ESBD_384429_1734668342150_CIO-25-229 Addendum 2.pdf

THE TEXAS A&M UNIVERSITY SYSTEM

Office of Procurement & HUB Program

ADDENDUM NO. 2

Date: December 19, 2024 RFP Due Date: December 27, 2024 RE: RFP01-CIO-25-229 – AI CoE

Respondents shall acknowledge receipt of addendum(s) on the Execution of Offer (Exhibit A).

Proposal Response:

As part of Section 4.2 – Vendor Qualifications, Respondent shall provide information for each of the items below (A-D).

A. Relevant supercomputing experience (i.e., experience in selling, installing, and/or operating high performance computing systems)

B. Relevant experience with integrating on-premise supercomputers with cloud computing (i.e., the ability to scale high performance computing to include cloud computing capacity)

C. Relevant experience with institutions of higher education and specifically with the Texas A&M University System (e.g., hardware, software, services, etc.)

D. Relevant experience with selling, installing, and/or supporting NVIDIA assets (i.e., hardware, software, networking, etc.)

Questions/Responses:

1. Can the due date be extended?

Response: No, the due date remains December 27, 2024.

2. How will the proposed solution integrate with existing power and cooling infrastructure at Texas A&M?

Response: Will be determined based on the proposed system design and implementation plan.

3. What specific data security measures are required to comply with Texas A&M’s policies and relevant regulations?

Response: All purchased information resources are required to comply with the requirements expressed in the Texas A&M Information Security Controls Catalog [https://u.tamu.edu/controls-catalog], which is based on NIST 800-53 and 800-171. Specific controls that should be noted: DC-5, SA-4, RA-2

4. What are the scalability requirements for the AI infrastructure to accommodate future growth?

Response: The SuperPOD should accommodate integration into those systems whether physically co-located in the same data center location or available in a nearby facility. Additionally, SuperPOD configuration should allow integration of cloud resources from major providers or national research platform.

5. Is there a target for the number of tokens in an AI model the TAMU system is hoping to be able to handle after completing this this effort that should be validated in the solution?

Response: The targets for use of Tokens will be determined as the capacity of the system is identified.

6. If the solution needs to grow over time, how will this be handled? What provisions, if any, need to be made in the RFP response in support?

Response: See response to question 3

7. What level of technical support is expected, and what are the response times for critical issues?

Response: At a minimum, Storage and Networking infrastructure should be on 24x7 support while HPC nodes could be on 8 x 5 (NBD). Alternative proposals for where a parts depot is stored locally, and technicians arrive on-site within a 3-hour response time can be proposed. Escalation to level 2 support should be part of the proposal

8. Will storage support be handled independently from the AI compute and networking or is there an expectation for “full solution support” beyond individual hardware elements and their respective support contracts?

Response: There is an expectation for “full solution support” for storage a reference architecture that provide integrated support between vendors is desirable.

9. Support listed in the RFP is for 3 years, how will additional years of support be handled? Would A&M like to see support costs for years 4 and 5 for TCO planning?

Response: TAMUS will need support costs beyond the first three years. You may provide support costs for years 4 and 5.

10. What is the planned useful life of this infrastructure investment? How will the infrastructure lifecycle be managed in terms of expected useful life?

Response: The useful life of this infrastructure investment will be no less than 5 to 7 years

11. To what extent can the solution be customized to meet Texas A&M’s specific needs?

Response: For purposes of the selection process, customizations will not be considered

12. What performance metrics will be used to evaluate the success of the implementation?

Response:

a) Completion on time

b) Successful commissioning

c) Successful completion of benchmark testing

d) End-user validation of functionality (i.e. components work as expected)

e) 90-day health check that validates that the platform is operating at rated levels

13. What disaster recovery and business continuity plans are expected?

Response: Network & Storage can be supported on UPS and Generator power while the HPC nodes will likely not incur that additional resiliency. System recovery and configuration recovery testing must be demonstrated.

14. What are the specific requirements for network integration and storage network connectivity?

Response: All network components must meet the requirements as defined in the RFP. Where not specified in the RFP, the SuperPod reference architecture should be used:

https://docs.nvidia.com/dgx-superpod/reference-architecture/scalable-infrastructure-h200/latest/index.html

15. The RFP makes mention of 10 PB of usable capacity in the technical specifications section, 3.4, however the DDN storage exhibit provided only lists 8PB of usable capacity. Can this be clarified?

Response: Use 8PB as the target number.

https://docs.nvidia.com/dgx-superpod/reference-architecture-scalable-infrastructure-h100/latest/storage-architecture.html https://docs.nvidia.com/dgx-superpod/reference-architecture-scalable-infrastructure-h100/latest/storage-architecture.html

16. How large should the storage environment be capable of growing to?

Response: TAMUS has not determined how large the storage environment must be capable of growing to.

17. Is tiering to S3/Object/Cloud storage a requirement or of value?

Response: Yes, being able to offload storage to other storage systems, and other object storage systems is desired. We would consider a Globus.org-based solution, but we are interested in an automated and integrated method of tiering, based on age of data or other useful metrics.

18. What are Texas A&M's requirements for storage being capable of handling metadata operations without the use of separate dedicated metadata servers?

Response: Either solution will be considered based on efficiency, performance, and data management.

19. Is storage intended to be scratch or persistent?

Response: Both

20. What is the expected storage growth rate? How will it be accommodated?

Response: Unknown and to be determined

21. How will sensitive data be secured both at rest and in transit?

Response: All data storage and processing must meet the controls described in the Texas A&M Information Security Controls Catalog, DC-5 and DC-6.

22. What is the expected bandwidth and latency between the compute and storage?

Response:

a) Less than a millisecond. Bandwidth should from a single node to storage should be (40 GBps read minimum, 80 GBps preferred; 20GBps write). 4 SU aggregate system should have 500 GBps read and 250 GBps write.

b) In other words, the storage performance should meet or exceed the SuperPod reference architecture described at https://docs.nvidia.com/dgx-superpod/reference-architecture/scalable-infrastructure-h200/latest/index.html

23. Line 12 in the BoM spreadsheet indicates user provided NFS storage for home directories. Is this something that should be quoted? If so, can the specifications be shared? If it is not to be quoted, but exists, can the details be provided for design considerations and validation?

Response: Not at this time.

24. The RFP references quantity 4 Dell R760 Management servers, what are the required specifications and support levels for these elements?

Response: The reference architecture from Nvidia:

Intel based x86 2 × Socket, 24 core or greater, 384 GB RAM, OS (2x480GB M.2 or SATA/SAS SSD in RAID 1), NVME 7.68 TB (raw), 4x HDR200 VPI Ports, TPM 2.0

25. What network topology will be used (e.g. Fat Tree, Clos, Rail)?

Response: we are requesting the design and implementation be provided by the awardee and any of the examples provided would work for an AI cluster. The DGX SuperPOD reference architecture should be followed: https://docs.nvidia.com/dgx-superpod/reference-architecture/scalable-infrastructure-h200/latest/index.html

26. What are the expectations for user training sessions and ongoing support?

https://docs.nvidia.com/dgx-superpod/reference-architecture-scalable-infrastructure-h100/latest/storage-architecture.html https://docs.nvidia.com/dgx-superpod/reference-architecture-scalable-infrastructure-h100/latest/storage-architecture.html https://docs.nvidia.com/dgx-superpod/reference-architecture/scalable-infrastructure-h200/latest/index.html https://docs.nvidia.com/dgx-superpod/reference-architecture/scalable-infrastructure-h200/latest/index.html

Response: The number and quality of training sessions and ongoing support will be “sufficient” to operate the SuperPOD. As such, specific metrics are difficult to define

27. Are there any specific hardware preferences or requirements beyond those listed in the RFP?

Response: No

28. What software compatibility requirements must be met?

Response: Be able to run, at minimum, Nvidia software suites.

29. What are the requirements for data migration from existing systems to the new AI infrastructure?

Response: These requirements are not currently known and are to be determined.

30. What compliance standards must the solution adhere to (e.g., FERPA, HIPAA)?

Response: Please reference to the published TAMUS security standards and controls. The systems will support Academic and Research programs spanning all types of data sources.

31. What are the expectations for network testing, validation, and system operational validation tests?

Response: The awardee shall validate the cluster, network, and storage operates per the design and at peak load. Additionally, the cluster must be benchmarked for submission to the HPC Top500 List. This will be accomplished prior to shipping the purchased hardware to Texas A&M. TAMU is open to review options for the benchmarking validation.

32. What specific AI models and frameworks need to be supported?

Response:

a) For commissioning, the system needs to be able to successfully run

1. Grok-1 (314B mixture of experts)

b) Llama 3.1 405B (or more recent, equally sized)

33. Are there any additional storage requirements beyond the 10 PB usable capacity? (i.e. snapshots, tiering, replication, deduplication, encryption?)

Response: Must be able to expand storage to address future needs.

34. Is there a measure in IOPS that the storage platform should achieve at minimum to add clarity to the request of “full IO utilization of 3 or 4 scalable units?

Response: Refer to question #30.

35. What are Texas A&M's requirements for utilizing a POSIX compliant parallel client that can stripe data across multiple storage servers in parallel?

Response: This will be determined at a later date.

36. What are Texas A&M's requirements that the high-performance storage environment not require tuning for best performance depending on the workload?

Response: All major components should be able to be tuned to optimize performance.

37. What are Texas A&M's requirements for instantaneous snapshots for the storage environment?

Response: Recommendations on best practices will be considered

38. How is Texas A&M planning to define storage failure domains?

Response: Not defined at this time.

39. What are Texas A&M's specific performance requirements for the mixed workloads that are prevalent in AI use cases?

Response: Not defined at this time.

40. What are Texas A&M's specific requirements for small file handling?

Response: Not defined at this time.

41. If the bandwidth target for 3-4 SU can be achieved with a usable capacity less than 10PB, will a smaller usable capacity design be considered?

Response: No

42. What level of stripe size is required for proper data protection?

Response: Not defined at this time.

43. Is there a need for integration with cloud services, and if so, which ones?

Response: Possibly. Any integration would include at a minimum AWS, Azure, GCP and OCI.

44. What are the requirements for monitoring and reporting tools?

Response: Not defined at this time.

45. What job scheduling systems are preferred or required (e.g., Kubernetes, Slurm, Bright)?

Response:

a) A hybrid system that supports some nodes running interactive workloads via Kubernetes and batch workloads via Slurm is preferred.

b) If there is an Nvidia preference for Bright vs. Slurm, we’d for the bidder to indicate this.

c) Flexibility in determining the number of nodes/GPUs assigned to each scheduling system is desired, as is guidance and recommendations from the bidder about the appropriate unit of allocation to each scheduling system method.

46. What are the specific access control requirements for the AI infrastructure?

Response:

a) Access to “head-nodes” should utilize best-practices for authenticating against enterprise systems such as Active Directory or LDAP (using software like SSSD or similar).

b) Authentication to compute nodes should utilize an authentication/authorization system that is internal to the SuperPOD system if there is a potential for latency or load impact on campus enterprise services.

47. What collaboration tools are expected to be included in the solution?

Response: None

48. What is the process for software updates and maintenance?

Response: Updates to the system need to be performed consistent with the TAMU Security Controls Catalog. In particular, this policy: https://it.tamu.edu/policy/it-policy/controls-catalog/controls/CM/CM-1/

49. What are the expectations for benchmarking, such as the TOP500 HPL benchmark?

Response: See answer to question 30

50. Given its prevalence with AI specific performance benchmarking, what are Texas A&M's requirements for MLPerf scoring?

Response: We would expect the system to perform at the level that Nvidia expects It to perform.

51. What are Texas A&M's specific performance requirements for AI Training?

Response: To be determined

52. What are Texas A&M's specific performance requirements for AI Inference?

https://it.tamu.edu/policy/it-policy/controls-catalog/controls/CM/CM-1/ https://it.tamu.edu/policy/it-policy/controls-catalog/controls/CM/CM-1/

Response: To be determined

53. What is the estimated timeline for the project from start to finish to ensure TAMU project goals are met?

Response: 9-12 months

54. Will additional capabilities for model training/tuning is expected?

Response: Yes

55. Will all racks be ordered/built/shipped together or in phases?

Response: It is preferred that all racks be shipped together with at least 2 scalable units’ worth of racks be shipped initially.

56. Financing Options – Is Texas A&M interested in ownership of the equipment at the end of the term?

Response: Yes

57. What is the current total available power in the data center?

Response: There is 6 MW available assuming some upgrades occur, 1 MW without upgrades

58. What is the cooling capacity within the data center?

Response: We have 14 24-ton CRACs for a total of 336 tons.

59. Is there a requirement for room neutrality in terms of ambient air temperature?

Response: Not defined at this time.

60. Which of the following does integration with the NVIDIA NeMo framework for Large Language Models (LLMs) and generative AI support mean? (1-7)

Response: All of the below (1-7).

1. Setup and Deployment

• Environment Configuration:

Install and configure NeMo within the organization's AI stack.

Ensure compatibility with existing hardware and software (Kubernetes, Slurm).

• Integration with NVIDIA AI Enterprise:

Incorporate NeMo into the existing NVIDIA AI Enterprise stack to streamline workflows.

2. Customization for LLMs and Generative AI

• Model Fine-Tuning:

Customize pre-trained models from NeMo for specific organizational needs (domain-specific vocabulary, customer data).

• Training Pipelines:

Build and optimize training pipelines for fine-tuning LLMs or training models from scratch.

Utilize GPU acceleration and distributed training capabilities of NeMo.

• Prompt Engineering:

Design prompts and configurations tailored for generative AI use cases such as chatbots, content generation, and summarization.

3. Integration with Existing Systems

• Data Pipelines:

Connect NeMo with data sources, including structured and unstructured data (databases, logs, text files).

Implement preprocessing pipelines for cleaning and preparing data.

• API Development:

Expose LLM capabilities via APIs to integrate with customer-facing or internal applications (e.g., virtual assistants, search engines).

4. Optimization and Scalability

• Performance Tuning:

Optimize model inference for latency-sensitive applications.

Leverage NVIDIA TensorRT for inference.

• Scalability:

Enable support for multi-node training or inference using NVIDIA's high-speed networking (InfiniBand).

Scale deployment across NVIDIA DGX SuperPODs or similar AI supercomputing clusters.

5. Generative AI Features

• Text-to-Image/Video Generation:

Leverage NeMo for multi-modal applications (e.g., generating images or video from text).

• Conversational AI:

Enable advanced conversational AI capabilities with real-time LLM inference for chatbots or customer support systems.

• Speech and Language Models:

Integrate NeMo’s capabilities for speech-to-text, text-to-speech, or other natural language processing (NLP) tasks.

6. Ongoing Support

• Monitoring and Maintenance:

Monitor system performance and usage.

Update NeMo versions to leverage new features or enhancements.

7. Security and Compliance

• Role-Based Access Control:

Implement access controls for managing sensitive AI resources and outputs.

61. Deploying NVIDIA AI Enterprise software could possibly involve a suite of tools and frameworks optimized for AI workloads. Which of the following apply? (1-9)

Response: All of the below (1-9)

1. Core AI Frameworks

• NVIDIA NeMo:

Tools for developing, fine-tuning, and deploying Large Language Models (LLMs) and other generative AI applications.

• TensorFlow and PyTorch (optimized versions):

Pre-integrated, GPU-accelerated versions of the popular deep learning frameworks.

• NVIDIA RAPIDS:

GPU-accelerated data processing and machine learning libraries for large-scale data preparation and analytics.

2. Inference and Deployment Tools

• NVIDIA TensorRT:

High-performance deep learning inference optimizer for reduced latency and higher throughput.

• Triton Inference Server:

A scalable, cloud-native inference solution that supports multiple frameworks and model types.

• NVIDIA TensorFlow Serving and TorchServe:

Optimized serving tools for TensorFlow and PyTorch models.

3. Cluster Management and Orchestration

• Kubernetes Support:

Tools for containerized AI workload orchestration in multi-node environments.

• NVIDIA GPU Operator:

Simplifies deployment and management of GPU resources in Kubernetes environments.

• Slurm Integration:

Support for HPC job scheduling in distributed computing environments.

4. AI Model Training and Experimentation

• NVIDIA Merlin:

Recommendation system framework for training and deploying high-performance recommendation engines.

• NVIDIA Clara:

Frameworks for AI applications in healthcare and life sciences.

• NVIDIA Riva:

Toolkit for building real-time speech AI applications (e.g., ASR, TTS).

5. Data Processing and Analytics

• Dask GPU:

Parallelized data processing framework integrated with RAPIDS.

• NVIDIA NVTabular:

Accelerated tabular data preprocessing for recommender systems and other use cases.

6. Monitoring and Visualization

• NVIDIA Nsight:

Profiling and debugging tools for CUDA applications and AI workloads.

• NVIDIA System Management Interface (nvidia-smi):

Monitoring GPU utilization, memory consumption, and application performance.

• Prometheus and Grafana:

Pre-integrated solutions for performance monitoring and visualization.

7. Multi-Modal and Generative AI

• NVIDIA Omniverse:

Platform for real-time collaboration and simulation, particularly for 3D design and digital twins.

• NVIDIA GauGAN and StyleGAN:

Pre-built generative AI models for creative applications.

8. Security and Compliance

• MIG (Multi-Instance GPU):

Partition GPUs into multiple isolated instances for improved multi-tenancy.

• NVIDIA AI Enterprise Security Features:

Encryption and role-based access control for secure AI deployment.

9. Pre-Trained Models and Workflows

• Hugging Face Transformers (optimized):

Integrated access to pre-trained transformer models for NLP.

• NVIDIA TAO Toolkit:

Transfer learning toolkit for customizing pre-trained models with minimal coding.

62. Can we provide a response to RFP01-CIO-25-229 for just the 92031-*Installation Of Computers, Peripherals, And Related Equipment (Including Software) portion?

Response: No

63. Where to ship the hardware? NVIDIA site? staging location? The Datacenter?

Response: Texas A&M University, College Station, Texas. Exact location still to be determined.

64. What are the expectations/deliverables regarding training sessions?

Response: We expect the training to enable researchers and SuperPOD administrators to maximize use of the cluster and to include reference diagrams and materials that can be shared with the research teams

65. What is the storage performance criteria for 1k+ GPUs, do all 10PB need to be All-NVMe?

Response: Refer to the reference architecture: https://docs.nvidia.com/dgx-superpod/reference-architecture/scalable-infrastructure-h200/latest/index.html

66. How is storage selection weighted (performance per RU, Feature set and flexibility, simplicity, Cost/performance)? Can you please provide guidance on these storage selection criteria?

Response: To be determined.

67. Are any particular features required from the storage platform (e.g. - Snapshots, S3 protocol, HDFS support, cloud hooks, etc.)?

Response: Performance is the primary driver for the storage system. Enterprise management features that do not compromise performance would be of interest. Please describe clearly what is proposed.

68. Is there a need for data migration from an existing file system at Texas A&M?

Response: Yes

69. Is Texas A&M open to proposals leveraging the same SuperPOD "rail-optimized" architecture but on their own Server hardware (e.g., Lenovo SR680a, Cray XD670, etc.)?

Response: No

70. If TAMUS is open to HGX solutions (with similar services such as PTAM), is A&M open to a liquid-cooled solution, and will the data center support that?

Response: Liquid cooled is not planned at this time

71. How much data center power can be supported per cabinet today?

Response: Current configuration is 35kW per rack.

72. Is the bidder also providing Rack/PDU infrastructure?

https://docs.nvidia.com/dgx-superpod/reference-architecture-scalable-infrastructure-h100/latest/storage-architecture.html https://docs.nvidia.com/dgx-superpod/reference-architecture-scalable-infrastructure-h100/latest/storage-architecture.html

If so, is A&M open to liquid-assisted cooling via Rear-door Heat exchange to improve density and thermals per cabinet?

Response: Yes, the bidder provides the racks, PDUs, cable management, cabling for IB Network, and NVMe storage connectivity.

73. Is there a preference for single-mode fiber over MMF?

Response: Single mode is preferred unless there is a multi-mode benefit that can be articulated.

74. Will Texas A&M leverage the full SuperPOD cluster management suite with BCM and Slurm integration?

Response: To be determined.

If not, will the university require the solution provider to assist with integration into the existing scheduling and cluster management framework?

75. Will SLURM be the default scheduler or is A&M interested in leveraging K8s in conjunction with something like Run:ai as a scheduler and workload orchestrator?

Response: We are interested in hearing more about options for scheduling and orchestrator options. TAMU already has experience with SLURM, so that should be considered, but should not be the final decision without further discussion.

76. Has the System evaluated the current infrastructure that will be supporting these SuperPods to ensure adequate power and cooling requirements are met?

Response: In process

If so, will the System provide those reports?

Response: Not at this time.

Will the selected respondent be allowed to validate all power and cooling requirements and system configurations to ensure compliance with OEM specifications?

Response: Yes

77. Has there been any initial site evaluation and or Data Center Infrastructure assessment for this project?

Response: Not at this time.

If so, will A&M System provide these reports and or evaluations?

If not, will the System supply any current limitations within the Data Center(s) that are to house these SuperPods? (number of racks per rows, hot aisle/cool aisle methodology, doorway height, etc).

78. Will all SuperPods be deployed within the same Data Center or will multiple sites be utilized?

Response: The SuperPOD will be deployed in a single data center.

Will the System provide any facility readiness reports specific to the locations that will host these systems?

Response: Not at this time.

79. How does The System envision the partnership between the respondent and Nvidia in terms of roles and responsibilities?

Specifically, since this is a SuperPod and Nvidia will perform the physical installation and logical configuration, what assistance does The System expect from the respondent regarding these activities?

Response: We expect both parties to actively participate in the design, implementation and support of the environment.

80. For the requested post-deployment on-site technical assistance, does The System wish to have a long-term, full-time resource staffed by the respondent to administer the SuperPod environment?

Response: Not at this time.

81. Does the System have an estimated timeframe of when implementation will begin?

Response: June 2025

82. Will the System provide detailed specifications for the Dell Management Server (INTEL-MGMT- SERVER) included in the provided spreadsheet?

Response: Refer to question #23.

83. Will you be evaluating and or requiring additional security solutions to support this deployment?

Response: Potentially there may be requirements associated with Federal CMMC compliance.

84. Per Section 3.3. Scope of Work: The selected respondent will be responsible for, “offering local, on-site technical assistance during and after deployment.” Would TAMUS please specify examples of work to be performed as part of technical assistance as part of the AI COE over the three years. This will allow respondents to appropriately estimate costs for these services.

Response: We would expect engineering support of new integrations with new technologies that may need to be implemented.

85. Per Section 3.4 and 3.5: All technical hands-on implementation work for standing up the AI cluster will be performed by NVIDIA. Can TAMU clarify if only technical PMO services are desired of the vendor or are there any implementation specific services also desired of the vendor?

Response: Expecting them to be integrated into the installation and configuration on site so that they are familiar with the internal workings of the environment

86. Per Section 3.6 and 3.7: Are the Accessibility requirements and TX-RAMP certifications provided by the OEM (Nvidia?) or are there submission requirements for the implementation partner as well?

Response: Accessibility requirements and TX-RAMP certifications will be the responsibility of the

OEM.

87. Is there public facing info on the data center specs that you can share (e.g., total MW, floor space, cooling types, layout, loading and staging areas, etc.)?

Response: There are no public facing specifications. Only the information contained in this RFP.

88. What is the max power per rack? (Assuming all racks are dual power capable.)

Response: Currently, 35KW

89. Does the DC currently meet a specific certification level?

Response: Not at this time.

90. Will this environment need to meet a different certification level? If so, will the environment be sectioned off/caged? Will additional security need to be implemented?

Response: At this time, the final decision on the need to section off/cage part of the environment has not been finalized. It is a possibility.

91. What are the delivery and staging capabilities of the DC?

Response: The data center has two loading docks for receiving. A 10,000 lb lift to get equipment to the raised floor or to an upstairs secured storage area. There is a staging/ burn-in room available, and additional staging area could be provided in the data hall.

92. Does the BOM account for any 'spare parts' to have on site in case of failure or damaged items in shipping or if they are faulty for any other reason? If so, is there a desired % of spare equipment needed?

Response: A spare parts depot onsite would be entertained at no cost to the Texas A&M and included in the maintenance & support agreements. All options will be entertained as the onsite depot is not a hard requirement.

93. There are 3 NVIDIA resources listed in the BOM, but they are all listed as remote; will NVIDIA be onsite for the deployment?

Response: Yes

94. Will NVIDIA have a project manager assigned, or will that be the Role of the TAMU System?

Response: NVIDIA will have a project manager assigned as will TAMUS.

95. Is there any cooling or electrical work expected to be performed outside of connecting to and powering the proposed gear?

Response: Electrical or cooling work will be done in preparation of receiving the cluster.

96. Is TAMU wanting managed support/services for the additional storage requested beyond normal the hardware warranty?

Response: Potentially, depends on the normal hardware warranty period.

97. Does TAMU want the respondent to have resources available beyond the deployment date, e.g., a storage engineer for 3 years of the contract, or just for the duration of the deployment?

Response: It would be desirable for the respondent to have resources available beyond the deployment date. That is part of the value-add that a respondent can provide to differentiate itself from others.

98. Can you give us more detail on the scope of the "assisting" and "supporting" in SOW?

Response: The respondent’s proposal of additional assistance/support is part of the value-add that a respondent can provide to differentiate itself from others.

99. Is there a proposed RACI for A&M/NVIDIA/Respondent?

Response: Not at this time.

100. For the Value-Add Services should those be added as line items at the bottom of the BOM?

Response: Yes

101. Is TAMU wanting managed support/services for the management nodes requested beyond normal hardware warranty?

Response: Yes

102. The BOM lists 8 PB useable storage for the DNN storage solution (customer supplied) but then also lists customer provided NFS (undetermined amount) supplied storage for home directories. Is the 10 PB listed in sections 3.4.1 and 3.4.2 total for both line items? Or is 10 PB required for each line item (total of 20 PB)?

Response: See question 14

103. Does TAMU want the respondent to receive the NVIDIA gear offsite, build out the racks, and then deliver the solution shrink-wrapped to the Data Center?

Response: See answer to question 91. If the response affords an onsite staging, that becomes an option. Preference is racked offsite for a shrink-wrapped delivery.

104. Do the counts in the BOM for the Legrand NVIDPD13 racks, cable, and PDU take into account the customer supplied storage solution and user directory storage requested? Or are those expected to come with their own racks, cable, and PDUs?

Response: TAMU can provide racks; PDUs and cabling for the storage needs to be provided.

105. The RFP requires TX-Ramp certification, as this is a NVIDIA managed service, is TAMU looking to the respondent to assist NVIDIA/TAMU with getting the final built solution TX-Ramp certified? If so, what part of the TX-Ramp requirements does TAMU feel this solution falls under?

Response: Accessibility requirements and TX-RAMP certifications will be the responsibility of the

OEM.

106. Will the NVIDA solution and any parts of it need to integrate with Active Directory?

Response: Active Directory is one available option available for authentication. If the Nvidia software specifies Active Directory to operate, then the TAMU Active Directory domain will be used

107. Is there any need to integrate this with directory services or security tooling with the university system?

Response: To allow for the campus identity to access and use services that will be provided to the resource

108. It is unclear in the NVIDIA informational section indicating they will do all networking, is this relative to integrating into the systems network? Can you provide details on network integration tasks and testing requirements?

Response: Not beyond what is in the RFP

109. Will the Respondent have access to raw, unencrypted student data?

Response: The Respondent will not have access to raw unencrypted student data.

110. Are the 3U and 4U clusters to be co-located - i.e. in the same datacenter? on the same floor? Or in disparate locations?

Response: Per section 4.4 A., the Respondent shall provide pricing on both the 3SU and 4SU for consideration by the A&M System. For clarification, A&M System will only move forward with one of those options, not both.

111. Do you accept equal solutions from Supermicro?

Response: No

112. Do you accept equal solutions for DDN storage?

Response: Yes

113. Do you accept any site visit?

Response: No

114. How do respondents meet the TX-RAMP Certification requirement if the respondent is not providing cloud services?

Response: The respondent is not required to meet TX-RAMP certification if they are not providing cloud services.

115. The proposal states a storage requirement of 10PB usable, but the Build of Materials states a requirement of 8PB. Which is correct?

Response: See question #14. Use 8PB as the target number.

116. Which is correct, the 95+127 from the spreadsheet or the 760 from the RFP document?

Response: The number of 760 is correct for the minimum number of GPUs as stated in the second bullet of section 3.4.1. Note also bullet #1 which states “Deployment of Nvidia DGX SuperPOD with DGX H200 Systems”. Therefore, a 3SU DGX H200 SuperPOD has 95 DGXs x 8 GPUs = 760 H200 GPUs. And the 4SU DGX H200 SuperPOD has 127 DGXs x 8 GPUs = 1,016 H200 GPUs.

END OF ADDENDUM NO. 2