
AMD Radeon for AI
Use a representative workload to assess your model, GPU memory needs and execution performance.
AI infrastructure · Comino GRANDO
From deskside workstations to rack-mounted multi-GPU systems, Comino GRANDO is built around a sealed liquid loop. EudTech reviews your workload, noise, space, power and cooling needs to prepare a configuration and quotation.


Comino Test Drive · Free remote evaluation
AMD ROCm™W7900 · AI PRO R9700
Apply through Comino’s official Test Drive program to remotely evaluate AMD Radeon PRO W7900, Radeon AI PRO R9700 and ROCm for your AI or HPC workload’s compatibility and performance.

Use a representative workload to assess your model, GPU memory needs and execution performance.

Explore libraries, development tools and runtimes for your AI and HPC software workflow.
Comino GRANDO · AMD Radeon

4U · Rackable workstation
AMD EPYC 9004/9005 · 64 cores
Comino liquid cooling
View manufacturer configuration
4U · Desktop workstation
AMD Threadripper PRO 7985WX
Comino liquid cooling
View manufacturer configuration
4U · Rackable workstation
AMD EPYC 9004/9005 · 64 cores
Comino liquid cooling
View manufacturer configuration
4U · Desktop workstation
AMD Threadripper PRO 7985WX
Comino liquid cooling
View manufacturer configurationConfigurations and chassis images are from Comino’s program page; test resources require Comino’s confirmation. GPU memory is summed across cards; usable capacity depends on multi-GPU software support.
Describe your use case, software and evaluation goals in the official form.
Qualifying applicants receive connection instructions from Comino by email.
After evaluation, discuss GPU configuration, site requirements and a quotation with EudTech.
Evaluation only; commercial use of the free computing time is prohibited. Comino may revoke remote access at any time. Eligibility, configuration and availability are subject to Comino’s confirmation.
Program and imagery: CominoStart with the site
A system needs more than floor or rack space. Heat, noise, power and handling affect where it can operate and how it will be maintained.
Review acoustics and distance from users before choosing deskside or rack deployment.
Consider power, heat removal, handling and service access together.
Short development sessions and sustained loads place different demands on cooling and operations.
Inside the cooling design
Coolant and airflow have different roles. Heat travels from high-load components through the coolant, then leaves the chassis radiator for the surrounding environment.

Cooling blocks transfer heat from high-load components, reducing the space required for individual GPU heatsinks.
A sealed liquid loop circulates coolant inside the system and carries heat to the chassis radiator. Blocks, pipework and chassis are designed as one system.
Airflow removes heat through the radiator, while intake air also cools components such as RAM.
No cooling tower does not mean no room cooling. The sealed-liquid-loop complete-system design described here does not require facility water or an external cooling tower. Heat still enters the room, so air conditioning, ventilation and power must be checked.
Two deployment priorities
Systems from the same family do not all have the same noise level. GPU power, GPU count, CPU choice and intake temperature must be reviewed together.

Deskside / lab
Review acoustics, room cooling, available space and distance from users before choosing the GPUs and system configuration.
Manufacturer workstation example; not every GPU configuration is suited to an office.

Rack / shared compute
Review power, heat removal, GPU density and service access. Confirm remote management and redundant power for the selected model and load.
Manufacturer rackable chassis example; GPU count and platform compatibility require review.
This Sales Kit 0911 matrix (11 September 2025) illustrates acoustic and environmental tradeoffs by GPU power, count and CPU platform. It is not a live compatibility list. Grey means “not presented”, not incompatible.

Start with your workload
GPU count is a configuration choice, not a software-performance guarantee. Understand how the software and data use the GPUs, then bring site requirements into the decision.
Start with the model, precision, concurrent users and response-time targets, then size GPU memory, GPU count, RAM and storage.
Bring: models, software and expected usage
Review software licensing and support, CPU/GPU roles, precision, inter-GPU data movement and sustained run time.
Bring: software versions and a representative workload
Plan compute, storage and networking around scenes, image workflows, GPU memory, asset volume and expansion needs.
Bring: scene scale, tools and output requirements
Evaluate H200, RTX PRO 6000 Blackwell and GeForce RTX 5090 within the complete system; a GPU model alone does not determine the deployment site. Review CPU platform, inter-GPU communication, memory and expansion against the workload.
Everyday operation
Manufacturer documentation shows cooling-system monitoring, quick-disconnect couplings and different power designs. EudTech confirms monitoring, maintenance responsibilities, warranty and support for the selected system.

Review temperatures, fans, pumps and cooling-system status. Monitoring and integration depend on the model.
Quick-disconnect couplings (QDC) support servicing under manufacturer procedures; they do not make every component hot-swappable.
Confirm power capacity and redundancy against the model, load and input voltage. Multiple supplies do not automatically provide redundant capacity.
Enterprise & government procurement · Compliance & trust
Understand Comino’s stated compliance and manufacturing advantages, then tie procurement requirements to the selected model, configuration and documentation. EudTech helps organise requirements and manufacturer information for an evidence-based review.
NASA SEWP’s public catalog lists the Comino RM V2S D9612 configuration as TAA compliant; see the record below. Review the actual supplied model, country of origin and compliance documentation for each purchase.
Evidence to confirm: model-specific origin and TAA declarations.
According to the manufacturer’s presentation, core closed-loop liquid-cooling modules and system integration are completed in the EU. This does not establish EU origin for every component; system and key-component origins require separate confirmation.
Evidence to confirm: manufacturing and integration locations, system origin and key-component list.
For enterprise or government purchases with sourcing restrictions, review restricted sources, key components and traceability during configuration. Confirm against the actual BOM and supplier evidence rather than assuming compliance from the brand or assembly location.
Evidence to confirm: BOM, sourcing-restriction review and component-substitution terms.
For security-sensitive environments, define management interfaces, access permissions, firmware updates, maintenance responsibilities and acceptance checks. Suitability must be assessed against the organisation’s security policy for each deployment.
Evidence to confirm: security requirements, support scope and acceptance records.
TAA compliance does not mean every component is made outside China, and it is not a cybersecurity certification. Manufacturer presentation statements do not replace declarations for a specific configuration or establish eligibility for every government tender. Confirm supply and compliance against the procurement terms and formal documentation.
Publicly verifiable procurement record
NASA’s U.S. government IT procurement catalog lists this Comino product through Blue Tech. The record below is transcribed from its public lookup.
Verified on
Comino Holding Ltd.
Liquid-cooled server platform
TAA field
CListed as compliant
Applies to the specific configuration in this catalog line item.

This is the product status reported by the contract holder in NASA SEWP, rather than a NASA-issued certificate. Other Comino/GRANDO models and custom configurations require their own origin and TAA documentation.
Lookup path: search Comino → Comino Holding Ltd. → Blue Tech → product item.
Listed configuration: Threadripper 3970X, 32 GB DDR4, 2 TB NVMe, one NVIDIA RTX 3080 and a 4U enclosure. This earlier-generation configuration illustrates the verifiable record; current requirements need separate sizing.
These fields describe separate statuses. EPEAT’s NC and SCRM’s N do not change this record’s TAA field of C.
Market research and requirements-justification reference
For users, IT and procurement teams: start with cooling, service, monitoring and site requirements, then review the actual build and manufacturer documentation.
Documents reviewed:
Traditional Chinese editions translated by EudTech are available in Chinese mode. They retain the original figures, values and page numbering; the English manufacturer originals prevail in case of discrepancy. Translation date:
v2.3 · Original manufacturer PDF in English · 14 pages
Cooling coverage, service, monitoring, power, 4U chassis and environmental conditions for functional and deployment review.
v2.0.2 · Original manufacturer PDF in English · 2 pages
Installation orientation, startup, cooling profiles, coolant maintenance and power-loss recovery settings for deployment and training.
The entries below identify model scope from Comino's public URLs and indexed text. The Chinese reference editions translate verifiable key fields and retain the original URLs. Obtain verifiable originals and cited reports from the manufacturer for procurement.
The Traditional Chinese reference editions were prepared by EudTech from publicly indexed manufacturer text. They are not Comino-issued documents or page, signature, or layout replicas of the source PDFs. The manufacturer host refused connections on 2026-10-04, so no source-file SHA-256 is available. A verifiable English original prevails.
CE · 2024-05
Model: RM-M
Submodels: RM-M-V1 · RM-M-V2
The manufacturer self-declaration covers RM-M-V1/V2 and cites EMC report LEITC-TR-23-161. The safety-test report number field is blank in the indexed source text.
CE · 2024-05
Model: RM-M-CRPS
Submodels: RM-M-CRPS-V1 · RM-M-CRPS-V2 · RM-M-CRPS-DUALBLADE
The manufacturer self-declaration covers RM-M-CRPS-V1/V2/DUALBLADE and cites EMC report LEITC-TR-23-163 and safety report 004TP24. Obtain copies of the cited reports when procurement requires them.
RoHS / REACH · 2024-05
Model: RM-M
Submodels: RM-M-V1 · RM-M-V2
A manufacturer RoHS/REACH self-declaration for RM-M-V1/V2, not a third-party test certificate. Review the final model, BOM, document date, and supply-chain evidence.
RoHS / REACH · 2024-05
Model: RM-M-CRPS
Submodels: RM-M-CRPS-V1 · RM-M-CRPS-V2 · RM-M-CRPS-DUALBLADE
A manufacturer RoHS/REACH self-declaration for RM-M-CRPS-V1/V2/DUALBLADE, not a third-party test certificate. Review the final model, BOM, document date, and supply-chain evidence.
FCC evidence boundary: As of 2026-10-04, no model-specific FCC declaration, authorization, or test report for GRANDO RM-M or RM-M-CRPS was verified in Comino's public official sources. If FCC evidence is required, obtain manufacturer evidence matching the final model and BOM. A legacy COMINO N1 declaration does not evidence GRANDO.
How to form public-procurement requirements
Product architecture evidence and needs assessment
Coolant circulates within sealed pipework, with an internal radiator and fans transferring heat to ambient air. Comino identifies this design for its eight-GPU RTX PRO 6000 GRANDO, without facility water connections; verify other models against their manufacturer documents.
Because the deployment site has no facility-water interface and cannot add such pipework, the offered complete system must not require a facility-water connection. The supplier shall describe the complete cooling architecture, system boundary, cooling of the CPU, GPU, GPU memory and power components, room heat rejection and maintenance procedures. At the agency-defined ambient temperature and workload, the system shall demonstrate sustained operation, no unintended throttling and no coolant leakage. Cooling technology and trade names are unrestricted; any solution achieving the same functional, site and acceptance outcomes may be offered.
Acceptance: match the supplied model and coolant layout, confirm independence from external water circuits, and record load temperatures and leak checks. Follow manufacturer maintenance procedures; QDC does not make all components hot-swappable. Room cooling and ventilation remain necessary. These self-contained requirements do not describe Integration Kits requiring external cooling equipment.
Manufacturer evidence: sealed liquid loop and maintainable loopEight-GPU market information and needs justification
The Comino eight-GPU RTX PRO 6000 complete system demonstrates that such a configuration exists in the market. It does not establish that every agency needs eight GPUs, 96GB per GPU or a particular cooling architecture. Necessity should first be tested against the model, precision, concurrency, completion deadline, inter-GPU communication and site conditions. If fewer GPUs, a different per-GPU capacity or another cooling approach can achieve the same purpose, those solutions should remain eligible.
| Configuration | Status and required confirmation |
|---|---|
| 8 × NVIDIA RTX PRO 6000 Blackwell Server Edition 96GB | Manufacturer-standard eight-GPU sealed-liquid-loop complete-system reference, configurator #23: 96GB per GPU, 768GB total. Verify the quoted model, BOM and manufacturer documents for each offer. Manufacturer eight-GPU product page |
| RTX A6000 / RTX 6000 Ada / RTX PRO 6000 Blackwell / A40 / L40S / A100 / H100 / H200 | GPU/accelerator families listed in the GRANDO Server datasheet. An up-to-eight-GPU configuration requires manufacturer confirmation of the motherboard, cold plates, power and cooling for the offered build; the family list does not establish every model as a confirmed standard eight-GPU complete system. |
| 8 × GeForce RTX 5090 | Evaluate as a project configuration and obtain written manufacturer confirmation of the complete eight-GPU system and cooling architecture before quotation. |
| Radeon PRO W7900 / Radeon AI PRO R9700 | Published presets include a six-GPU R9700 workstation; W7900 is also a workstation option. These must not be treated as standard eight-GPU GRANDO Server options; confirm GPU count and complete-system architecture separately. |
| 8 × H200 Integration Kit | An Integration Kit has a different system boundary from a self-contained sealed-liquid-loop complete system and may require external cooling equipment. Compare them under documented site conditions and the same acceptance thresholds. H200 availability and lead time both require current written manufacturer confirmation. |
In the agency-defined software environment, the offered system shall complete the agreed training or inference workload using the specified model, precision, input data, concurrency and test duration, and meet the published thresholds for throughput, latency, completion time, stability and error rate. The supplier shall provide the complete-system model and BOM, accelerator count and memory per device, interconnect method, power draw, cooling architecture and site requirements. Accelerator brand, count and cooling technology are unrestricted; any solution achieving the same function, outcome, compatibility, site constraints and acceptance results may be offered.
The agency’s needs analysis has established that the specified workload requires eight independent GPU execution units concurrently and that each unit must hold a model shard or data set larger than lower-capacity alternatives can accommodate. The requirement is therefore at least eight GPUs with at least 96GB of usable memory per GPU. The agency should retain the workload, capacity calculation, multi-GPU allocation method and comparison of alternatives in its need record, and validate every bid with the same test. A different architecture achieving the same function and outcome should be assessed under the published equivalence criteria.
768GB is the sum of memory across eight GPUs, not a single shared memory pool or a guarantee that any model can use the full capacity directly. Specify workload, framework compatibility and inter-GPU communication requirements separately, and verify them by testing. Internal heat rejection still releases heat into the room; assess room cooling, ventilation and power capacity together.
| No. | Requirement | Suggested review | Evidence and acceptance |
|---|---|---|---|
| 1 | Sealed liquid loop, cooling coverage and sustained compute | Define sustained-compute, intake-temperature, site-interface, acoustic and maintenance outcomes first. Require a self-contained liquid-cooling architecture only when supported by needs justification. A sealed liquid loop may be identified as a market solution while allowing other architectures that meet the same functional and acceptance criteria. | Cooling layout, component temperatures, power, throughput and throttling records.Manufacturer document · pp. 4, 8, 10 |
| 2 | Serviceable liquid cooling | Provide procedures for component replacement, coolant handling, leak prevention and return to service. Demonstrate service time when required. | Service manual, coupling demonstration and checks after replacement. QDC does not make every component hot-swappable.Manufacturer document · pp. 4 |
| 3 | Monitoring and event records | List monitored and logged temperatures, fans, pumps and cooling states, with standard or optional features and alert methods. | Monitoring views, events and history. Confirm optional coolant-level sensing; review security functions separately.Manufacturer document · pp. 2 |
| 4 | Power redundancy at the specified load | When redundancy is required, verify that essential operation continues after one power-module failure at the specified CPU/GPU load. | Input voltage, available-power calculations and a failure test. 4+0 has no redundancy; size 3+1 or 2+2 for the load.Manufacturer document · pp. 10 |
| 5 | Site, density and acoustics | Set space, GPU capacity and acoustic requirements from the rack, power, room cooling and user distance. Allow service access. | Complete BOM, dimensions, weight, heat removal and acoustic test conditions. Platform maxima are not guarantees for every build.Manufacturer document · pp. 8–12 |
| 6 | Conformity evidence matching the model | Request safety, electromagnetic-compatibility and restricted-substance evidence matching the supplied model and configuration under applicable requirements. | Review model, date, scope and referenced reports in applicable CE, FCC, RoHS or REACH documents; confirm Taiwan product-inspection requirements separately. |
Manufacturer PDFs retain their original wording; the summary clarifies conditions for citation. Performance, energy-saving and temperature claims for retrofit kits require the complete modified system, external cooling equipment and test conditions.
Datasheets and guides are technical information. Review CE/FCC conformity declarations, RoHS/REACH declarations and relevant test reports against the confirmed model and BOM. Declarations and test reports serve different purposes; do not assume every GPU build has undergone the same tests.
For public market research and requirements justification. For procurements at or above the publication threshold, tender documents should be based on actual functions or benefits, with the necessity of special or stricter thresholds documented and acceptance measured objectively. Competitive restriction is assessed from necessity and from the purpose and effect of the specification, not supplier count alone; the words ‘or equivalent’ do not cure a threshold that lacks needs justification. Government Procurement Act, Article 26
EudTech selection guidance based on manufacturer documents. Review each cited page together with the actual model and BOM.
Not automatically. A 4+0 arrangement has no redundancy. A 3+1 or 2+2 arrangement still requires checking input voltage, available power per module and total system load. If redundancy is required, verify essential operation after one power-module failure at the agreed CPU/GPU load.
No. The manufacturer datasheet gives a platform reference range of 39–70dB. Confirm the selected model, GPU/CPU configuration, intake temperature, workload and measurement method before agreeing acoustic acceptance criteria. For an office or lab, also assess the distance between the system and its users.
QDC alone does not establish GPU hot-swap support. It is a liquid-cooling service interface. Replacement must follow the manufacturer's shutdown, coolant-handling and return-to-service procedures. Any requirement for maintenance without shutdown needs explicit support for the complete system and affected components.
No. These are manufacturer technical documents and EudTech's Traditional Chinese translations, not manufacturer-issued conformity declarations or test reports. Review applicable CE/FCC documents separately against the supplied model, BOM, date and scope. A datasheet does not establish that every GPU configuration underwent the same tests.
No. 768GB is the sum of physical memory across eight cards, not a guarantee that one application can use it as a single pool. Specify the model, precision, context length, concurrency and software environment, then validate multi-GPU allocation with the actual workload. Check each GPU model, count and capacity at acceptance.
No. Internal radiators still reject system heat into the room. Size room cooling and ventilation for the offered configuration’s power draw and heat output. Require the complete system cooling architecture and site conditions, then verify stable operation at the agreed ambient temperature and load. If documented site constraints require operation without external cooling equipment, an integration kit that requires such equipment does not meet that requirement. Without that justified constraint, compare solutions under the same workload, site conditions and acceptance thresholds.
Comino × EudTech
Comino provides the hardware and liquid-cooling system design. EudTech handles local requirements, configuration review and delivery planning so users, IT and procurement work from the same information.
Record models, software, users, data scale and run time, then document GPU, CPU, RAM, storage and networking choices.
Check racks, power, heat removal, noise, handling and networking. Keep a shareable configuration and version for users, IT and procurement.
Confirm supply, installation and documentation, and agree hardware inventory, stress tests, thermals, error logs and workload acceptance.
Define warranty, maintenance responsibilities, training and support. Check origin and procurement requirements against the selected model and manufacturer documents.
Actual supply, installation, testing and support are defined by the formal quotation and agreed scope.
Plan your next step
Use the configurator if you already have a specification. Otherwise, share your software, GPU requirements, run time, site and budget range. EudTech helps identify what needs to be confirmed next.
Share your existing hardware, chassis and use case. Compatibility, modification scope and warranty implications must be reviewed before proposing a retrofit.
Discuss existing hardware