IMPOSSIBLE PC/ULTIMATE PC

🔥 IMPOSSIBLE PC — FINAL MASTER COMPONENT & ENGINEERING LIST

This is the fully consolidated final specification for the machine: the compute platform, six-GPU architecture, 512GB physical GPU memory, 144.64TB SSD subsystem, 5kW PSU system, 15kW-class cooling plant, UPS, custom chassis, networking, monitoring, AI software and the supporting engineering hardware. I also incorporated the corrections from the last compatibility pass rather than carrying forward known mismatches. The WRX90E-SAGE SE provides 8-channel ECC RDIMM memory, six PCIe 5.0 x16 slots plus one x8-mode slot, four PCIe 5.0 M.2 interfaces, two PCIe 4.0 x4 SlimSAS NVMe interfaces and four SATA ports. (ASUS Global)

🔥 IMPOSSIBLE PC — FINAL MASTER COMPONENT & ENGINEERING LIST

#	Component / Subsystem	🏆 Exact Final Choice / Specification	Qty	Problem	Final Status
1	🧠 CPU	AMD Ryzen Threadripper PRO 9995WX — 96C/192T, 350W, sTR5	1	None	🟢
2	🧩 Motherboard	ASUS Pro WS WRX90E-SAGE SE — WRX90, sTR5	1	None	🟢
3	🔧 BIOS	BIOS 1106 or newer for 9995WX support	1	None	🟢
4	🧬 System RAM	128GB DDR5 ECC RDIMM, ASUS-QVL-qualified exact P/N	8	Exact QVL P/N must be verified at procurement	🟡
5	🧬 Total RAM	1TB ECC DDR5 RDIMM, 8-channel populated	—	None	🟢
6	🧬 RAM speed target	DDR5-6400 MT/s, highest stable QVL-supported speed	—	6400 is target, not an unconditional guarantee	🟡
7	🤖 AI GPU #1	NVIDIA RTX PRO 6000 Blackwell Workstation Edition — 96GB ECC GDDR7, 600W	1	None	🟢
8	🤖 AI GPU #2	Same	1	None	🟢
9	🤖 AI GPU #3	Same	1	None	🟢
10	🤖 AI GPU #4	Same	1	None	🟢
11	🤖 AI GPU #5	Same	1	None	🟢
12	🎮 Gaming / Creator GPU	NVIDIA GeForce RTX 5090 Founders Edition — 32GB GDDR7	1	None	🟢
13	🧠 Aggregate AI VRAM	480GB across 5× RTX PRO 6000	—	VRAM remains distributed across GPUs	🟢
14	🧠 Total physical GPU VRAM	512GB = 480GB AI + 32GB 5090	—	Not one unified hardware memory pool	🟢
15	🔗 GPU topology	6× direct PCIe 5.0 x16 CPU-connected links	—	Six-GPU configuration exceeds ASUS's published VGA validation	🟡
16	🔗 GPU slot mapping	Slots 1,2,3,4,5,7 occupied; slot 6/x8 unused	—	Requires custom external GPU mounting	🟢
17	🔌 PCIe risers	Cooler Master PCIe 5.0 x16 300mm — MCA-U000C-KPCI50-300	6	Six-cable Gen5 stability must be physically tested	🟡
18	🏗️ GPU physical mounting	Custom six-bay reinforced GPU deck	1 system	Custom fabrication required	🟡
19	🤖 PRO 6000 water block	Bykski N-RTXPRO6000-WS-SR — Workstation Edition specific	5	Exact GPU PCB/SKU must match	🟢
20	🎮 RTX 5090 water block	EK-Quantum Vector³ FE RTX 5090 – Plexi	1	Must use 5090 Founders Edition	🟢
21	🧠 CPU water block	Watercool HEATKILLER IV PRO for Threadripper Copper/Nickel — MPN 18026	1	None	🟢
22	⚡ PSU #1	FSP CANNON PRO 2500W — FSP2500-57APB	1	None	🟢
23	⚡ PSU #2	FSP CANNON PRO 2500W — FSP2500-57APB, identical	1	None	🟢
24	⚡ Combined PSU capacity	5,000W	—	Actual final consumption must be measured	🟢
25	🔌 GPU power	1 dedicated native 12V-2x6 cable per GPU	6	None	🟢
26	🔌 PSU reserve connectors	Additional 12V-2x6 outputs left unused	—	None	🟢
27	🔌 Dual-PSU control	ASUS documented dual-PSU architecture, identical PSU models	1 system	Must follow ASUS wiring exactly	🟢
28	❄️ Cooling plant	Koolance ALX-15KA — 15,000W cooling capacity, up to 32L/min	1	None	🟢
29	💧 Primary pump	Koolance PMP-600	1	None	🟢
30	💧 Redundant pump	Koolance PMP-600	1	Requires proper failover plumbing	🟡
31	💧 Pump topology	Parallel redundant pumps with check valves + isolation	1 system	Custom plumbing	🟡
32	🌡️ Cooling flow monitoring	ALX-15KA integrated flow monitoring	1 system	External INS-FM18 removed from control path	🟢
33	💧 GPU QDCs	Koolance QDT4 high-flow 13mm-ID system	1 system	Exact quantity depends on branch layout	🟢
34	🧪 Coolant	Koolance LIQ-705	System-fill	None	🟢
35	🔩 Coolant manifold	Custom CNC copper/brass/nickel 7-branch high-flow manifold	1	Fabrication + pressure test required	🟡
36	💧 Cooling branches	1× CPU + 5× PRO 6000 + 1× RTX 5090	7	Requires balanced branch design	🟡
37	🔧 Branch isolation	Individual isolation valves	7	Custom plumbing	🟡
38	🚰 Drain/fill	Dedicated fill and drain ports	1 system	Custom plumbing	🟡
39	🔗 Cooling tubing	13mm ID / 16mm OD EPDM-class tubing	Custom	Chosen to match QDT4-M13/F13	🟢
40	🚨 Cooling safety	ALX-15KA flow/temp/RPM/reservoir alarms + emergency shutdown relay	1 system	Final relay integration required	🟡
41	🧠 Cooling controller	Aqua Computer OCTO for auxiliary fans/sensors	1	No longer responsible for Koolance flow sensing	🟢
42	🌬️ Chassis fans	Noctua NF-A14 industrialPPC-3000 PWM	CAD-dependent	Final quantity depends on chassis CFD/thermal design	🟡
43	🌬️ Motherboard airflow	Dedicated VRM/RDIMM/M.2/BMC airflow channel	1 system	Custom chassis design	🟢
44	🧊 GPU cooling architecture	Full-cover liquid cooling for all six GPUs	1 system	Exact PCB/block matching required	🟡
45	💾 NVMe #1	Samsung 9100 PRO 8TB PCIe 5.0 x4	1	None	🟢
46	💾 NVMe #2	Same	1	None	🟢
47	💾 NVMe #3	Same	1	None	🟢
48	💾 NVMe #4	Same	1	None	🟢
49	💾 Gen5 NVMe total	32TB	—	None	🟢
50	💾 Enterprise NVMe #1	Micron 9550 MAX 25.6TB U.2	1	Runs through motherboard PCIe 4.0 SlimSAS path	🟢
51	💾 Enterprise NVMe #2	Same	1	Same	🟢
52	💾 Enterprise NVMe total	51.2TB	—	None	🟢
53	🔗 U.2 cable #1	SFF-8654 4i → SFF-8639 U.2, 85Ω, ~300mm enterprise cable	1	Exact vendor P/N should match chassis routing	🟡
54	🔗 U.2 cable #2	Same	1	Same	🟡
55	🗄️ U.2 mounting	Dedicated 2.5″ U.2 enterprise carriers	2	Custom mounting	🟡
56	💿 SATA SSD #1	SSSTC ER4 15.36TB SATA III	1	None	🟢
57	💿 SATA SSD #2	Same	1	None	🟢
58	💿 SATA SSD #3	Same	1	None	🟢
59	💿 SATA SSD #4	Same	1	None	🟢
60	💿 SATA total	61.44TB	—	None	🟢
61	💾 TOTAL SSD	144.64TB raw	—	None	🟢
62	🌐 Primary network	Dual onboard 10GbE	2 ports	None	🟢
63	🛡️ Management network	Dedicated 1GbE BMC/IPMI	1	None	🟢
64	🚀 High-speed NIC	Not installed in base build; future ConnectX-class option	Optional	Adding one would require PCIe redesign	⚪
65	🛡️ BMC	ASUS/ASPEED BMC with IPMI/remote management	1	None	🟢
66	🔋 UPS	Schneider Electric Galaxy VS GVSUPS10KRFS, 10kW-class	1	Exact site voltage/configuration must match	🟡
67	🔋 UPS topology	N+1 power-module architecture	1 system	None when correctly configured	🟢
68	🔋 Battery target	≥2kWh usable end-of-life target, aiming for ~15 min	1 system	Exact cabinet depends on final UPS configuration/load	🟡
69	⚡ PDU	High-current metered PDU matched to Galaxy VS output	1	Exact model depends on UPS phase/output	🟡
70	⚡ Facility power	Dedicated AC distribution for dual 2500W PSUs + ALX-15KA	1 system	Professional electrical design required	🟡
71	🏗️ Main chassis	Custom six-GPU liquid-cooled super-tower	1	Custom CAD/fabrication	🟡
72	🏗️ GPU chassis zone	Six individually accessible GPU bays	1	Custom fabrication	🟡
73	🧩 Motherboard zone	Isolated WRX90 platform chamber	1	Custom chassis	🟡
74	⚡ PSU zone	Two dedicated PSU compartments	1	Custom chassis	🟡
75	💾 Storage zone	Dedicated M.2/U.2/SATA service compartment	1	Custom chassis	🟡
76	💧 Cooling zone	Pump/manifold/QDC/drain/fill service section	1	Custom chassis	🟡
77	🔗 PCIe routing	Six controlled 300mm Gen5 riser paths	6	Physical Gen5 validation required	🟡
78	🧰 GPU serviceability	Independent power/PCIe/coolant access for every GPU	1 system	Custom chassis	🟡
79	🔒 Security	Secure Boot + firmware security + platform TPM capabilities	1	None	🟢
80	🧪 AI interconnect	NVIDIA NCCL	1 stack	Actual topology must be benchmarked	🟢
81	🧠 AI framework	PyTorch	1 stack	None	🟢
82	🤗 Model ecosystem	Hugging Face Transformers	1 stack	None	🟢
83	⚡ Distributed AI	DeepSpeed + FSDP + tensor/pipeline parallelism	1 stack	Workload-dependent	🟢
84	🐧 Primary OS	Ubuntu 26.04 LTS	1	None	🟢
85	🪟 Secondary OS	Windows 11 Pro	1	None	🟢
86	🎮 Gaming GPU role	RTX 5090 primarily for display/gaming/creator workloads	—	None	🟢
87	🤖 AI GPU role	Five RTX PRO 6000s as dedicated AI compute pool	—	None	🟢
88	🧠 AI memory model	480GB distributed GPU memory, usable through supported multi-GPU frameworks	—	Not physically unified VRAM	🟢
89	🔬 P2P validation	nvidia-smi topo + NCCL tests + GPU P2P validation	1 procedure	Can only be measured after assembly	🟡
90	🧪 PCIe validation	Gen5 link/error/stability testing for all six paths	1 procedure	Physical commissioning required	🟡
91	🧪 Cooling validation	Pressure/leak/flow/failover testing before full power	1 procedure	Physical commissioning required	🟡
92	🧪 Power validation	Full-load PSU distribution + transient/load testing	1 procedure	Physical commissioning required	🟡
93	🧪 Memory validation	8×128GB memory training/stability test	1 procedure	Physical commissioning required	🟡
94	🗄️ Backup server	QNAP TS-h3087XU-RP-class enterprise NAS	1	Exact drive population remains configurable	🟡
95	🛡️ Backup filesystem	QuTS hero / ZFS-based storage	1	None	🟢
96	🔄 Backup strategy	Snapshots + versioned backups + independent storage	1 system	None	🟢
97	🧪 Expansion architecture	Custom PCIe-class modular expansion	1	Custom engineering	🟡
98	🔬 CXL	Removed from required architecture; future experimental only	—	None	🟢
99	🖥️ Display	ASUS ROG Swift OLED PG32UCDM3	1	None	🟢
100	🖥️ Display specification	31.5″/32″ class, 3840×2160, 240Hz Tandem QD-OLED	—	None	🟢
101	🧰 Mechanical support	CNC GPU carriers, riser mounts, manifold brackets, cable supports	1 system	Custom fabrication	🟡
102	🧪 Pressure-test hardware	Low-pressure coolant-loop pressure/leak testing equipment	1	Commissioning equipment	🟡
103	🧯 Emergency shutdown	Hardware-controlled shutdown path for cooling/power faults	1 system	Custom electrical/control integration	🟡
104	❄️ Room HVAC	2× ~18,000 BTU/h-class inverter AC, N+1	2	Exact capacity depends on room/environment	🟡
105	🌡️ Room thermal monitoring	Dedicated ambient temperature/humidity sensors	1 system	None	🟢
106	⚡ Power metering	UPS/PDU-level real-time power monitoring	1 system	Exact meter depends on PDU	🟡
107	🛠️ Commissioning sequence	Motherboard → RAM → 1 GPU → 6 GPUs → storage → cooling → full load	1 procedure	None	🟢
108	🧪 Burn-in target	Extended CPU/GPU/RAM/storage/PCIe/NCCL simultaneous stress validation	1 procedure	Must be done physically	🟡

🔢 FINAL CAPACITY SUMMARY

Category	Final specification
🧠 CPU	96 cores / 192 threads
🧬 System RAM	1TB ECC DDR5 RDIMM
🤖 AI GPUs	5× RTX PRO 6000 Blackwell 96GB
🎮 Gaming/Creator GPU	1× RTX 5090 FE 32GB
🧠 Aggregate AI VRAM	480GB
🧠 Total physical GPU VRAM	512GB
💾 Samsung Gen5 NVMe	32TB
💾 Micron enterprise NVMe	51.2TB
💿 SATA SSD	61.44TB
💾 Total raw SSD	144.64TB
🔗 GPU PCIe	6× PCIe 5.0 x16
⚡ PSU	2× 2500W = 5,000W
❄️ Cooling	Koolance ALX-15KA = 15,000W class
💧 External pumps	2× PMP-600
🌐 Main networking	2× 10GbE
🛡️ Management	1GbE BMC/IPMI
🔋 UPS	Galaxy VS 10kW N+1
🖥️ Display	4K 240Hz Tandem QD-OLED
🌬️ Room cooling	2× ~18,000 BTU/h-class N+1 AC

🧩 FINAL PHYSICAL ARCHITECTURE

GPU topology

The six GPUs are assigned to the six x16 paths:

PCIe Slot 1 → RTX PRO 6000 #1
PCIe Slot 2 → RTX PRO 6000 #2
PCIe Slot 3 → RTX PRO 6000 #3
PCIe Slot 4 → RTX PRO 6000 #4
PCIe Slot 5 → RTX PRO 6000 #5
PCIe Slot 7 → RTX 5090
PCIe Slot 6 → UNUSED / x8

This uses all six x16 paths while avoiding the x8 slot. ASUS officially lists six PCIe 5.0 x16 slots plus one PCIe 5.0 x16 physical slot operating at x8. (ASUS Global)

Each GPU then leaves the motherboard through its own 300mm PCIe 5.0 x16 riser.

The chosen riser is the Cooler Master MCA-U000C-KPCI50-300, which Cooler Master specifies as a PCIe 5.0 x16, 300mm cable with EMI shielding and a 90° downstream connector. (Cooler Master)

⚡ POWER ARCHITECTURE

                    Galaxy VS
                    10kW N+1
                        │
                 Metered PDU
                        │
          ┌─────────────┴─────────────┐
          │                           │
   CANNON PRO 2500W            CANNON PRO 2500W
          │                           │
    GPU + platform               GPU + platform

Each FSP CANNON PRO is rated at 2500W, accepts 200–240V, and provides four native PCIe 12V-2x6 connectors. Two identical units therefore provide 5000W aggregate capacity and eight native 12V-2x6 outputs, while the six GPUs consume six dedicated connections. (FSP Lifestyle)

❄️ COOLING ARCHITECTURE

The cooling plant is now:

Koolance ALX-15KA

It is rated for:

15,000W cooling capacity

up to 32L/min

700W maximum electrical consumption

8.5L reservoir

and includes built-in monitoring for coolant flow, temperature, pump RPM, fan RPM and reservoir level, along with alarm/relay triggers. (Koolance)

The loop is:

ALX-15KA
    │
    ▼
Pump A ─────┐
            ├── Supply manifold
Pump B ─────┘
                 │
        ┌────────┼─────────┐
        │        │         │
       CPU     PRO1 ...  PRO5   RTX5090
        │        │         │      │
        └────────┴─────────┴──────┘
                 │
            Return manifold
                 │
              ALX-15KA

The previous INS-FM18 + OCTO flow-monitoring arrangement is removed. The ALX-15KA already provides integrated flow monitoring and relay alarms, making that additional control path unnecessary. (Koolance)

💧 COOLANT INTERFACE

The selected QDT4 system is now paired with:

13mm ID / 16mm OD EPDM tubing

rather than the old 13×19mm specification.

The loop therefore has:

QDT4 → 13×16mm EPDM → QDT4

throughout the serviceable GPU branches.

Koolance recommends its own LIQ-702/LIQ-705 coolant family for this type of system and warns about incompatible mixed-metal combinations. (Koolance)

💾 STORAGE ARCHITECTURE

The four Samsung drives occupy the four native PCIe 5.0 x4 M.2 interfaces:

M.2_1 → 9100 PRO 8TB
M.2_2 → 9100 PRO 8TB
M.2_3 → 9100 PRO 8TB
M.2_4 → 9100 PRO 8TB

Samsung's 9100 PRO is PCIe 5.0 x4/NVMe 2.0 and is available in 8TB, with the 8TB model specified at up to 14,800MB/s sequential read and 13,400MB/s sequential write. (Samsung Global Newsroom)

The two Micron 9550 MAX 25.6TB U.2 drives occupy the two motherboard SlimSAS NVMe paths. Micron's 9550 documentation includes 25.6TB U.2 15mm variants and identifies the U.2 interface as SFF-8639. (Micron Technology)

The motherboard's SlimSAS links are PCIe 4.0 x4, so the 9550 drives intentionally operate through that Gen4 path. (ASUS Global)

The four ER4 drives use the motherboard's four SATA ports. SSSTC lists the ER4 15.36TB model as a 2.5″ SATA III 6Gb/s SSD. (SSSTC SSDs)

🤖 AI ARCHITECTURE

RTX PRO 6000 #1 ─┐
RTX PRO 6000 #2 ─┤
RTX PRO 6000 #3 ─┼── NCCL
RTX PRO 6000 #4 ─┤    │
RTX PRO 6000 #5 ─┘    ├── PyTorch
                       ├── FSDP
                       ├── Tensor Parallelism
                       ├── Pipeline Parallelism
                       └── DeepSpeed

The five PRO cards provide:

5 × 96GB = 480GB aggregate GPU memory

The RTX 5090 remains outside that AI pool by default:

gaming / display / creator

This is distributed GPU memory, not physically unified 480GB VRAM.

🖥️ DISPLAY

The final display remains:

ASUS ROG Swift OLED PG32UCDM3

with:

3840×2160

240Hz

Tandem QD-OLED

and current DisplayPort/HDMI connectivity. The display therefore fits naturally with the dedicated RTX 5090 role.

• ### 🔋 UPS + FACILITY

The UPS remains:

Schneider Electric Galaxy VS GVSUPS10KRFS

with a 10kW N+1 architecture.

The UPS is not sized merely for the two computer PSUs; the design also accounts for the ALX-15KA's own electrical consumption, pumps, controls and display.

The ALX-15KA itself can consume up to 700W, so its power draw must be included in the UPS/load budget. (Koolance)

🌡️ ROOM COOLING

The liquid CDU doesn't make the heat disappear; the heat is ultimately rejected into the surrounding environment.

For a roughly 5kW-class computer + cooling system, the room requires its own thermal-management layer.

Final:

2× ~18,000 BTU/h-class inverter AC

configured as:

AC #1 → primary
AC #2 → N+1 redundancy

This keeps the cooling infrastructure from having a single room-temperature failure point.

🧪 FINAL COMMISSIONING ORDER

The finished machine should not be switched directly from assembly to a simultaneous 5-GPU/5090/CPU load.

The correct progression is:

1. Motherboard + CPU
        ↓
2. BIOS / firmware
        ↓
3. 8×128GB memory
        ↓
4. Memory stability
        ↓
5. One GPU
        ↓
6. PCIe Gen5 riser verification
        ↓
7. Remaining GPUs one-by-one
        ↓
8. Six-GPU enumeration
        ↓
9. P2P / NCCL topology
        ↓
10. Samsung NVMe
        ↓
11. Micron U.2
        ↓
12. SATA SSDs
        ↓
13. Coolant pressure test
        ↓
14. Leak / flow test
        ↓
15. Pump failover test
        ↓
16. Cooling stress test
        ↓
17. PSU load distribution
        ↓
18. CPU + GPU simultaneous load
        ↓
19. AI multi-GPU stress
        ↓
20. Long-duration burn-in

🏁 FINAL VERDICT

The component architecture is now complete.

The remaining 🟡 entries are not "we don't know what to buy" problems anymore. They are things that intrinsically require fabrication, installation or physical testing:

six-GPU Gen5 signal integrity

exact ASUS-QVL memory P/N verification

custom coolant-manifold fabrication

pump failover testing

custom chassis fabrication

UPS/PDU/site electrical commissioning

NCCL/P2P benchmarking

long-duration burn-in

The major component choices themselves are now fixed:

9995WX + WRX90E-SAGE SE + 1TB ECC + 5× RTX PRO 6000 + RTX 5090 FE + 144.64TB SSD + 2×2500W FSP + ALX-15KA + 2×PMP-600 + custom six-GPU chassis + 10kW N+1 UPS + dedicated room HVAC.

Bottom line: this is now the final master specification, while the remaining work belongs to engineering fabrication and commissioning rather than another round of swapping core components. 🔥