Created bymario-alka.dePowered bygodcore.denoob2claw.detricoma.de
Contributed byMario Alkamicrosoft

Phi-3.5-mini-instruct

Performance benchmark · measured on 07.10.2026 23:19

Benchmark-IDrun-20261008-105917-69e251
Timebench 3 - Kombi (Prefill + Generation)Dense3.8BRuntime: llama.cppQuantisierung: Q4_K_M
ℹ
Public note: Dell PowerEdge R820 . 2x AMD Radeon RX 7900 XTX (48 GB VRAM gesamt, ROCm/HIP) . RAM: 32x 16 GB SK Hynix DDR3-1600 ECC RDIMM (HMT42GR7AFR4A-PB) = 512 GB @ 1600 MT/s
Generation172,34tok/s
Prefill5.225,59tok/s
Time to First Token11.907,50ms
Total duration82,99s
Concurrency5parallel
Ranking in the field
735of 1469 systems

Performance benchmark · Primary metric: Generation-Speed (tok/s) · 5× concurrent

This run is better than 50 % of all comparable systems.
Generation 172,3 tok/s
-30 % vs Ø 246,5
Prefill 5.225,6 tok/s
-3 % vs Ø 5.365,8
Time to First Token 11.908 ms
-61 % vs Ø 30.155
Distribution in the field0 – 1.349 tok/s
Ø 247 Median Dieser Lauf

Wie schlägt sich dieser Benchmark mit anderen Modellen?

gemma-4-E2B-itNVIDIA RTX PRO 6000 Blackwell Workstation Edition · run-20260727-181606-0adf6d
1.349,3 tok/s
gemma-4-E2B-itNVIDIA GeForce RTX 5090 · run-20260728-184455-5b4937
1.301,9 tok/s
gemma-4-E2B-it3× 3x NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition · run-20260728-184456-4111a9
1.195,7 tok/s
Nemotron-3-Nano-4BNVIDIA RTX PRO 6000 Blackwell Workstation Edition · run-20260727-181607-30d864
1.188,7 tok/s
Nemotron-3-Nano-4BNVIDIA GeForce RTX 5090 · run-20260728-194135-5432cd
1.164,8 tok/s
gpt-oss-20bNVIDIA RTX PRO 6000 Blackwell Workstation Edition · run-20260728-140954-9e81f9
1.136,3 tok/s
gpt-oss-20bNVIDIA RTX PRO 6000 Blackwell Workstation Edition · run-20260730-035052-09627f
1.135,0 tok/s
Nemotron-3.5-Lightning-30B-A3BNVIDIA GeForce RTX 3090 Ti · run-20261021-093815-b22926
1.135,0 tok/s
gpt-oss-20bNVIDIA GeForce RTX 5090 · run-20260729-032121-058a31
1.113,5 tok/s
Nemotron-3-Nano-4B3× 3x NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition · run-20260728-194136-836a91
1.090,6 tok/s
Nemotron-3-Nano-4B3× 3x NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition · run-20260727-181606-8b381e
1.077,4 tok/s
Nemotron-3-Nano-4B3× 3x NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition · run-20260728-194136-b8a818
1.071,0 tok/s
Nemotron-3-Nano-4B3× 3x NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition · run-20260728-194136-038e92
1.051,1 tok/s
gpt-oss-20b3× 3x NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition · run-20260729-032121-ffb18a
1.015,5 tok/s
Phi-3.5-mini-instruct this run2× 2x AMD Radeon RX 7900 XTX · run-20261008-105917-69e251
172,3 tok/s

Hardware

GPU: 2x AMD Radeon RX 7900 XTX · 24 GB VRAM
CPU: 4x Intel(R) Xeon(R) CPU E5-4657L v2 @ 2.40GHz
RAM: 504 GB
Mainboard: Dell Inc. PowerEdge R820

Setup

Runtime: llama.cpp
Quantization: Q4_K_M
Model: Phi-3.5-mini-instruct

Configuration

benchmark-konfiguration — run-20261008-105917-69e251
# LLM-Benchmark Konfiguration # Modell : Phi-3.5-mini-instruct # Engine : llama.cpp # Run-ID : run-20261008-105917-69e251 # GPU : 2x AMD Radeon RX 7900 XTX # CPU : 4x Intel(R) Xeon(R) CPU E5-4657L v2 @ 2.40GHz # RAM : 504 GB bench@llm-benchmark:~$ llama-server \ -m Phi-3.5-mini-instruct-Q4_K_M.gguf \ --alias Phi-3.5-mini-instruct \ -ngl 999 \ -sm layer \ -fa on \ -c 40960 \ -np 10 '(2x' RX 7900 XTX 'ROCm/HIP)'
Engine?Die Inferenz-Software, die das Modell ausliefert (z.B. vLLM oder llama.cpp). Sie bestimmt Geschwindigkeit, unterstuetzte Modellformate und welche Parameter ueberhaupt verfuegbar sind.llamacpp
Modellalias?Der Name, unter dem das Modell ueber die API angesprochen wird. Genau dieser Wert muss im Request-Feld 'model' stehen.Phi-3.5-mini-instruct
Kontextlaenge?Maximale Anzahl Tokens (Eingabe + erzeugte Ausgabe zusammen), die das Modell pro Anfrage verarbeiten kann.40960
Modellpfad?Pfad zur GGUF-Modelldatei, die geladen und ausgeliefert wird.Phi-3.5-mini-instruct-Q4_K_M.gguf
GPU-Layer?Anzahl der auf die GPU ausgelagerten Modell-Layer. Hoeher = mehr VRAM und schneller; der Rest laeuft auf der CPU. 999 = alles auf GPU.999
Split-Mode?Verteilung ueber mehrere GPUs: none (nur eine GPU), layer (Layer aufteilen) oder row (Tensoren zeilenweise).layer
faon
Kontext?Groesse des Kontextfensters in Token. 0 = der beim Training verwendete Kontext des Modells.40960
np10
execution_typelocal

All benchmarks of this model To leaderboard

Anzeige
Model comparison

Phi-3.5-mini-instruct on various hardware

All published performance runs of this model – each bubble a variant: position = prefill (X) × generation (Y), bubble size = number of runs. Closer to the top right = faster. ★ Marked gold = this benchmark.

GPUby graphics card

3323173022872728.0128.3538.6949.035Prefill (tok/s)Generation (tok/s)AMD Radeon RX 7900 XTX - 302,0 tok/s Generation, 8.523 tok/s Prefill, TTFT 9.068 ms (6 Laufe) | DIESER LAUF★ AMD Radeon RX 7900 XTX
★ AMD Radeon RX 7900 XTX 302,0 tok/s this run

CPUby processor

3383132872622376635.90311.14416.384Prefill (tok/s)Generation (tok/s)AMD Ryzen Threadripper PRO 3955WX 16-Cores - 302,0 tok/s Generation, 13.699 tok/s Prefill, TTFT 1.150 ms (3 Laufe)AMD Ryzen Threadrippe...Intel(R) Xeon(R) CPU E5-4657L v2 @ 2.40GHz - 272,9 tok/s Generation, 3.348 tok/s Prefill, TTFT 16.986 ms (3 Laufe) | DIESER LAUF★ Intel(R) Xeon(R) CPU ...
AMD Ryzen Threadripper PRO 3955WX 16-Cores 302,0 tok/s★ Intel(R) Xeon(R) CPU E5-4657L v2 @ 2.40GHz 272,9 tok/s this run

MBby mainboard

3383132872622376635.90311.14416.384Prefill (tok/s)Generation (tok/s)ASUSTeK COMPUTER INC. Pro WS WRX80E-SAGE SE WIFI - 302,0 tok/s Generation, 13.699 tok/s Prefill, TTFT 1.150 ms (3 Laufe)ASUSTeK COMPUTER INC....Dell Inc. PowerEdge R820 - 272,9 tok/s Generation, 3.348 tok/s Prefill, TTFT 16.986 ms (3 Laufe) | DIESER LAUF★ Dell Inc. PowerEdge R...
ASUSTeK COMPUTER INC. Pro WS WRX80E-SAGE SE WIFI 302,0 tok/s★ Dell Inc. PowerEdge R820 272,9 tok/s this run

ENGby engine

3383132872622376635.90311.14416.384Prefill (tok/s)Generation (tok/s)vLLM - 302,0 tok/s Generation, 13.699 tok/s Prefill, TTFT 1.150 ms (3 Laufe)vLLMllama.cpp - 272,9 tok/s Generation, 3.348 tok/s Prefill, TTFT 16.986 ms (3 Laufe) | DIESER LAUF★ llama.cpp
vLLM 302,0 tok/s★ llama.cpp 272,9 tok/s this run

DRVby driver

3323173022872728.0128.3538.6949.035Prefill (tok/s)Generation (tok/s)unbekannt - 302,0 tok/s Generation, 8.523 tok/s Prefill, TTFT 9.068 ms (6 Laufe)unbekannt
unbekannt 302,0 tok/s
💰 Economics

Economics of this run

Operating cost, TCO and comparison with the next-best runs of the same model at identical concurrency (5× concurrent). Methodology →

⚙️ ConfigurationAll metrics and charts below follow these settings – based on a 24-month runtime.Save to URLReset
⚡ Electricity price EUR/kWh
⚙️ System utilization 100 %
🖥️ Acquisition EUR
🔌 Idle 35 W
⚡ TDP 749 W
☁️ External LLM (API)
Electricity0.30 EUR/kWh
Avg power (incl. idle)749 W estimated (TDP)GPU 710 + CPU 29 + Board 10 W full load
Avg cost / hourEUR 0.22
Electricity / 1M tokensEUR 0.36
Token / kWh828.61K
Acquisition (system)EUR 4,353 partial priceGPU EUR 2,098 · CPU EUR 59 · RAM EUR 2,016 · PSU EUR 180
Electricity (2 years)–
TCO (2 years)EUR 8,288
Output tokens (2 years)10.87B
☁️ External LLM (API) – comparison
External LLM cost (2 years)–
Savings vs. external (2 years)–

All values above and the charts below take the configured system utilization into account: at X% the system generates only X% of the time, the rest it idles (35 W). Cost per hour drops (more idle), cost per token rises.

Cost over 2 years – electricity only

Cost over 2 years – incl. acquisition (TCO)

Speed vs. tokens per euro

Euro per 1M tokens

Comparison vs. API – economics per benchmark

Phi-3.5-mini-instruct2x AMD Radeon RX 7900 XTXPhi-3.5-mini-instruct2x AMD Radeon RX 7900 XTX
Electricity cost (24 mo.)––
Acquisition cost––
Total cost (TCO)––
Generated tokens (24 mo.)––
Token price via API––
Break-even point (days)––
Result (savings / extra cost)––

Comparison with up to 3 next-best runs of this model at the same concurrency (at least one on different hardware). Power = GPU TDP + CPU (idle + 15 %) + board (estimated), acquisition = full system (GPU + CPU + board + RAM + PSU), prices = stored market prices.

Contributed by

Mario Alka Administrator

@marioalka

Ich bin Unternehmer, Softwareentwickler und KI-Enthusiast. Seit vielen Jahren entwickle ich Unternehmenssoftware und beschäftige mich inzwischen fast täglich mit lokalen LLMs, KI-Agenten und leistungsfähiger KI-Hardware.

Mit LLM-Benchmark.de möchte ich eine Plattform schaffen, auf der Modelle, GPUs und Agenten objektiv und reproduzierbar miteinander verglichen werden.