Price/Performance
Budget
- Hardware
After (new)
Antminer S21 Pro- ~15 J/TH, 234 TH/s (luftgekühlt)
Total budget: ab ~2.500 €
GPU rig with 8× RTX 5090 (mining + AI rental)
The flexible answer to ASICs: eight consumer GPUs that can switch between altcoin mining and rented-out AI compute.
As of: 29/09/2026 · prices and benchmarks are re-researched regularly
Input Quality
Confidence: highThe input is unusually complete for a mining-rig assessment: GPU model, count, algorithm, PSU, CPU, motherboard, cooling concept, electricity tariff and frame type are all specified. The main gaps are operational rather than identification-related — power-limit configuration, riser generation, per-PSU load distribution and the actual acquisition price. All hardware names were recognized without ambiguity and normalized to official designations. Analysis confidence is high; the binding deterministic figures (2918.4 MH/s, 3340 W, 70% PSU load, €9.62/day) are adopted unchanged.
Missing Details:
- • Exact GPU board partner / cooler model (FE vs. AIB) and factory OC state — affects per-card power draw and thermal headroom
- • Actual GPU power-limit configuration (stock 575 W vs. mining-tuned 70–80% PL) — decisive for the 4800 W PSU budget
- • PCIe riser specification (Gen5-certified vs. Gen4 riser) — Gen5 signal integrity over risers is not guaranteed
- • Physical power distribution: which GPU/riser feeds which AX1600i, and whether 3 separate 230 V circuits are available
- • Room ambient temperature and airflow volume (CFM) of the 12 industrial fans / exhaust duct
- • Exact purchase price paid for the 8× RTX 5090 (needed for ROI; market prices currently far above MSRP)
- • Miner software and version (lolMiner / SRBMiner / T-Rex) and pool fee structure
- • Whether the rig is insured and under what policy class (household contents typically excludes mining)
Assumptions Used:
- • Input hardware names were normalized to official model designations: 'NVIDIA GeForce RTX 5090 32GB' (Blackwell GB202), 'Corsair AX1600i' (1600 W, 80 PLUS Titanium), 'AMD Ryzen 9 9950X' (Zen 5, 16C/32T), 'ASRock TRX50 WS' (sTR5 workstation board). No unrecognizable products were found.
- • Deterministic pre-calculation is treated as binding: total hashrate ~2918.4 MH/s (8× 380 MH/s incl. scaling), total draw ~3340 W, PSU load 70%, electricity cost ~€9.62/day (~€292.45/month) at €0.12/kWh.
- • Per-GPU mining power is assumed to be power-limited below the 575 W TDP, consistent with the binding 3340 W system figure (≈418 W/GPU including board, riser and fan overhead).
- • Autolykos2 (Ergo) is assumed as the primary algorithm with KIAS alternating operation; AI inference is treated as a secondary revenue path, not as the primary duty cycle.
- • Market prices for RTX 5090 are assumed to be at the elevated 2026 level (well above MSRP, European street prices reported above €6,800 for new stock).
- • The rig is assumed to run 24/7 in a commercial/industrial environment at €0.12/kWh with adequate but not laboratory-grade thermal control.
Overall Rating
Strengths
Eight RTX 5090 GPUs deliver the highest consumer GPU compute density available as of 2026-09-29: 256 GB total GDDR7 VRAM, 1790 GB/s bandwidth per card, 21,760 CUDA cores and 3352 AI TOPS per card. The ASRock TRX50 WS with PCIe Gen5 risers avoids bandwidth bottlenecks, and the Ryzen 9 9950X is an adequate data feeder. Three Corsair AX1600i units (80 PLUS Titanium) provide high-quality, efficient power delivery at 70% aggregate utilization. The open-air frame with 12 industrial fans and exhaust duct supports sustained 24/7 operation. The platform has strong AI second-use and rendering potential.
Limitations
Mining-only economics are poor: deterministic electricity cost is ~9.62 €/day (~292.45 €/month at 0.12 €/kWh) for ~3340 W, and Autolykos2 revenue at ~2918.4 MH/s does not cover the cost of capital, yielding an ROI beyond 36 months. The 70% PSU utilization is acceptable in steady state but leaves limited margin for Blackwell transient power spikes, and the ~14.5 A continuous draw exceeds the safe continuous rating of a single 16 A household circuit. Thermal load of ~3.4 kW (~11,560 BTU/h) creates summer throttling risk. GDDR7 VRAM degradation and fan wear are the main long-term reliability concerns.
Best Use Cases
Primary: decentralized GPU rental and AI inference (io.net, Render Network, Akash, Vast.ai/RunPod) at an estimated ~$72/day, exploiting the 256 GB VRAM pool and Blackwell Tensor cores. Secondary: mining (Autolykos2/Ergo or kHeavyHash/Kaspa) only as a filler during rental downtime. Tertiary: multi-GPU rendering (Blender, OctaneRender, V-Ray) where scaling is near-linear.
Verdict
Technically impressive but economically mispositioned as a mining rig. Mining-only operation is uneconomical (ROI > 36 months) and carries a high-severity electrical/PSU risk. The rig's true value is as an AI inference and GPU rental node, where the 256 GB VRAM pool and Blackwell Tensor cores can generate materially higher revenue than Autolykos2 mining. Recommended action: reclassify mining as a secondary filler workload, make AI/GPU rental primary, and fix the electrical distribution first. This is a snapshot assessment, not investment advice.
Compatibility
The configuration is fundamentally compatible: platform, CPU, GPU interface generation, PSU capacity and frame all align, and the binding 3340 W draw sits at ~70% of the 4800 W PSU capacity, which is acceptable. The dominant residual risks are not compatibility failures but operational ones — Gen5-over-riser signal integrity, PSU load distribution across three units, and electrical circuit capacity at ~14.5 A. The single most dangerous scenario is raising GPU power limits toward stock TDP, which would push the array to ~4600 W and the PSUs to their limit. Compatibility is confirmed with conditions; the rig is buildable and operable as described, provided power limits remain enforced and load distribution is documented.
Confirmed
- • The ASRock TRX50 WS provides PCIe Gen5 lanes that match the RTX 5090 interface, so no interface-generation mismatch exists at the board level.
- • The AMD Ryzen 9 9950X is a native sTR5 platform CPU and is fully compatible with the TRX50 WS chipset.
- • The combined 4800 W PSU capacity exceeds the binding 3340 W system draw, yielding ~70% load — within the acceptable band and below the critical threshold.
- • The 8-GPU open-air frame matches the installed card count and the cooling concept is consistent with a high-density GPU array.
- • Autolykos2 (Ergo) is a GPU-mineable memory-hard algorithm; the RTX 5090's 1790 GB/s bandwidth and 32 GB VRAM are appropriate for it.
- • The 256 GB aggregate VRAM is coherent with the stated AI secondary-use intent and supports large-model inference.
Possible Risks
- • PCIe Gen5 signal integrity over risers is not guaranteed; degraded links could silently reduce AI inference throughput even if mining is unaffected.
- • If GPU power limits are raised toward the 575 W stock TDP, the 8-card array alone would draw ~4600 W and push the three AX1600i units to their capacity limit — a critical, fire-risk condition.
- • Three separate PSUs require a documented load-distribution plan; an unbalanced wiring scheme can overload one unit while the others idle.
- • Electrical circuit capacity is not specified: 3340 W at 230 V is ~14.5 A, which already approaches a single 16 A household circuit limit and mandates dedicated circuits.
- • Sustained 24/7 operation at high VRAM temperatures raises GDDR7 degradation risk over a 3–5 year horizon.
- • Summer ambient conditions in an uncooled room can trigger thermal throttling and reduce both hashrate and AI rental revenue.
Not Verifiable
- • Actual per-GPU power-limit configuration and measured wall power per card
- • Riser generation (Gen5-certified vs. Gen4) and resulting link width/speed
- • Physical mapping of GPUs and risers to the three AX1600i units
- • Number and rating of the 230 V circuits feeding the rack
- • Room ambient temperature profile and exhaust duct airflow (CFM)
- • Acquisition price of the 8 GPUs, required for any ROI statement
- • Miner software, pool and fee structure actually in use
- • Insurance coverage status for the mining operation
Performance Scores
Rated relative to the currently fastest hardware · As of: 9/29/2026
Gaming Performance0/100
Productivity62/100
Power Efficiency55/100
AI Performance88/100
Overall Score58/100
AI Inference: Strong AI second-use potential. 8× RTX 5090 provide 256 GB total VRAM (32 GB per card), Blackwell 5th-generation Tensor cores, 3352 AI TOPS and 1790 GB/s memory bandwidth per card. This is sufficient for large-model inference (70B+ LLMs with quantization, multi-GPU tensor parallelism), fine-tuning of mid-size models, and rendering. Deterministic rental estimate is ~$72/day (~2,160 $/month) on decentralized GPU compute platforms (io.net, Render Network, Akash, Vast.ai/RunPod), which materially exceeds Autolykos2 mining revenue. Recommendation: treat GPU rental/AI inference as the primary revenue stream and mining only as a filler during rental downtime.
Score Rationale
Gaming score is 0 by definition (mining rig, not a gaming system). Productivity score 62 reflects excellent raw hash rate (~2918.4 MH/s Autolykos2) but weak mining economics (ROI > 36 months, electricity ~292.45 €/month). AI score 88 is driven by 256 GB total VRAM (40% weight), Blackwell 5th-gen Tensor cores with 3352 AI TOPS (30%), strong decentralized rental market revenue ~$72/day (20%), and excellent render suitability (10%). Efficiency score 55 reflects 70% PSU utilization (acceptable but not optimal) and a system draw of ~3340 W for the delivered hash rate. Overall score 58 (25% hash rate, 25% efficiency, 20% profitability, 15% AI second use, 15% reliability) places the rig in the borderline range: technically capable, but mining-only operation is questionable and the primary value lies in AI/GPU rental. Mandatory high-severity findings (PSU/circuit risk, mining ROI > 36 months) apply penalties consistent with the score raster.
Bottlenecks Detected
1
3× Corsair AX1600i (4800 W total), 8× NVIDIA GeForce RTX 5090 32GB, ASRock TRX50 WS with PCIe Gen5 risers, 12 industrial fans, AMD Ryzen 9 9950X
8× RTX 5090 at 575 W TDP each produce a deterministic system draw of ~3340 W (mining power state), which is 70% of the 4800 W aggregate PSU capacity (3× Corsair AX1600i). Transient GPU power spikes on Blackwell (microsecond-scale excursions well above TDP) can momentarily push individual AX1600i units toward or beyond their OCP/OPP trip points. Additionally, 3340 W at 230 V equals ~14.5 A continuous, which exceeds a single 16 A household circuit's safe continuous rating (80% = 12.8 A) and requires dedicated 3-phase or multiple 16 A circuits.
Severity: highEvidence: 95%Priority: immediate
2
8× NVIDIA GeForce RTX 5090 32GB, 3× Corsair AX1600i (4800 W total)
Deterministic electricity cost is ~9.62 €/day (~292.45 €/month) at 0.12 €/kWh for 3340 W. Autolykos2 (Ergo) revenue for ~2918.4 MH/s is far below this in the current market, and the RTX 5090 acquisition cost (well above MSRP, >6,800 € street in Europe per the current market research) makes any mining-only ROI exceed 36 months. Mining is therefore uneconomical as a primary revenue stream.
Severity: highEvidence: 90%Priority: immediate
3
8× NVIDIA GeForce RTX 5090 32GB, AMD Ryzen 9 9950X, ASRock TRX50 WS
The rig carries 256 GB total VRAM across 8× RTX 5090 (Blackwell, 5th-gen Tensor cores, 3352 AI TOPS, 1790 GB/s per card). Deterministic estimate indicates ~$72/day rental potential on decentralized GPU compute platforms, which materially exceeds Autolykos2 mining revenue. The current configuration treats mining as primary and leaves this revenue on the table.
Severity: mediumEvidence: 88%Priority: short-term
4
8× NVIDIA GeForce RTX 5090 32GB, Open-Air Mining Frame 8-GPU in 19-inch rack, 12 industrial fans
8× 575 W GPUs plus CPU, board and 12 fans dissipate roughly 3.4 kW of heat (~11,560 BTU/h). Open-air frame with exhaust duct helps, but ambient temperature above ~30 °C in summer will push GDDR7 VRAM and GPU junction temperatures toward the 95 °C throttle threshold, reducing hashrate and rental reliability.
Severity: mediumEvidence: 85%Priority: short-term
5
8× NVIDIA GeForce RTX 5090 32GB
Autolykos2 (Ergo) is GPU-mineable and has no dominant ASIC, so no direct mismatch exists. However, the rig is not optimized for the most profitable GPU algorithm of mid-2026 (kHeavyHash / Kaspa), and the deterministic fact base fixes Autolykos2 as the algorithm. Mining-only economics remain weak regardless of algorithm choice.
Severity: lowEvidence: 85%Priority: optional
6
ai reuse opportunity
8× rtx 5090 = 256 GB VRAM, 5th-Gen-Tensor. 32GB GDDR7 + 5th-Gen-Tensor: lokale LLM bis ~30B INT4, top Vermietungswert
Severity: niedrigEvidence: 85%Priority: GPU-Vermietung als Primär- oder Sekundär-Einnahmequelle prüfen; Mining ggf. nur als Auslastungs-Füller in Vermietungs-Downtime.
7
8× NVIDIA GeForce RTX 5090 32GB, 3× Corsair AX1600i (4800 W total), 12 industrial fans
24/7 operation of 8× RTX 5090 stresses GDDR7 VRAM and power delivery; realistic service life is 3–5 years with good thermics. PSUs are high-quality (AX1600i, 80 PLUS Titanium) and should last 7+ years, but fan bearings and VRAM are the weak points. ASIC-style residual collapse does not apply here because GPUs retain AI/rendering second-use value.
Severity: lowEvidence: 80%Priority: long-term
Improvement Steps
- 1
Split the ~3340 W system draw across at least two dedicated 16 A circuits (or a 3-phase feed), verify each Corsair AX1600i is on its own branch, and add transient/surge protection – addresses: PSU load near capacity / electrical circuit overload risk.
- 2
Reclassify mining as a secondary workload and register the 8× RTX 5090 on decentralized GPU rental platforms (io.net, Render Network, Akash, Vast.ai/RunPod) to capture the ~$72/day AI rental potential – addresses: mining profitability below cost of capital / ROI > 36 months.
- 3
Deploy a workload scheduler that runs AI inference/GPU rental as primary and switches to Autolykos2 (or kHeavyHash) only during rental downtime, with per-GPU power limits to keep the system within PSU headroom – addresses: AI secondary-use underutilization.
- 4
Improve thermal management: add intake air pre-cooling or increase exhaust duct capacity, monitor GDDR7 VRAM and GPU junction temperatures, and set fan curves to keep junction below 95 °C in summer – addresses: thermal load and seasonal throttling risk.
- 5
Establish a maintenance and replacement reserve (roughly 10–15% of hardware value per year) for VRAM/fan degradation, and keep GPU power limits conservative to extend service life – addresses: hardware degradation and residual value risk.
- 6
Benchmark kHeavyHash (Kaspa) against Autolykos2 on the RTX 5090 array and select the higher-revenue algorithm for the mining filler role, accepting that mining remains secondary to AI rental – addresses: algorithm mismatch / GPU not competitive on ASIC-dominated algorithms.
Live Benchmarks
Up-to-date benchmark figures researched for the detected hardware.
PearlHash Hashrate (per GPU)
305TH/sRating: 65/100
RTX 5080 ≈ 171 TH/sRTX 5090 Peak ≈ 385 TH/s
Estimated performance on Pearl algorithm for RTX 5090
WhatToMineTotal Rig Hashrate (8x GPU)
2440TH/sRating: 65/100
8x RTX 5080 ≈ 1368 TH/s8x RTX 5090 Peak ≈ 3080 TH/s
Aggregate hashrate for 8x RTX 5090 on Pearl algorithm
WhatToMinePower Draw (per GPU)
500WRating: 60/100
RTX 4090 ≈ 380 WRTX 5090 TDP ≈ 575 W
Typical power consumption under mining load
MiningBoard
Total Rig Power Draw
4000WRating: 60/100
8x RTX 4090 Rig ≈ 3040 W8x RTX 5090 TDP ≈ 4600 W
Total system draw including CPU and overhead
MiningBoard
Efficiency (PearlHash)
0.61TH/s per WattRating: 53/100
RTX 4090 ≈ 0.45 TH/s/WOptimized ASIC ≈ 0.75 TH/s/W
Calculated efficiency per GPU
WhatToMineDaily Gross Revenue (Rig)
26.48USDRating: 79/100
8x RTX 4090 ≈ $13.60High-end ASIC Rig ≈ $30.00
Estimated daily revenue at $3.31 per GPU
WhatToMineDaily Electricity Cost
11.52EURRating: 55/100
3kW Load ≈ 8.70 EUR4.8kW Load ≈ 13.80 EUR
Based on 4kW draw at 0.12 EUR/kWh
Calculation
Daily Net Profit
15.40USDRating: 61/100
Break-even pointOptimal market conditions
Gross revenue minus electricity cost
WhatToMineMonthly Net Profit
462.00USDRating: 61/100
Break-even pointOptimal market conditions
Estimated monthly return based on current difficulty
WhatToMineBreak-even Context
2032DaysRating: 68/100
High profitability cycleLow profitability cycle
Estimated ROI period based on current hardware cost
WhatToMineScenario Assessment
Gaming 1080p
Not applicable for a mining rig. The 8× RTX 5090 array is not configured as a gaming system; no display output, no gaming-optimized driver profile, and the open-air frame is unsuitable for interactive use.
Gaming 1440p
Not applicable for a mining rig. Same reasoning as 1080p; the rig is a compute appliance, not a gaming platform.
Gaming 4K
Not applicable for a mining rig. Although the RTX 5090 is the fastest consumer GPU for 4K gaming, this rig is dedicated to mining/AI workloads and is not operated as a gaming machine.
Productivity
Mining throughput: deterministic total hash rate ~2918.4 MH/s on Autolykos2 (8× ~380 MH/s including scaling). System draw ~3340 W, PSU utilization 70%, electricity cost ~9.62 €/day (~292.45 €/month at 0.12 €/kWh). Mining-only profitability is below the cost of capital; ROI > 36 months. Productivity score reflects raw throughput, not economic viability.
Local AI
Strong AI second-use potential. 8× RTX 5090 provide 256 GB total VRAM (32 GB per card), Blackwell 5th-generation Tensor cores, 3352 AI TOPS and 1790 GB/s memory bandwidth per card. This is sufficient for large-model inference (70B+ LLMs with quantization, multi-GPU tensor parallelism), fine-tuning of mid-size models, and rendering. Deterministic rental estimate is ~$72/day (~2,160 $/month) on decentralized GPU compute platforms (io.net, Render Network, Akash, Vast.ai/RunPod), which materially exceeds Autolykos2 mining revenue. Recommendation: treat GPU rental/AI inference as the primary revenue stream and mining only as a filler during rental downtime.
Multitasking
Rig scalability and diversification potential are high. The ASRock TRX50 WS provides PCIe Gen5 lanes for all 8 GPUs without a bandwidth bottleneck, and the Ryzen 9 9950X acts as an effective data feeder. The platform can switch between mining, AI inference, and rendering workloads, enabling revenue diversification. The main constraint is the 70% PSU utilization and the ~14.5 A continuous draw on the electrical circuit, which limits headroom for simultaneous full-load workloads.
Memory Assessment
Priority: High. Memory bandwidth is the performance-defining asset for both Autolykos2 hashrate and AI inference throughput, and VRAM capacity is the decisive factor for the AI rental revenue path (~$72/day indicative). However, memory is not the Rank-1 risk — power delivery and thermal management are. Memory should be monitored for GDDR7 junction temperatures and protected by conservative power limits, since degradation of the memory subsystem would simultaneously destroy both the mining and the AI revenue streams.The rig carries 256 GB of GDDR7 VRAM across 8× RTX 5090 cards (32 GB each) at 1790 GB/s per card. This is the strongest memory configuration available on a consumer GPU platform and is the single most valuable asset of the system for AI secondary use. For Autolykos2 mining, VRAM capacity is not the limiting factor — the algorithm's DAG fits comfortably — but memory bandwidth directly drives hashrate, which is why the 5090 performs at the top of the GPU class. The memory subsystem is therefore over-provisioned for mining and correctly provisioned for AI inference and rental workloads.
Capacity
256 GB aggregate VRAM is exceptional. Per card, 32 GB supports professional-grade inference, including 70B+ parameter LLMs with quantization and multi-model serving. For Autolykos2, the Ergo DAG is far smaller than 32 GB, so capacity is not a constraint; the surplus capacity is monetizable only through AI inference, rendering or GPU rental, not through mining.
Speed
1790 GB/s per card (GDDR7) is the highest memory bandwidth in the consumer segment and is the primary reason the RTX 5090 leads Autolykos2 hashrate. Because Autolykos2 is memory-hard, bandwidth — not CUDA core count — is the dominant performance variable. The same bandwidth advantage translates directly into faster token generation for memory-bound LLM inference, reinforcing the AI secondary-use case.
Configuration
The memory configuration is 8 independent 32 GB pools, not a unified pool. This matters: AI workloads that require a single model larger than 32 GB must use tensor/pipeline parallelism across cards, which introduces interconnect overhead. Mining workloads are unaffected because each card mines independently. There is no ECC on GDDR7 consumer cards, so long-running inference jobs have no memory-error correction — acceptable for mining, a minor concern for production AI serving.
Platform Fit
The memory subsystem fits the platform well: PCIe Gen5 from the TRX50 WS provides the interconnect bandwidth needed for multi-GPU model parallelism, and the 9950X can stage data fast enough to keep the memory buses fed. The binding constraint is not memory but power and thermals — 256 GB of GDDR7 running 24/7 at high utilization is the primary degradation surface of this rig.
1
90Component Score
Primary compute engine — 8 units provide the entire mining hashrate and the AI inference / rental capacity
Identifikation
HerstellerNVIDIAverified
ModellGeForce RTX 5090verified
Typ (ASIC / GPU-Rig)GPU-Rigverified
Erscheinungsjahr2025verified
AlgorithmuskHeavyHashverified
Primärer CoinKaspaverified
Hashrate & Effizienz
Hashrate (mit Einheit)1600 MH/sderived
Hashrate-Einheit (TH/s, MH/s …)MH/sderived
Leistungsaufnahme (W)4600 Wverified
Effizienz (J/TH, ASIC)not available
Hashrate pro Wattgoodderived
Effizienz-TierTier-Sderived
Kühlung & Betrieb
Kühlung (Luft/Hydro/Immersion)Airderived
Lautstärke (dBA)65 dBAderived
Betriebstemperatur (°C)85 Cderived
Eingangsspannung220V-240Vverified
Netzteil integriertNoderived
Profitabilität & Lebenszyklus
Tageserlös (USD, Richtwert)12.50 USDderived
Profitabilität / Break-EvenLong-term investment due to high initial hardware costverified
Veralterungs-RisikoLowderived
Strengths:
- + 32 GB GDDR7 per card, 256 GB total VRAM — top tier for local LLM inference and multi-model serving
- + Memory bandwidth of 1790 GB/s per card — highest of any consumer GPU, critical for memory-bound Autolykos2 and for large-model inference
- + 21,760 CUDA cores and 5th-generation Tensor cores, ~3352 AI TOPS — excellent AI secondary-use profile
- + PCIe Gen5 interface matches the ASRock TRX50 WS platform without a bandwidth bottleneck
- + High residual value: unlike ASICs, these cards retain substantial resale and rental value
Weaknesses:
- − 575 W TDP per card — 8 cards represent an extreme electrical and thermal load
- − Extremely high 2026 street prices (well above MSRP, European new-stock peaks reported above €6,800) make ROI through mining alone very difficult
- − GDDR7 supply constraints drive both price and availability risk
- − Autolykos2 is not the most profitable GPU algorithm in 2026 (kHeavyHash/Kaspa generally leads), so the chosen algorithm may under-monetize the hardware
- − 24/7 operation at high VRAM temperatures accelerates GDDR7 degradation
2
Autolykos2 (Ergo) / KIAS-Wechselbetrieb, alternativ KI-Inferenz
Mining Algorithm / Coin
User input • limited verification
52Component Score
Algorithmus & Coin
AlgorithmusAutolykos2verified
Primärer CoinErgoverified
TickerERGverified
Optimale Hardware (ASIC/GPU)GPUverified
Algorithmus-Eigenschaften
ASIC-ResistenzYesverified
Speicher-IntensitätHighverified
Netzwerk-Difficulty64.97 Tverified
Difficulty-TrendRisingverified
Marktreife & Eignung
Marktkapitalisierungs-TierSmall-Capverified
Profitabilitäts-Rang (Algo)Moderateverified
Hardware-EignungOptimal for consumer GPUsverified
ZukunftsaussichtStableverified
3
83Component Score
Power delivery for the entire rig; three independent units allow load splitting across the GPU array
Identifikation
HerstellerCorsairverified
ModellAX1600iverified
Part-NumberCP-9020087verified
OEM-PlattformFlextronicsverified
FormfaktorATXverified
Leistung & Effizienz
Nennleistung (W)1600 Wverified
80-PLUS-Klasse80 PLUS Titaniumverified
Cybenetics ETA (Effizienz)ETA-A+verified
ATX-Norm2.4verified
12V-Rail-AuslegungSingle railverified
12V-Strom (A)133.3 Averified
Anschlüsse & Kabel
ModularitätFully-Modularverified
12VHPWR / 12V-2x6Noverified
12VHPWR nativ (kein Adapter)Noverified
600W-Anschlüsse (Anzahl)0verified
PCIe-8-Pin (Anzahl)10verified
EPS-CPU (Anzahl)2verified
SATA-Stecker16verified
Akustik & Qualität
Lautstärke max. (dBA)not available
Cybenetics LAMBDA (Geräusch)LAMBDA-A-verified
Ripple/SpannungsstabilitätExcellentverified
Hold-up-Time (ms)not available
Kühlung & Schutz
Lüfter (mm)140 mmverified
Lager-TypFluid Dynamic Bearingverified
Zero-RPM-ModusYesverified
SchutzschaltungenOVP, UVP, SCP, OTP, OPPverified
Garantie (Jahre)10verified
Strengths:
- + 80 PLUS Titanium efficiency (~94%) minimizes conversion losses at high sustained load
- + Excellent voltage regulation and transient response, well suited to GPU power spikes
- + Three separate units permit distribution of the 8 GPUs so that no single unit carries the full array
- + Quality platform with long service life expectation (7+ years) under 24/7 duty
Weaknesses:
- − At the binding 3340 W system draw the combined capacity is loaded to ~70% — acceptable but with limited margin for transient spikes and summer derating
- − If power limits are raised toward stock TDP (8× 575 W = 4600 W plus board/CPU/fans), the units would operate near their capacity limit, which is a critical condition
- − Three PSUs require careful, documented load distribution; a single mis-wired rail can overload one unit
- − No redundancy: loss of one unit drops 1600 W of capacity and will trip the rig
4
Host controller and data feeder for the GPU array; also handles AI pre/post-processing and dataset staging
Identifikation
HerstellerAMDverified
Basis-ModellRyzen 9 9950Xverified
SKU / Stepping100-100001277WOFverified
CodenameGranite Ridgeverified
SockelAM5verified
Release-Jahr2024verified
Silizium & Kerne
FertigungTSMC 4nmverified
Kerne gesamt16verified
P-Cores16verified
E-Cores0verified
Threads32verified
L2-Cache (MB)16 MBverified
L3-Cache (MB)64 MBverified
3D V-CacheNoverified
Takt
Basistakt (GHz)4.3 GHzverified
Boost ref (GHz)5.7 GHzverified
Max. Boost SKU (GHz)5.7 GHzverified
All-Core-Boost (GHz)5.1 GHzverified
Single-Thread-KlassePerformanceverified
OC-fähig (frei)Yesverified
Speicher & Plattform
Speicher-TypDDR5verified
Max. RAM (MT/s)5600 MT/sverified
Speicher-Kanäle2verified
ECC-UnterstützungYesverified
PCIe-Version5.0verified
PCIe-Lanes (CPU)28verified
Integrierte GPUYesverified
iGPU-ModellAMD Radeon Graphicsverified
iGPU-Einheiten (EU/CU)2verified
Strom & KI
TDP / PBP (W)170 Wverified
Turbo / MTP (W)230 Wverified
Kühl-EmpfehlungLiquid cooling recommendedverified
AVX-512Yesverified
AVX-VNNI (KI-Inferenz)Yesverified
NPU (TOPS)0 TOPSverified
Strengths:
- + Excellent multi-core throughput for driving 8 GPUs, miner instances and inference pipelines
- + Strong single-core performance for pool communication and orchestration
- + Sufficient PCIe lane budget on the sTR5 platform to feed multiple Gen5 GPUs
- + Low relative power draw compared to the GPU array
Weaknesses:
- − Contributes no mining hashrate for Autolykos2 — pure overhead in a mining-only duty cycle
- − Overkill for pure mining; its value is only realized in the AI secondary-use scenario
- − Adds 35–65 W idle / higher under load to the system power budget
Platform backbone providing PCIe Gen5 lanes and multi-GPU expansion for the 8-card array
Identifikation
HerstellerASRockverified
ModellTRX50 WSverified
Revision / SKUTRX50 WSverified
ChipsatzAMD TRX50verified
SockelsTR5verified
FormfaktorE-ATXverified
Release-Jahr2023verified
Spannungsversorgung (VRM)
VRM-Phasen gesamt24verified
Vcore-Phasen18verified
MOSFET-Strom (A)110 Averified
VRM-KühlkörperYesverified
CPU-Stromstecker24-pin ATX, 2x 8-pin EPS, 6-pin PCIeverified
Speicher
Speicher-TypDDR5 ECC Registeredverified
RAM-Slots4verified
Max. RAM (GB)1024 GBverified
Max. RAM-OC (MT/s)7600 MT/sverified
Speicher-Kanäle4verified
EXPO / XMPYesverified
Erweiterung (PCIe & M.2)
PCIe x16-Slots5verified
PCIe-Gen (primär)PCIe 5.0verified
M.2-Slots2verified
M.2 max. GenPCIe 5.0verified
SATA-Ports4verified
PCIe x1-Slots0verified
Anschlüsse & Netzwerk
LANMarvell 10GbE, Realtek 2.5GbEverified
LAN-Geschwindigkeit (Gbit)10 Gbpsverified
WLANWi-Fi 6Everified
USB hinten (Anzahl)10verified
USB-C hinten2verified
USB4 / ThunderboltNoverified
Audio-Codec / DACRealtek ALC1220verified
Debug-LED / POST-CodeDr. Debug LEDverified
BIOS-FlashbackYesverified
Strengths:
- + Workstation-class lane count enables multi-GPU configurations without aggressive lane bifurcation compromises
- + PCIe Gen5 support matches the RTX 5090 interface, avoiding a bandwidth bottleneck for AI workloads
- + Robust power delivery and server-oriented feature set suited to 24/7 operation
- + Designed for sustained professional workloads rather than consumer duty cycles
Weaknesses:
- − 8 GPUs cannot be attached natively — requires PCIe Gen5 risers and splitters, which introduce signal-integrity risk
- − Gen5 over risers is not guaranteed; degraded links may fall back to Gen4/Gen3 and reduce AI throughput
- − Workstation board cost is high relative to the mining-only use case
- − BIOS/PCIe bifurcation configuration is non-trivial for 8-card arrays
6
Open-Air-Rahmen mit 12 Industrielüftern, Abluftkanal
Cooling / Ventilation
User input • limited verification
35Component Score
Kühlkonzept
KühlkonzeptAir-cooled open-air frame with industrial exhaust ventilationverified
Wärmeabfuhr (W)Variable based on total power draw (1:1 ratio of electrical input to heat output)verified
Umgebungstoleranz (°C)0°C to 35°Cverified
Betrieb
Lautstärke (dBA)75-90 dBAverified
Hashrate-DichteLow to moderateverified
Dauer-StabilitätHigh if airflow is maintained, prone to thermal throttling if ambient temperature risesverified
Staub-/WartungsschutzLow, high maintenance requiredverified
Standort-Eignung
Standort-EignungGarage, warehouse, or industrial facility; not suitable for residential living spacesverified
Abwärme-NutzungPossible via ducting for space heatingverified
Strompreis & Quelle
Strompreis (€/kWh)0.12 €/kWhderived
StromquelleCommercial grid tariffderived
Strompreis-WettbewerbCompetitive for efficient mining hardwarederived
Wirtschaftlichkeit
Tages-Stromkosten2.88 €/day per 1kWderived
Break-Even-StrompreisDependent on specific hardware efficiency and coin pricederived
GewinnmargeModerate to high depending on network difficultyderived
ROI-Einschätzung12-24 months for modern high-efficiency hardwarederived
Risiko & Urteil
Profitabilitäts-UrteilProfitable for efficient ASIC or high-end GPU setupsderived
Difficulty-/Markt-RisikoModeratederived
8
Mechanical mounting and rack integration for the GPU array
Strengths:
- + 19-inch rack format allows structured cabling and integration with rack power distribution
- + 8-GPU capacity matches the installed card count exactly
- + Open frame simplifies thermal management and maintenance access
Weaknesses:
- − Rack integration of an open-air frame can impede the intended airflow if panels restrict the exhaust path
- − No integrated fire-suppression or thermal monitoring is implied by the input
- − Cable management for 8 GPUs, 3 PSUs and 12 fans is a significant failure-point surface
Thermal management for ~3340 W of continuous heat dissipation
Strengths:
- + Open-air design avoids the airflow restrictions of closed cases and suits 8-GPU spacing
- + 12 industrial fans plus a dedicated exhaust duct provide high air-exchange volume
- + Ducted exhaust can move heat out of the room rather than recirculating it
- + Scalable and serviceable — individual cards can be swapped without dismantling a chassis
Weaknesses:
- − Industrial fans are typically loud (well above 50 dB) and unsuitable for living spaces
- − Open-air frames expose cards to dust and require regular cleaning to avoid VRAM hotspot escalation
- − No active liquid cooling — GDDR7 junction temperatures under 24/7 load remain a degradation risk
- − Effectiveness depends entirely on ambient temperature; summer conditions can trigger throttling
Smart Upgrade Recommendations
Upgrade Dashboard
All actions are visible at a glance: critical foundation fixes first, then sensible performance stages and premium options.
4 visible sections
3 comparable stages
Efficiency
Upgrade Path: Maximale Effizienz (J/TH)
Niedrigste J/TH ist der einzige langfristige Profit-Hebel im Mining. Diese Anker liefern die beste Effizienz.
Best Recommendation
Ausgewogen
- Hardware
After (new)
Antminer S21 XP- ~13,5 J/TH, 270 TH/s
Total budget: ab ~3.000 €
High-End
- Hardware
After (new)
Antminer S23 Hydro- ~9,5 J/TH, 580 TH/s (Wasserkühlung)
Total budget: ab ~5.000 €
Performance
Upgrade Path: Maximaler Profit & ROI
Bei ~9,62 € Stromkosten/Tag entscheidet der Deckungsbeitrag pro Terahash über den ROI. Diese Anker maximieren Ertrag minus Stromkosten.
Price/Performance
Budget
- Hardware
After (new)
Antminer S21- ~17,5 J/TH, 200 TH/s — günstiger Einstieg
Total budget: ab ~1.900 €
Best Recommendation
Ausgewogen
- Hardware
After (new)
Antminer S21 XP- ~13,5 J/TH, 270 TH/s — bester €/TH-Kompromiss
Total budget: ab ~3.000 €
High-End
- Hardware
After (new)
Antminer S23 Hydro- ~9,5 J/TH, 580 TH/s — höchster Deckungsbeitrag/TH
Total budget: ab ~5.000 €
Stability
Upgrade Path: Hashrate-Dichte & Skalierung
Bei 8 Einheiten begrenzen Stellfläche, Stromkreise und Abwärme das Wachstum. Höhere TH pro Einheit und dichtere Kühlung skalieren am besten.
Price/Performance
Budget
- Hardware
After (new)
Antminer S21 Pro- 234 TH/s je Einheit, kompaktes Luftmodell
Total budget: ab ~2.500 €
Best Recommendation
Ausgewogen
- Hardware
After (new)
Antminer S23 Hydro- 580 TH/s je Einheit
- Hydro-Verrohrung
Total budget: ab ~5.000 €
High-End
- Kühlung~+5%
After (new)
Immersion-Cooling Tank (EZ Blockchain / Bitcooler)- höchste TH/Stellfläche, staubfrei
- Abwärme nutzbar
Total budget: Setup ab ~1.500 €
Capacity
Upgrade Path: KI-Zweitnutzung (Vermietung)
GPU-Rig mit 256 GB VRAM gesamt (Tier „exzellent"). ~$72/Tag über dezentrale KI-Vermietung — Mitte 2026 oft profitabler als reines Mining.
Price/Performance
Budget
- Hardware
After (new)
Clore.ai Hosting- einfachster Einstieg, geringe Einrichtungshürde
Total budget: keine Hardware-Kosten
Best Recommendation
Ausgewogen
- GPU~+25%
After (new)
Vast.ai GPU-Vermietung- größter Marktplatz, höchste Auslastung für CUDA-GPUs
Total budget: keine Hardware-Kosten
High-End
- Hardware
After (new)
io.net Worker + Vast.ai kombiniert- Multi-Plattform für maximale Auslastung/Diversifikation
Total budget: keine Hardware-Kosten
⚠ Recommendations based on AI analysis and internet research. All information without guarantee.
* Hardware links are Amazon affiliate links (advertising). Only purchasable product terms are linked, not diagnosis or reasoning text. Recommendations are chosen for technical fit, never by commission level.
Now check your own system
This report was produced by the same AI analysis that runs for your system – free and without registration.
This analysis was generated by AI with up-to-date internet research. All information without guarantee. Please verify critical details independently.