IdealTech RTX 5090 ComfyUI service-node purchase decision#

Status: Historical decision log (pre-order, 2026-09-22/23). Superseded by: rtx5090-ram-and-cost-performance-decision.md, which records what PC1, PC2 and PC3 were actually ordered with, and the PSU decision (Vertex PX-1200 fleet standard, not the HX1200i default below). The "already built first worker" below is PC1 (GIGABYTE card, two 9100 Pro drives without RAID). The O11 Vision/HydroShift build, Gen4 OS SSD, conditional Tomahawk and price gates were never ordered. The "x16 is a hard requirement" rule below came from the founder's keep-x16 instruction during purchasing and no longer applies: x8 is acceptable when it buys something. The reconciled, current versions of this doc's material live in: rtx5090-node-procurement.md (channels, package pricing, price gates, quote history), rtx5090-node-motherboard-and-pcie-lanes.md (lane topology, motherboard, 10GbE) and the PSU decision (site power and UPS). Original status: CPU/GPU/PSU platform retained; local model-drive count and the resulting motherboard upgrade remain conditional Decision date: 2026-09-22 Lane-topology addendum: 2026-09-23; optional CPU-direct Gen5 OS-drive upgrade investigated and rejected at current cost; retaining the RTX 5090 at PCIe 5.0 x16 is a hard requirement Power addendum: 2026-09-23; 1200 W retained, same-price peers audited, Corsair HX1200i made the value default under a quote gate, and UPS purchase deferred pending power-quality evidence Storage/NAS addendum: 2026-09-23; start with one local 2 TB Samsung model cache plus a separate 2 TB OS SSD, use a shared HDD-backed NFS origin, and add neither per-worker HDDs nor RAID 0 without a capacity or measured-performance trigger Network addendum: 2026-09-23; use onboard 5GbE initially and defer the RM499 10GbE adapter until measured network demand justifies it Scope: Second Malaysian RTX 5090 worker purchased through IdealTech for ComfyUI-as-a-service Pricing status: RM38,647 is the historical two-Samsung/1 TB-OS quote; with a 2 TB OS drive, the provisional Tomahawk/Seasonic total is RM36,398-RM36,498 for one model SSD or RM39,147-RM39,247 for two, pending a written re-quote

Context#

Malaysian retailers were requiring an RTX 5090 to be purchased as part of a complete system. The goal was therefore not to find the cheapest standalone GPU, but to extract the best useful ComfyUI service node from the builder's mandatory bundle without paying for components that do not improve service capacity or latency.

One worker had already been built with the following independently selected specification:

Ryzen 9 9950X
MSI MAG X870E Tomahawk Max WiFi
Zotac RTX 5090 Solid OC 32 GB
T-Create Expert 64 GB DDR5-6000 CL34
2x Samsung 9100 Pro 2 TB RAID 0
Seasonic Vertex PX-1200 ATX 3.1 Platinum
Arctic Liquid Freezer III Pro 360
Fractal Design North XL

The second purchase introduced a RM30,999 promotion based on a Ryzen 7 9850X3D, Palit/Zotac RTX 5090, MSI PRO X870E-S EVO, G.Skill 64 GB, one Kioxia 1 TB SSD, Antec GSK1000, HydroShift 360 and O11 Vision. The promotion allowed component upgrades by paying the difference. This changed the answer from a simple standalone price-list optimization into a marginal-cost decision around a heavily discounted base.

Procurement update: source the spread, not necessarily the whole machine#

The component decision below remains valid, but the purchasing decision is no longer “buy every upgrade from IdealTech.” The RM38,647 quote remains a historical comparison point, not the current target, because the local RAID decision and system drive have changed. Obtain live, written quotes from the Malaysian distributors and authorised dealers and split the order only where the net saving survives tax, delivery, warranty and assembly costs.

Direct distributor purchasing is possible in principle, but “distributor price” is not normally a public one-unit consumer price. Distributors commonly require a reseller, system-integrator or corporate account, company documents, a minimum order, cash terms or recurring volume. They also avoid competing openly with the dealer channel that they supply. We should approach them truthfully as a company deploying its own AI infrastructure, offer cash payment and expected repeat purchases, and ask either for a corporate account or an introduction to their lowest-margin authorised dealer. We should not represent the company as a retailer if it is not one.

RTX 5090 channel#

Sun Cycle is the most relevant first contact because it distributes both Palit and Zotac, operates the local warranty portal, and publishes current SRPs. On 2026-09-22 its list showed:

These are SRPs, not confirmed dealer costs or offers to sell directly. Contact Sun Cycle sales on +60 10-392 7393 and request a live cash quote, available quantity, exact SKU, any attachment requirement and the warranty path. The target SKUs are NE75090019R5-GB2020G for the Palit GameRock and ZT-B50900J-10P for the Zotac Solid OC.

Standalone authorised-retailer supply is not impossible. Dotatech's RTX 50 collection advertised the Zotac Solid OC “from RM20,799” on the same date, although its individual product page showed a conflicting higher price. Dotatech appears on Zotac's Malaysia where-to-buy list, so it is worth requesting a written GPU-only quote rather than relying on either web price.

The IdealTech promotion nevertheless remains difficult to beat. The previously reconstructed list value of the base package was approximately RM35.2k, including a RM22,999 RTX 5090. Holding the other list values constant implies that the RM30,999 promotion values the bundled GPU at roughly RM18.8k. This is an approximation rather than an invoice allocation, but it means that finding a standalone 5090 at RM20.5k does not automatically save money: it gives up approximately RM4.2k of package discount. A full independent build only becomes compelling if the live 5090 quote approaches RM18.8k or distributor savings on the rest of the BOM more than offset the gap.

The claim that bundling exists solely to stop scalpers should not be treated as the complete explanation. There is no evidence here of a Malaysian law or universal NVIDIA rule requiring a full system. Scarce 5090 allocation, dealer-margin protection, warranty/fraud exposure and moving attached inventory can all produce the same retailer policy. The presence of at least one authorised GPU-only listing confirms that bundling is a channel policy, not an inherent product requirement.

The RTX 5090 itself remains non-negotiable for this workload. Its 32 GB VRAM and measured ComfyUI throughput are why the machine is being purchased; a 16 GB RTX 5080 is not an equivalent substitute merely because it is sold without a bundle.

Samsung SSD channel and the importance of live prices#

M-Link System's 17 March 2026 dealer list listed the Samsung 9100 Pro 2 TB (MZ-VAP2T0BW) at RM1,630 with a five-year warranty. That historical price would have made direct sourcing attractive. However, M-Link's 17 August 2026 list raised the same drive to RM2,592 and marked the 9100 Pro range as limited stock. This is the relevant recent comparison; the March price must not be used as a purchasing assumption.

The split-order calculation at the August published price is:

This table uses the current Seasonic quote. A confirmed Corsair or Lian Li substitution at the provisional list-price delta would reduce both compared totals by the same RM310 or RM620 respectively, so neither would change the RM414 SSD-sourcing spread.

RM414, or about 1.1%, is not enough to justify split invoicing and warranty handling unless the live M-Link quote improves. Any excluded tax or delivery charge could erase it. The arithmetic also gives a reusable rule: the two independently purchased drives together must cost less than RM5,598 to beat the IdealTech storage transaction at all. To justify the extra procurement path with at least RM1,000 saving, each drive should be RM2,299 or less, all-in. At RM2,592, buy the drives through IdealTech unless there is another material benefit.

M-Link explicitly publishes dealer prices, warranty contacts and an invitation to join as a reseller. Ask for a current September quote, whether the price includes every applicable tax, stock reserved for two matching units, delivery, five-year local warranty, and whether a corporate end-user or system-integrator account is acceptable. Samsung Malaysia states that SSD warranty service is carry-in to the point of purchase, so the invoice and authorised local channel matter.

Quote sequence and decision gates#

1. Ask IdealTech for the base promotion plus the CPU and motherboard substitutions, no Samsung upgrade or add-on, and the included Kioxia retained. Request side-by-side installed-system prices for the regular Corsair HX1200i, Seasonic Vertex PX-1200 and Lian Li SX1200P; do not assume the standalone list-price differences will be honored inside the promotion.
2. Ask M-Link for a same-day, all-in quote for two MZ-VAP2T0BW drives. Split the SSD purchase only at RM2,299 per drive or below; between RM2,300 and RM2,799, decide whether the small saving is worth separate warranty handling; at RM2,800 or above, use IdealTech.
3. In parallel, ask Sun Cycle for a corporate/dealer cash quote for one Palit and one Zotac 5090, with no attached parts. Ask Dotatech and the other Zotac-authorised retailers for the same GPU-only quote.
4. Price the complete independent BOM only if the best authorised 5090 quote is near RM18.8k, or if written distributor quotes on the remaining parts demonstrate enough additional saving to compensate.
5. Compare final landed totals, not headline component prices. Require exact SKU, local authorised invoice, serial numbers, stock confirmation, DOA terms and warranty owner before placing an order.
6. Use the board's onboard 5GbE initially. Defer the TX401 and any onboard-10GbE substitution until measured network traffic shows that 5GbE is a material bottleneck.

This sourcing approach keeps the scarce 5090 bundle available as a fallback while testing the two realistic places to remove margin: a GPU bought through its actual distributor/dealer channel and high-markup add-ons bought from their specialist distributor.

Current storage and network direction#

The completed single-NVMe versus RAID 0 experiment supersedes the earlier two-drive default. The fixed application matrix did not show a repeated practical RAID advantage, and later schema-v4 controls proved that a single physical Gen5 NVMe can serve every cold model byte and complete the workload without RAID.

The current starting configuration is therefore:

Active models must run from local NVMe. The NAS is a shared source used to populate each worker before a job is admitted; normal customer jobs should not read their model files directly over NFS. The complete architecture, performance arithmetic, pricing, NFS layout and upgrade gates are maintained in the dual-worker model storage and NFS/NAS decision.

Both the included MSI PRO X870E-S EVO and the Tomahawk provide onboard 5GbE, one CPU Gen5 x4 M.2 path, chipset M.2 storage and a Gen5 x16 GPU slot. If RAID is rejected and 2 TB is demonstrably sufficient before the order is placed, the base board can implement the initial topology and saves RM450. Retain the Tomahawk while the second drive remains a real possibility: that RM450 preserves the independent second CPU Gen5 x4 path and avoids a later motherboard replacement.

Using the prior promotional substitutions and the two plausible 2 TB OS drives gives this provisional Seasonic-based range; it must be re-quoted because the original written quote used a 1 TB OS drive:

No RM499 network adapter is included: onboard 5GbE is the current decision.

Historical two-drive component decision and quote#

The following was the decision before the later RAID experiment and the move to a 2 TB OS drive. It is retained to preserve the quote history, not as the current storage order. The platform reasoning and exact GPU request remain useful.

The historically quoted system was:

The price calculation is:

This table does not price the current 2 TB OS requirement or one-drive starting configuration. The seller should previously have been asked to retain the Kioxia during the SSD upgrade if possible. The arithmetic totals RM38,647 if IdealTech first credits its RM899 list price and then charges RM899 to add it back. The current procurement request must instead quote the three scenarios in the preceding section.

Why the promotional package changes the answer#

The initial comparison used the standalone component prices in the extracted IdealTech list. At those prices, the Ryzen 9 9950X, premium PSU and Samsung Gen5 drives appeared to be the obvious places to reduce cost.

The promotion changes the marginal economics. An indicative reconstruction of the base package from the same list is approximately RM35.2k before assembly, compared with its RM30,999 promotional price. The exact fan-pack accounting may differ, but the package clearly carries roughly RM4k or more of discount. The real discount is concentrated in the base configuration; most upgrade charges are ordinary retail price differences:

This means it is advantageous to enter through the promotion, even if several base components are then upgraded.

The approximately comparable standalone build was RM41,482 without a separate OS drive. Adding the RM899 Kioxia would make it RM42,381. The RM38,647 promotional build therefore saves about RM3,734 while retaining the desired CPU, PSU and model-cache layout. The cases, coolers and RAM brands differ, so this is a practical system-level comparison rather than an identical bill of materials.

Workload and benchmark evidence#

Current interpretation: The historical schema-v3 comparison below is retained for chronology, but its physical-read interpretation and two-drive recommendation are superseded. Later schema-v4 work discovered a buffered loop-backing-file cache on some Vast hosts, and the completed dual-workload matrix found no repeated application-level RAID advantage. See the current experiment conclusion.

The relevant workload is a ComfyUI service that changes large model sets. Warm inference is primarily a GPU problem, but cold requests also depend on how quickly approximately 42.47 GB of model and input files can traverse the complete storage, filesystem, container and model-loading path.

The storage investigation is documented in model-read-storage-layouts.md. Its benchmark protocol:

• flushes and evicts every resolved file with fsync and POSIX_FADV_DONTNEED;
• reads the real 42,471,968,783-byte file set in normal 8 MiB buffered reads;
• verifies physical reads using process and cgroup counters;
• separately runs five cold-fresh-process/warm-same-process workflow pairs; and
• requires warm samples to perform zero physical model reads.

The important observed layouts were:

The supporting captures are:

• repeatable dual-SN850X RTX 5090 result
• same-host dual-SN850X repeat
• dual-9100 RTX Pro 5000 result
• single-9100 RTX 5090 result
• four-drive negative result

All observed hosts sustaining approximately 10 GiB/s or more on the real file set used two NVMe devices in parallel. No measured single-drive host crossed the threshold through the complete application path.

Storage decision and correction made during review#

Superseded decision: This section explains why the earlier two-drive quote was selected. The current decision is one 2 TB Samsung model cache initially, with a second drive triggered by capacity or a local same-machine application A/B result—not by this historical table.

An early recommendation proposed replacing the Samsung drives with two MSI Spatium M480 2 TB Gen4 drives. That would save RM2,700 at standalone prices, and two similar Gen4 SN850X drives had already demonstrated more than 11 GiB/s.

That recommendation was withdrawn after reviewing the storage postmortem. The M480 alternative is plausible but not proven:

• no captured M480 host demonstrates the required application-level model-read rate;
• nominal per-drive sequential specifications do not predict this workload reliably;
• direct or high-queue-depth fio results can materially exceed normal buffered multi-file reads; and
• filesystem placement, container backing, PCIe topology, thermal state and RAID configuration all affect the result.

The selected two-drive 9100 Pro configuration has direct supporting evidence and is the layout already recommended by the storage investigation. For a production purchase, the unvalidated saving is not worth giving up the best-supported route to the 10 GiB/s target.

A single 4 TB 9100 Pro was also rejected. It preserves capacity and avoids RAID 0, but the captured single-drive machine achieved only 5.93 GiB/s and added 21.31 seconds to the cold path. The two-drive array is deliberately a disposable cache, so redundancy is less valuable than model-switch latency.

Required physical storage topology#

The current initial layout is:

Ryzen 9 9950X
├── PCIe 5.0 x16 ── Zotac RTX 5090
│
├── M2_1, CPU Gen5 x4 ── Samsung 9100 Pro 2 TB ── /srv/model-cache
├── M2_2, CPU Gen5 x4 ── empty; reserved for a justified second model drive
│
├── M2_3, chipset Gen4 x4 ── 2 TB Gen4 OS SSD
│                             Linux, Docker, application, downloads and outputs
├── M2_4, chipset Gen4 x4 ── available for future scratch/output storage
└── onboard 5GbE ──────────── shared NFS/NAS model origin for cache population

On the MAG X870E Tomahawk Max, leaving M2_2 empty keeps rear USB4 available. If a second model drive is later justified, configure M2_2 for full Gen5 x4 and accept that the rear 40 Gbps USB-C ports will be disabled. This does not reduce the GPU's x16 link. The relevant lane-sharing details are in MSI's specification.

Use the following initial storage configuration:

• a direct XFS or ext4 filesystem on the single Samsung, not a buffered loop-backing file;
• XFS or ext4 with 1 MiB partition alignment and noatime;
• periodic fstrim, not continuous discard on the hot path;
• motherboard heat spreader and direct airflow over the Gen5 drive;
• a read-only shared NFS/NAS origin from which the cache can be rebuilt; and
• 15-20% local free-space headroom with cache-aware eviction.

Separating the OS remains important. Container layers, logs, downloads and output writes must not contend with cold model reads. The Samsung model volume is a disposable cache, so no irreplaceable data belongs there. If a second Samsung is later added, use mdadm RAID 0 with the previously tested 512 KiB starting chunk only after the local before/after workload test or a demonstrated 2 TB capacity shortfall.

Why the motherboard upgrade stays#

The base MSI PRO X870E-S EVO provides one CPU-connected Gen5 x4 M.2 slot and two chipset-connected Gen4 x4 slots. See MSI's PRO X870E-S EVO specification.

It cannot reproduce a future topology of two model-cache members on independent CPU-connected M.2 paths. It can, however, implement the current initial layout: one CPU Gen5 x4 Samsung, one chipset Gen4 OS drive, onboard 5GbE and the RTX 5090 at Gen5 x16. The RM450 Tomahawk upgrade is therefore now option value, not a current one-drive requirement. Retain it while the second-drive test or capacity decision remains open; drop it and save RM450 if RAID is rejected and a 2 TB cache is proven sufficient before ordering.

10GbE network review#

Current decision: Use onboard 5GbE first. The adapter and motherboard audit below establish the correct future upgrade path, but the RM499 TX401 is not part of the initial order.

Units and actual role#

The motherboard provides 5 Gbit/s Ethernet (5GbE), not 5 GB/s. Likewise, the proposed upgrade is 10 Gbit/s Ethernet (10GbE), whose raw line rate is 1.25 GB/s and whose useful TCP payload is normally lower. If the requirement were literally 10 GB/s, neither an onboard 10GbE controller nor the proposed adapter would satisfy it; that would require an 80-100GbE-class design.

The selected model cache reads locally at several GiB/s, so 10GbE is not a substitute for local NVMe. Its value would be faster model replication, cache warming, backup, output transfer, and communication between workers and storage. Moving the 42.47 GB model set over an ideal 10 Gbit/s link has a raw lower bound of about 34 seconds, versus about 68 seconds at 5 Gbit/s, before protocol and storage overhead. The current choice is to defer the adapter until measurement shows that 5GbE cache fills are a material production bottleneck.

Why the add-in adapter preserves the important lanes#

MSI specifies three expansion slots on the MAG X870E Tomahawk Max:

• PCI_E1: CPU-connected PCIe 5.0 x16 for the RTX 5090;
• PCI_E2: chipset PCIe 3.0 x1; and
• PCI_E3: chipset PCIe 4.0 x4.

The TP-Link TX401 is a PCIe 3.0 x4, single-port 10GBASE-T adapter supporting 100M/1G/2.5G/5G/10G speeds and Linux. IdealTech lists it at RM499. Installed in PCI_E3, it has substantially more host-interface bandwidth than one 10GbE port requires and uses chipset lanes rather than the CPU's PEG block. It therefore does not reduce the RTX 5090 from x16 and does not consume either CPU-direct Gen5 x4 path used by the Samsung RAID members. MSI's Tomahawk specification does not disclose an M.2 lane-sharing penalty for PCI_E3 on this board.

The NIC and Kioxia system drive ultimately share finite chipset/uplink bandwidth with other chipset devices. That is acceptable: 10GbE consumes at most 1.25 GB/s raw, normal OS traffic is far below a sustained SSD benchmark, and the high-throughput model array remains CPU-direct. Concurrently benchmarking every chipset device could create contention, but the production data path is not designed around doing that.

Onboard-10GbE motherboard audit#

Prices are from the 2026-09-22 IdealTech extraction. The meaningful comparison is not motherboard price alone: it is whether onboard 10GbE preserves the RTX 5090 at x16 and both Samsung drives at CPU Gen5 x4.

ASUS explicitly states in the ProArt specification that enabling M.2_2 changes PCIEX16(G5)_1 to x8. MSI's Godlike specification confirms two CPU Gen5 M.2 sockets, 10GbE, and the same USB4/M2_2 BIOS trade used on the Tomahawk. The Godlike is the only unambiguously identified IdealTech 10GbE listing confirmed here to satisfy all three non-negotiable link states, but paying RM2,801 to integrate a replaceable RM499 function has no economic justification.

There is also a product-description mismatch that must not be papered over. IdealTech's RM4,120 Dark Hero line says 5GbE + 2.5GbE and three Gen5 M.2 sockets, while ASUS's current Dark Hero specification says 10GbE + 5GbE and two CPU Gen5 M.2 sockets; it allows the second socket to run x4 by disabling USB4 and does not take lanes from the x16 GPU. If IdealTech confirms the exact current SKU, revision, and rear I/O, it is a valid alternative—but still costs RM1,922 more than Tomahawk plus TX401. Gigabyte's current Aorus Master X3D Ice specification also conflicts with the retailer summary: it now shows 10GbE + 5GbE, but its second CPU M.2 is limited to x2 when populated, so it remains disqualified. The broader-market ASRock X870E Taichi White has onboard 10GbE but only one Gen5 x4 M.2 socket, so it cannot replace the Tomahawk for this storage design.

An add-in NIC also has an operational advantage: if it fails or a later network moves from 10GBASE-T to SFP+, the NIC can be replaced without replacing or RMAing the motherboard. Conversely, the TX401 is the right choice only if the switch side is RJ45 10GBASE-T. If the network switch exposes SFP+ instead, quote a Linux-supported SFP+ NIC and DAC/fibre link rather than adding a hot 10GBASE-T transceiver merely to use the TX401.

Physical-fit gate#

The electrical lane allocation is valid, but physical clearance must be confirmed before treating the adapter as final. Zotac specifies the Solid OC as a 3.5-slot, 67.8 mm-thick card, while the O11 Vision has six horizontal expansion slots. PCI_E3 is the only suitable x4 slot; PCI_E2 is x1 and cannot provide the TX401's required interface.

Ask IdealTech to install the actual Zotac in PCI_E1 and dry-fit the TX401 in PCI_E3. Accept the layout only if the NIC seats normally, the case bracket aligns, the GPU support does not interfere, and the NIC does not touch the GPU or materially obstruct its fans. A photo and link-state evidence are more useful than an assurance that both items appear separately in compatibility lists. Do not solve a clearance problem by moving the GPU to another slot or reducing it to x8. If the dry-fit fails, revisit case/slot placement or an SFP+ NIC with a smaller cooler before paying thousands for a motherboard.

Network decision#

1. Use onboard 5GbE on both workers initially. Test it against the actual NAS and simultaneous cache fills before buying adapters.
2. Do not add the RM499 TX401 now. One card per worker would consume RM998 while the shared origin may remain limited by one or two HDDs well below even 5GbE.
3. If 10GBASE-T is later justified, retain the selected board and use the TX401, subject to the PCI_E3 dry-fit. Do not replace the motherboard merely to integrate a replaceable NIC.
4. Do not select the ASUS ProArt for a possible two-drive layout. Enabling its second Gen5 M.2 path reduces the RTX 5090 to x8.
5. Do not buy the Dark Hero or Godlike for its NIC. Their premiums do not produce useful ComfyUI throughput or availability return.

Optional lane-expansion review: moving the OS drive to CPU-direct Gen5#

Current relevance: This analysis applies only if the optional second Samsung model drive is installed. With the initial single-model-drive plan, the OS still belongs on an inexpensive chipset Gen4 path; there is no measured reason to spend CPU lanes or Gen5 money on it.

This question was reopened on 2026-09-23 to explore whether a motherboard or platform upgrade could move the OS drive from its chipset-connected Gen4 path to a third CPU-direct PCIe 5.0 x4 path without sacrificing any RTX 5090 bandwidth. This is an optional optimization, not a requirement of the selected build.

The selected build still has exactly three M.2 drives:

• two Samsung 9100 Pro 2 TB drives on CPU-direct Gen5 x4 paths, forming the RAID 0 model cache; and
• one Kioxia Exceria Basic 1 TB OS drive on a chipset-connected Gen4 x4 path.

The exploration does not add a third cache member or a fourth drive. For the optional upgrade to count as a complete CPU-direct Gen5 solution:

• two Samsung 9100 Pro drives must each connect directly to the CPU at PCIe 5.0 x4 and form the RAID 0 model-cache volume;
• the OS drive must move to the third CPU-direct PCIe 5.0 x4 path and must itself be a Gen5 SSD;
• the Zotac RTX 5090 must retain PCIe 5.0 x16;
• a chipset-attached socket, a Gen4 socket, or a socket behind a PCIe switch does not count as a third direct Gen5 path; and
• the existing two-drive benchmark evidence must not be traded for an unmeasured topology merely to increase the visible drive count.

No AM5 motherboard satisfies those upgrade conditions. The limitation comes from the AM5 CPU's lane budget, not from MSI's motherboard design. The correct present decision remains the Ryzen 9 9950X and X870E Tomahawk Max: both cache drives stay CPU-direct at Gen5 x4, the OS remains on the chipset at Gen4 x4, and the RTX 5090 retains Gen5 x16. Threadripper/TRX50 is the technically clean route to a third CPU-direct Gen5 M.2 path, but it is a platform migration rather than a sensible motherboard-only upgrade.

What running the RTX 5090 at x8 instead of x16 means#

The selected Zotac RTX 5090 Solid OC has a PCIe 5.0 x16 host interface. PCIe width describes the connection between the CPU and GPU:

Changing to x8 does not disable CUDA cores, reduce the 32 GB VRAM capacity, or halve the GPU's internal GDDR7 bandwidth. A fully resident inference can therefore show little or no proportional slowdown. It does, however, halve the maximum bandwidth for CPU-to-GPU and GPU-to-CPU transfers. That path matters during model upload, CPU offload, tensors that spill beyond VRAM, pinned-memory transfers and workloads that stream data rather than leaving it resident.

The performance loss is workload-dependent rather than automatically 50%, but the link capability is unambiguously halved. This service is being built around a flagship 32 GB GPU and explicitly optimizes the cold model-loading path. Preserving x16 is therefore a hard acceptance condition; an x8 configuration is rejected even if a gaming benchmark or a warm, fully resident workflow happens to show a small difference.

Why the Ryzen 9 9950X cannot supply x16 plus three x4 paths#

AMD specifies 28 native lanes in total and 24 usable lanes for devices on the Ryzen 9 9950X. The remaining four form the chipset uplink. The 24 device lanes can be allocated as:

AM5 Ryzen 9 9950X device lanes
├── 16 lanes: PEG block, normally RTX 5090 at PCIe 5.0 x16
├──  4 lanes: dedicated CPU NVMe path
└──  4 lanes: general-purpose path used for a second NVMe or USB4

The Tomahawk makes the best allocation for this node:

RTX 5090       = x16
M2_1           = x4 CPU Gen5
M2_2           = x4 CPU Gen5, with rear USB4 disabled at full speed
Total          = 24 usable device lanes

The hypothetical all-CPU-direct layout would consume 16 + 4 + 4 + 4 = 28 device lanes before the chipset link: x16 for the GPU, x4 for each of the two Samsung RAID members, and x4 for the OS drive. A 24-lane AM5 CPU cannot provide that allocation. No BIOS update, X870E chipset, or premium VRM can create four missing CPU lanes.

The non-X3D Gigabyte X870E Aorus Elite WiFi7 exposes three physical CPU-connected Gen5 sockets by bifurcating the 16-lane PEG block. Its actual allocation is:

RTX 5090       = x8
M2A_CPU        = x4 from the dedicated NVMe path
M2B_CPU        = x4 taken from the PEG block
M2C_CPU        = x4 taken from the PEG block
USB4 controller = x4 from the remaining general-purpose path
Total          = 24 usable device lanes

Gigabyte explicitly states that populating either M2B_CPU or M2C_CPU makes PCIEX16 operate at no more than x8 in the official Aorus Elite specification. The board does not avoid GPU degradation; it exposes the user's choice between x16 graphics and the extra CPU M.2 paths. The same basic x8/x4/x4 trade applies to the other qualifying three-socket AM5 boards.

Audit of the IdealTech AM5 options#

Prices below are from the normalized IdealTech list dated 2026-09-22. Manufacturer lane tables, rather than the retailer's short descriptions, determine whether a board qualifies.

The Aorus Master specification, ASUS B850-E specification, ASUS Crosshair Hero specification and MSI Godlike specification all disclose the relevant sharing or bifurcation. The Godlike can keep x16 while using its two onboard CPU M.2 sockets, but adding the third via its included XPANDER card occupies the second CPU PEG slot and changes the PEG allocation to x8/x8.

The retailer list also contains misleading socket counts. The Gigabyte Aorus Master X3D, Aorus Elite X3D Ice and Aero X3D Wood, ASUS X870E Dark Hero, and ASUS X870E-E Gaming WiFi7 Neo are described as having three Gen5 M.2 sockets in the short listing, but their current manufacturer specifications expose only two CPU Gen5 M.2 paths. They are not alternatives for this optional upgrade.

Whether to change the CPU and platform#

Changing to another AM5 CPU does not solve the problem. Ryzen 9000 and Ryzen 9000 X3D processors use the same AM5 I/O model; buying a 9950X3D-class part changes cores and cache, not the number of lanes available to this topology. Additional X870E chipset lanes are Gen4 lanes behind the shared chipset uplink and are not substitutes for another CPU-direct Gen5 path.

If the optional OS-path upgrade were pursued, the least-expensive lane-correct platform visible in the IdealTech list would be:

AMD's Threadripper 9960X specification lists 92 total and 88 usable lanes. AMD's Threadripper platform description states that TRX50 exposes up to 80 PCIe 5.0 lanes. The TRX50 AERO D specification confirms that M2A_CPU, M2B_CPU and M2C_CPU are each PCIe 5.0 x4, while both primary CPU expansion slots support PCIe 5.0 x16. This produces the optional target topology without bifurcating the GPU:

Threadripper 9960X / TRX50 AERO D
├── PCIEX16_1, CPU Gen5 x16 ── RTX 5090
├── M2A_CPU, CPU Gen5 x4 ───── Samsung 9100 Pro 2 TB ─┐
│                                                      ├─ RAID 0 model cache
├── M2B_CPU, CPU Gen5 x4 ───── Samsung 9100 Pro 2 TB ─┘
├── M2C_CPU, CPU Gen5 x4 ───── Gen5 OS SSD
└── M2D_CPU, CPU Gen4 x4 ───── available

The IdealTech description incorrectly summarizes the TRX50 AERO D as 4x Gen4 M.2; Gigabyte's primary specification shows three Gen5 plus one Gen4 socket, all CPU-connected. The board is 264 mm wide and fits within the O11 Vision's published E-ATX limit of 280 mm. The selected Zotac card is 329.7 mm long, below the case's 430 mm GPU limit. Physical fit is therefore plausible, though the builder must still validate radiator, cable and 3.5-slot GPU clearance in the complete assembly.

The CPU-and-motherboard subtotal changes from RM4,698 for the 9950X plus Tomahawk to RM10,498 for the 9960X plus TRX50 AERO D: RM5,800 more before RDIMM, cooler and any promotional-package effects. Threadripper also raises CPU TDP from 170 W to 350 W. The 1,200 W PSU, CPU cooler and sustained case thermals must be re-validated around a 575 W RTX 5090 rather than assumed adequate from connector wattage alone. IdealTech lists compatible RDIMMs as ASK, so a defensible final system total cannot yet be calculated. The quoted Kioxia Exceria Basic system drive is a PCIe 4.0 device. Moving it to M2C_CPU would make it CPU-direct, but it would still negotiate only Gen4 x4. Achieving an actual CPU-direct Gen5 OS link would therefore require a replacement Gen5 OS SSD as well as the Threadripper platform.

The ARCTIC Freezer 4U-M listed by IdealTech is explicitly sTR5-compatible and rated for processors up to 350 W, but cooler choice still needs a noise and sustained boost review. The existing workload averaged fewer than two CPU cores during cold loading, so Threadripper is not justified by present CPU utilization. Its value here is PCIe topology alone.

Final lane-expansion decision#

1. Retain the Ryzen 9 9950X. For the initial single-drive layout, either the base PRO X870E-S EVO or the Tomahawk keeps the RTX 5090 at Gen5 x16, places the Samsung on CPU Gen5 x4 and leaves the 2 TB OS drive on chipset Gen4 x4. Retain the Tomahawk only while its second independent CPU Gen5 x4 path has real option value for the unresolved expansion test.
2. Do not replace the Tomahawk with the Aorus Elite, Aorus Master, ASUS B850-E, Crosshair Hero or MSI Godlike merely to obtain a third CPU M.2 socket. Every such AM5 path reduces the GPU to x8.
3. Do not change to another AM5 CPU for lanes. It cannot change the platform lane budget.
4. Do not migrate to Threadripper solely to accelerate the OS drive without measured evidence of an OS bottleneck. Threadripper remains a future option only if a measured workload justifies its RM5,800 CPU-and-board premium plus new RDIMM, cooling and Gen5 OS storage.

If a future Threadripper build is evaluated, its topology test should simultaneously show the RTX 5090 at 32.0 GT/s x16, both Samsung RAID members at 32.0 GT/s x4, and the replacement Gen5 OS SSD at 32.0 GT/s x4. It should also confirm that all three NVMe devices enumerate beneath CPU root ports rather than the chipset. An x8 GPU result remains unacceptable regardless of synthetic SSD throughput.

CPU decision#

The Ryzen 7 9850X3D included in the package is already an 8-core/16-thread Zen 5 CPU with a 5.6 GHz boost clock and 120 W default TDP. It is sufficient for the measured workload. Resource traces from the cold model-loading captures averaged fewer than two process CPU cores, while warm execution averaged about 1.16 cores. The 5090 determines most resident throughput.

The Ryzen 9 9950X upgrade therefore is not required for the current ComfyUI benchmark. It was retained because:

• the marginal promotional upgrade is only RM700;
• it doubles the core count for future preprocessing, encoding and concurrent service work;
• it brings the second worker closer to the existing worker's platform; and
• under the current Seasonic quote, removing both the CPU and PSU upgrades saves RM1,600 before accounting for storage changes; the saving would be RM980 if the provisional Lian Li PSU substitution is accepted.

For a strictly minimum-cost worker, retaining the 9850X3D remains technically reasonable. The final choice favors operational headroom and fleet consistency over the last RM700 of savings. AMD's published 9850X3D specification is available here.

PSU, incoming-power and UPS decision#

Decision summary#

• Use a 1200 W ATX 3.1/PCIe 5.1 PSU with its own supplied 600 W 12V-2x6 GPU cable. This is the capacity sweet spot for this machine. Do not pay for 1500-2200 W unless the configuration later gains another major load.
• Keep a Seasonic Vertex PX-1200 that has already been purchased. It is an excellent unit and there is no economic case for replacing it.
• For the unbuilt IdealTech node, the regular Corsair HX1200i is the balanced value default. Select it if the final installed-system quote is at least RM250 lower than the Seasonic and confirms the exact ATX 3.1 revision, supplied 600 W 12V-2x6 GPU cable, and local ten-year warranty. At the extracted list prices it saves RM311 without a meaningful electrical or capacity sacrifice.
• Retain the Lian Li SX1200P as the cost-first option. Select it if the installed quote is at least RM500 below the Seasonic and the exact black SX1200P SKU carries the local ten-year warranty. Otherwise choose the Corsair or Seasonic according to the RM250 gate above.
• Spend first on the electrical installation and switchboard surge protection. Those measures address failure modes that no premium PC PSU can address alone.
• Do not buy two large UPS units yet solely to “regulate voltage.” Instrument the circuits and record actual outages, sags and swells first. Add UPS capacity when the measured interruption cost justifies it.

The complete evidence, same-tier comparison, limitations, site-power plan, and procurement decision tree are maintained in the dedicated power-supply and site-power decision.

Why 1200 W is the balanced capacity#

The selected Zotac RTX 5090 Solid OC is specified at 575 W and recommends a 1000 W system PSU. NVIDIA's own RTX 5090 specification also gives 575 W total graphics power and a 1000 W system recommendation based on a Ryzen 9 9950X host. The included Antec GSK1000 therefore meets the published requirement, but it is the floor rather than a generous continuous-service design point.

The real ComfyUI captures averaged fewer than two CPU cores during cold loading, so this worker normally will not combine a 575 W GPU load with maximum CPU package power. Nevertheless, the system also contains the pump, fans, motherboard, memory and three NVMe drives, and future preprocessing could raise concurrent CPU use. Moving from 1000 W to 1200 W gives 20% more nominal capacity, lower sustained loading and useful thermal/aging margin without the waste of a 1600 W unit. A larger PSU does not make the computer draw its nameplate wattage, but capacity beyond 1200 W provides no throughput or meaningful additional protection for this parts list.

PSU model comparison and price gate#

The prices are from the IdealTech list dated 2026-09-22. The provisional system totals below apply the list-price difference from the included RM799 Antec GSK1000; IdealTech must confirm the actual promotional substitution price before the total is changed.

Cybenetics' 230 V testing confirms that the Antec GSK1000, Antec GSK1200, Lian Li SX1200P, Corsair HX1200i and Seasonic Vertex PX-1200 all operate within their tested limits. The Corsair sample was quieter, regulated 12 V more tightly, and had essentially the same full-load 12 V ripple as the tested Vertex; the Vertex held output longer after AC loss and carries two more years of warranty. These are small trade-offs, not evidence of a reliability gulf. The Lian Li is noisier and has higher ripple, but its results remain comfortably suitable for this system. None is a mains conditioner.

At the list-price deltas, the Corsair would produce a provisional RM38,337 total and the Lian Li approximately RM38,027, compared with the quoted RM38,647 Seasonic ceiling. Request all three prices, but do not delay the node or accept an ambiguous SKU, warranty, or cable revision. The current ceiling remains in force until the builder supplies written substitution quotes.

What a PSU can and cannot protect#

The Seasonic, Corsair, Lian Li and Antec models all provide the normal internal protection classes: over-current, over-voltage, under-voltage, short-circuit, over-power and over-temperature shutdown. These protect the PSU and downstream components when the PSU output or load becomes abnormal. Better ripple control, transient response and hold-up time can improve operating margin, but the difference is measured in milliseconds and does not bridge a grid outage.

Protection is layered because no single device covers every failure mode:

Malaysian residential electrical arrangement#

Malaysia's Energy Commission specifies 230 V single-phase and 400 V three-phase low-voltage supply, with normal steady-state tolerance of +10%/-6%. TNB describes domestic service as either single-phase two-wire or three-phase four-wire; “two-phase residential supply” is not the normal service classification. If the plan means using two separate phases from a three-phase board, each workstation must still be connected phase-to-neutral at approximately 230 V. It must never be connected phase-to-phase at approximately 400 V.

The two dedicated home runs from the main distribution board are a good reliability choice because a fault or overload on another appliance circuit is less likely to trip a worker. Have a Malaysian licensed electrical contractor design, install and test them rather than specifying breaker or conductor sizes from the server BOM. Dedicated branches do not condition utility power: both machines still share the premises' incoming service, distribution board, neutral and earthing system. The acceptance scope should include:

• one dedicated, labelled final circuit per machine, with no unrelated receptacles or loads;
• compliant over-current and 30 mA residual-current protection for each socket circuit, preferably arranged so one worker's fault does not disconnect the other;
• verified polarity, protective-earth continuity, earth-loop performance and RCD trip time;
• a correctly selected and earthed SPD at the distribution board; the Energy Commission's residential guidance encourages an SPD near the incoming supply for lightning and overvoltage protection;
• a suitable local outlet/isolator and connector for the actual PSU or UPS input, without daisy-chained extension strips or improvised plug adapters; and
• a branch energy/power-quality meter, or equivalent logging, to record voltage range, interruptions and peak current under real GPU workloads.

If the premises later changes to three-phase service, keep each computer as a 230 V single-phase load and place the two dedicated circuits on different phases after the electrician balances the whole property. The phase migration does not require changing the PC PSUs. It may improve fault isolation and whole-site capacity, but it does not eliminate the need for earthing, residual-current protection or surge protection. TNB's published minimum supply scheme allows up to 12 kVA on a 230 V single-phase service. The two workers alone are not a reason to buy a three-phase upgrade; make that decision from measured whole-premises maximum demand, cooling load, planned worker count and the electrician's/TNB's assessment.

Relevant primary guidance is the Energy Commission's residential wiring guide and TNB's supply description.

UPS decision and economic gate#

A UPS would provide two useful capabilities: AVR for supply outside its preferred pass-through band, and battery time to ride through a short interruption or shut Linux down cleanly. It is not required merely because the utility remains within Malaysia's declared supply tolerance: all shortlisted PSUs are universal input designs that include the local 230/240 V nominal service. Site logging must still confirm the actual voltage behavior. A switchboard SPD and correct earth are also more direct protections against incoming transients than paying more for a PC PSU.

Do not use a common 1500 VA/900 W desktop UPS for this machine. Its watt rating can be lower than the worker's possible input demand even though its VA number looks large. A suitable per-machine starting class is approximately 2200 VA with at least 1800 W continuous output, pure-sine output, AVR, USB or network shutdown support, and user-replaceable batteries. The Eaton 5PX 2200i G2 is a concrete reference: 2200 VA/2200 W, line-interactive, pure sine wave, automatic shutdown support, 98% stated full-load efficiency and ten minutes of stated runtime at half load. M-Link's 19 August 2026 dealer promotion listed it at RM3,790.

Two such UPS units would cost RM7,580, about 9.8% of the two RM38,647 workers, before replacement batteries and conversion losses. Amortized over five years, the purchase price alone is approximately RM126 per month for the pair. Sealed lead-acid batteries normally require replacement in roughly three to five years, and APC notes that every 8 °C increase above 25 °C can approximately halve expected battery life. This matters in a warm equipment room.

For this bootstrapped deployment, the recommended sequence is:

1. Install the dedicated protected circuits, switchboard SPD and power-quality logging now.
2. Make the OS/configuration reproducible, keep application state backed up off-machine, and make customer jobs retryable. The local model volume is a disposable cache whether it remains single-drive or later becomes RAID 0.
3. Record at least 30 days of voltage and outage data, including storms, and record the recovery time and lost gross margin from every interruption.
4. Buy UPS capacity when the expected monthly loss from interruptions exceeds its monthly ownership cost, or when paying-customer commitments make graceful shutdown a baseline requirement.
5. If purchasing, prefer one properly sized UPS per machine for fault isolation. Also protect the ONT, router and switch; server runtime has no service value if the network path is dead. Configure automatic job draining and shutdown rather than treating the UPS as long-duration generation.

A single central 5 kVA online UPS may look cheaper than two premium tower units, but it adds a shared failure point, continuous conversion loss, fan noise, special distribution and more complex maintenance. Revisit that architecture only after the deployment becomes a real equipment room with three or more workers or a generator. Online double conversion is not the 80/20 choice unless logging shows persistent out-of-tolerance voltage or frequency that line-interactive AVR cannot handle.

RAM decision#

Keep the included 64 GB kit. A 64 GB host completed the benchmark workflow, and warm inference does not benefit simply from unused system memory. The extracted list priced the available 128 GB kit at RM8,999, approximately RM4,000 more than the 64 GB kits. That is poor value without evidence of production memory pressure.

Monitor process RSS, swap activity, major faults and cache eviction in production. If real workloads regularly exceed available memory or CPU model offload becomes common, replace the two-DIMM kit with a validated 2x64 GB kit. Do not assume adding a second 2x32 GB kit will retain DDR5-6000 stability.

GPU decision#

The RTX 5090 and its 32 GB VRAM are the non-negotiable performance component. At standalone prices, a Gigabyte Windforce OC was RM700 cheaper than the Zotac and would have been a reasonable equivalent GPU class. The promotion instead supplies Palit or Zotac.

Request the Zotac explicitly because the existing worker uses it. Identical GPU models simplify power, thermal and failure comparisons between the two nodes. If IdealTech will not guarantee the brand, either 5090 should execute the workload, but the quote and acceptance record must identify the exact supplied SKU.

Cooler, case and airflow decision#

The HydroShift LCD and O11 Vision are not the independently selected price/performance choices; they are part of the discounted package, and replacing them would erode the package value.

The O11 Vision supports a side-mounted 360 mm radiator. Lian Li lists the HydroShift as supported in this position but notes that its performance is less optimal than a top mount. See the HydroShift compatibility guidance and O11 Vision layout.

Ask the builder to follow the supported side orientation and use an airflow plan equivalent to:

• bottom fans as intake for the GPU and motherboard;
• side radiator as intake, using the supported tube orientation;
• rear fans as exhaust; and
• positive pressure without a closed recirculation loop.

The builder must verify sustained NVMe temperatures because the storage target depends on avoiding Gen5 thermal throttling, not merely on keeping the CPU cool.

Alternatives considered#

Cost-reduced promotional build: RM36,947#

This version would keep the 9850X3D and Antec GSK1000, upgrade only the motherboard for RM450, retain the Kioxia, and add two Samsung drives at RM2,749 each:

RM30,999 + RM450 + RM2,749 + RM2,749 = RM36,947

It is technically sound and RM1,700 cheaper than the Seasonic-based ceiling, RM1,390 cheaper than the provisional Corsair alternative, or RM1,080 cheaper than the provisional Lian Li alternative. It was not selected because the 9950X has broader capability and a 1200 W PSU provides useful continuous-service power margin.

Two MSI M480 2 TB drives#

This was the leading standalone price/performance alternative. It was rejected because it has no application-level validation in the captured dataset. It may be revisited only if the exact physical layout passes the same model-read acceptance test before the return window closes.

One 4 TB Samsung 9100 Pro#

The old rejection is superseded. Later exact-drive and physical-layer controls support beginning with one drive, although the current purchasing constraint now limits each NVMe device to 2 TB. The current question is therefore one versus two 2 TB Samsung devices, with the second justified by capacity or a local A/B test.

Upgrade to 128 GB immediately#

Rejected because it costs approximately RM4,000 more, 64 GB completes the workload, and there is no production memory-pressure evidence yet.

Retain the base motherboard#

Conditionally valid. It has one CPU-connected Gen5 M.2 slot, chipset Gen4 M.2 slots, onboard 5GbE and a Gen5 x16 GPU slot, so it satisfies the initial one-model-drive design and saves RM450. Retain the Tomahawk only while an independent second CPU Gen5 x4 path remains valuable for the unresolved two-drive expansion.

Retain the base CPU and PSU#

Technically acceptable and the best minimum-cost choice. Retaining both saves RM1,600 against the current Seasonic-based quote or RM980 against the provisional Lian Li alternative. It was not selected because the CPU has operational value beyond the current benchmark and the move from the minimum 1000 W PSU to 1200 W adds useful continuous-service margin.

Instructions to put on the IdealTech order#

1. Supply the Zotac RTX 5090 SKU if available; record the exact SKU on the invoice.
2. Quote a separate 2 TB OS SSD, with the MSI M480 Gen4 TLC and Kioxia Exceria Basic Gen4 QLC shown side by side; do not add a Gen5 premium to the OS path.
3. Install one Samsung 9100 Pro 2 TB in CPU-connected M2_1 as the initial model cache.
4. Install the 2 TB OS SSD in a chipset-connected M.2 slot and put Linux, containers, logs, downloads and outputs there.
5. If the Tomahawk is retained, leave M2_2 empty and rear USB4 enabled initially. Add a second Samsung and disable USB4 only after the documented capacity or performance gate is met.
6. Confirm the RTX 5090 still negotiates its full intended x16 link.
7. Use onboard 5GbE. Do not add the TX401 to the initial order.
8. Record the exact PSU model and revision. Use only that PSU's supplied direct 12V-2x6 GPU cable, fully seated, without the multi-8-pin adapter where possible; never reuse modular cables from a different PSU model.
9. Install the Samsung motherboard heat spreader and ensure direct case airflow reaches it.
10. Configure the O11 Vision fans for bottom/side intake and rear exhaust.
11. Do not place the operating system or irreplaceable output data on the local model cache.

Acceptance test before treating the node as production-ready#

The parts list alone does not prove the desired model-loading performance. Validate the completed machine in this order:

1. PCIe topology
   • The initial 9100 Pro negotiates 32.0 GT/s at x4 on a CPU-connected path.
   • The RTX 5090 negotiates its intended x16 width under load.
   • The 2 TB OS SSD is on a chipset path.
2. Direct filesystem and physical reads
   • Put models on a direct XFS or ext4 filesystem on the Samsung, not a buffered loop-backing file.
   • Run the cache-evicted model reader and confirm the Samsung namespace counters rise by approximately the resolved model byte count. Process counters alone do not prove physical-NVMe reads.
   • Record direct and buffered synthetic results as diagnostics; do not restore the withdrawn 10 GiB/s purchase threshold.
3. End to end
   • Run five verified cold/warm ComfyUI pairs.
   • Every cold run must be verified at the process and physical-NVMe layers.
   • Every warm run must record zero model reads.
   • Establish the local machine's distribution before setting the production threshold.
4. Optional second-drive A/B
   • Add the second Samsung only if 2 TB capacity is insufficient or a same-machine single-versus-RAID test is being run.
   • Keep the same model set, filesystem, GPU/CPU settings and physical-read verification. Require a repeated improvement of roughly two seconds on both production workloads, not merely higher fio throughput.
5. Thermals
   • Read at least 200 GB continuously while recording installed NVMe temperatures and link states.
   • Reject or correct a configuration that briefly passes before thermal throttling below the target.
6. Electrical and PSU
   • Record idle, sustained ComfyUI and combined CPU/GPU stress input power at the branch meter.
   • Confirm the dedicated circuit remains stable without abnormal plug/connector heating, nuisance trips or material voltage sag; have the electrician correct the installation rather than increasing breaker size ad hoc.
   • After initial thermal cycling, shut down and isolate mains power before visually checking the native 12V-2x6 connection for full insertion, discoloration or deformation. Follow the PSU and GPU makers' cable-bend guidance; do not repeatedly disconnect a healthy connector without cause.
   • If a UPS is later installed, test its self-test, monitoring and automated job-drain/shutdown path under controlled conditions before relying on its advertised runtime.
7. Network and NAS cache fill
   • Confirm onboard Ethernet negotiates 5GbE full duplex with no increasing CRC or drop counters.
   • Run iperf3 to a RAM- or SSD-backed NAS endpoint and target at least about 4.5 Gbit/s when testing one worker at a time. Test HDD-backed NFS separately so disk and network limits are not confused.
   • Copy the real model set from an empty local cache, verify hashes, and repeat with both workers filling simultaneously. Record the disk, NAS-uplink and worker-link bottlenecks.
   • Simulate loss of the NAS: already-cached jobs must continue, while cache misses wait or fail cleanly.

The already-built first worker should be subjected to the same direct-filesystem and block-layer test. If it already has RAID 0, retain that configuration while measuring it; do not infer that the second worker needs the same drive count. Fix M.2 placement, filesystem/container placement or cooling before buying additional storage for fleet symmetry.

Boundaries of this decision#

• Prices are from an IdealTech extraction and promotional quote dated 2026-09-22; they will change.
• Cross-host end-to-end workflow times remain confounded by CPU, GPU power, kernel, filesystem, container placement and background load. Storage conclusions rely primarily on the dedicated physical-read probes and captured topology.
• The earlier 10 GiB/s threshold has been withdrawn as a purchase requirement. Other model formats and loaders can bottleneck elsewhere, so local end-to-end acceptance remains mandatory.
• RAID 0 is acceptable only as an optional disposable cache with another source of truth; it is not the initial configuration.
• The 9950X and 1200 W PSU are operational-margin decisions, not requirements for current warm ComfyUI inference. The specific PSU model remains subject to the documented price gate.
• RM38,647 is the historical two-Samsung/1 TB-OS Seasonic quote. The current one- versus two-model-drive scenarios with a 2 TB OS SSD must be re-quoted; the provisional totals are RM35,948-RM39,247 depending on board, OS SSD and model-drive count.
• Onboard 5GbE is the initial network. Any future 10GbE decision must be based on measured NAS cache-fill demand and the actual RJ45 or SFP+ fabric.
• UPS hardware, switchboard work, surge protection and branch metering are site-infrastructure costs and are not included in the computer price.
• The final price should be re-quoted in writing with the third drive, exact GPU SKU and exact drive-slot placement before payment.