment
- NAICS
- 541519
— Other Computer Related Services
- Contracting Office
- Department of Energy, Fermi National Accelerator Laboratory, Fermilab, P.O. BOX 500, MAIL STOP 210 (WH4W), BATAVIA, Illinois, 60510
- ZIP Code
- 60510
- Solicitation Number
- 277111WPK
- Archive Date
- 11/30/2017
- Point of Contact
- William P Koncelik, Phone: 630-840-4173, Steven Gaugel, Phone: 630-840-5782
- E-Mail Address
-
billk@fnal.gov, sgaugel@fnal.gov
(billk@fnal.gov, sgaugel@fnal.gov)
- Small Business Set-Aside
- N/A
- Description
- Amendment No. 1 This Amendment No. 1 to Request for Proposal No. 277111WPK is being issued to incorporate the following clarification questions and answers into the RFP by reference. Clarification Questions and Answers Question 1 The Classical computing requirements specify that each CPU core needs 2GB of RAM and 10GB of "local disk". Compute on Google Compute Engine is consumed via virtual machines to which we can attach standard persistent disk, persistent SSD, or local SSD. The "persistent" disks are the same type of network-based storage that you would find in any virtual environment while the local SSD are direct attached. In previous conversations with Fermi and our internal high performance computing specialists, it was determined that Persistent SSD was the way to go. I think the SOW needs a clarification to indicate exactly which type of disk is being required. In a literal reading of the SOW, you could believe that the Local SSD is the required disk format and that class of storage is significantly more expensive than Persistent SSD. Answer Attachment A - Statement of Work, Classical Section, paragraph 4, change the following: Change the words "10 GB of local disk" to "10 GB of persistent SSD dick" Question 2 Similar to our concern about the Classical computing disk requirements, the Machine Learning requirements specify 10GB SSD per core. Again, it would be helpful to have this clarified to either Persistent SSD or Local SSD to ensure that Fermi doesn't receive pricing that is too high or performance that is too low. Answer Attachment A - Statement of Work, Machine Learning Section, paragraph 1, change the following: Change the words "10 GB SSD" to "10 GB persistent SSD disk" Question 3 For each of the three machine learning workloads, how many CPU's are there per workload? Answer One CPU per 8 GPU cores in use. Thus, "Development" workflows utilizing 1-2 GPU would need an associated single CPU core; "Small-scale" workflows utilizing O(10) GPUs would need 2 CPU cores; "Intermediate" workflows utilizing O(100) GPUs would be provisioned with 13 CPU cores. Question 4 For the total GPU hours, what is the ratio that makes up those GPU hours? Answer Here's the ratio year by year by workload: Yr Dev Small Intermed 2017 100% 0% 0% 2018 32% 68% 0% 2019 16% 80% 4% 2020 7% 34% 59% 2021 4% 22% 74% 2022 4% 22% 74% Note that the SOW indicates O(1 TB) data storage for Intermediate workflows, and O(10 TB) for Large workflows. We are not asking for pricing on Large workflows. Development and Small workflows are assumed to require no additional storage. Question 5 Is the firm fixed price on each burst or the resources in each burst? Answer Vendors should price out the resources within the burst which should sum to the total price per burst. Question 6 For the classical computing cycle, can you provide details on number of instances that you would like to provision. Since the RFP does not clearly state the number of instances, current requirements can be achieved by using fewer instances having many cores. You get the benefit of lower cost but availability is low. Using more instances having relatively less number of cores. You get the benefit of high availability but at higher cost. It will be helpful if you can provide what would be the optimal choice if you do not know the number of instances. Answer The optimal choice is custom-16-32768 (16 core, 32 GB of memory). Question 7 For the quantum computing cycle, can you provide additional details Approx. hours of Quantum cycle needed for each workshop Approx. support hours needed for each workshop from the quantum computing provider Answer For the Quantum Computing work, we are requesting unlimited access to the machine during the course of each of the four workshops/experiments per year. Each workshop/experiment will last one week. Google will provide hardware and expertise access to enable these experiments. Question 8 Is Fermilab asking for our list price for a single unit of the Google Cloud SKU: CP-PROF-SVC-CREDITS?" Answer This SKU is approved to use for the five year Quantum R&D engagement - so each year it will be procured using that same SKU. Additional clarification: The RFP requests pricing for an intial year with 4 optional years. The SOW document lists resource requests for years fiscal year 2017 to 2022.
- Web Link
-
FBO.gov Permalink
(https://www.fbo.gov/spg/DOE/FERMI/FERMI/277111WPK/listing.html)
- Place of Performance
- Address: Fermilab, PO Box 500, Batavia, Illinois, 60510, United States
- Zip Code: 60510
- Zip Code: 60510
- Record
- SN04668304-W 20170910/170908231443-e64b01d16a223bafe1248f6eb27b8935 (fbodaily.com)
- Source
-
FedBizOpps Link to This Notice
(may not be valid after Archive Date)
| FSG Index | This Issue's Index | Today's FBO Daily Index Page |