GPU Registry
This is the canonical answer to “do we have this card, and where is it?”. It reconciles the benchmark fleet and the relic shelf, and counts the one card recorded in both places once.
Click any column header to sort. Use the filters to narrow by machine, architecture, project, or status. The machine column shows every recorded location, so the shared GTX 960 appears as “Tortuga + Relic shelf”.
50owned cards
53GPU devices
32benchmark fleet
8unassigned
19relic shelf cards
22relic devices
| Card | Machine | Cards | Devices | VRAM | Bandwidth | Arch | CC | Status | Projects | Role |
|---|---|---|---|---|---|---|---|---|---|---|
| GTX 1060 6 GB | Colmena | 3 | 3 | 6 GB | ~192 GB/s | Pascal | 6.1 | installed | LocoBanco | 6 GB floor; bridges into pre-RTX coverage |
| RTX 2060 Super | Colmena | 3 | 3 | 8 GB | ~448 GB/s | Turing | 7.5 | installed | LocoBanco | 8 GB Turing floor; three cards for result-consistency validation |
| RTX 4060 Ti 16 GB | Colmena | 1 | 1 | 16 GB | ~288 GB/s | Ada | 8.9 | installed | LocoBanco | 16 GB consumer floor; documents the memory-bus penalty |
| RTX 5050 | Colmena | 1 | 1 | 8 GB | ~320 GB/s | Blackwell | — | installed | LocoBanco | Blackwell 8 GB floor |
| GTX 950 | Tortuga | 1 | 1 | 2 GB | ~105 GB/s | Maxwell | 5.2 | installed | LocoBanco | 2 GB floor; TinyLlama only |
| GTX 980 Ti | Tortuga | 1 | 1 | 6 GB | ~336 GB/s | Maxwell | 5.2 | installed | LocoBanco | 6 GB Maxwell floor |
| GTX Titan X | Tortuga | 1 | 1 | 12 GB | ~336 GB/s | Maxwell | 5.2 | installed | LocoBanco | 12 GB Maxwell floor |
| GTX 750 Ti | Tortuga | 1 | 1 | 2 GB | ~86 GB/s | Maxwell | 5.0 | installed | LocoBanco | 2 GB Maxwell baseline |
| GTX 1050 Ti | Tortuga | 1 | 1 | 4 GB | ~112 GB/s | Pascal | 6.1 | installed | LocoBanco | 4 GB Pascal floor |
| GTX 1060 3 GB | Tortuga | 1 | 1 | 3 GB | ~192 GB/s | Pascal | 6.1 | installed | LocoBanco | 3 GB Pascal floor |
| GTX 1060 6 GB | Tortuga | 1 | 1 | 6 GB | ~192 GB/s | Pascal | 6.1 | installed | LocoBanco | Additional 6 GB Pascal card |
| Tesla V100 16 GB | Hidra | 1 | 1 | 16 GB | ~900 GB/s | Volta | 7.0 | installed | LocoBanco | 16 GB server tier |
| Tesla P100 16 GB | Hidra | 1 | 1 | 16 GB | ~732 GB/s | Pascal | 6.0 | installed | LocoBanco | 16 GB server tier; HBM2 |
| Tesla M40 24 GB Arrived; awaiting installation | Unassigned | 1 | 1 | 24 GB | ~288 GB/s | Maxwell | 5.2 | installed | LocoBanco | 24 GB server floor; Ollama only |
| Tesla P40 24 GB Arrived; awaiting installation | Unassigned | 1 | 1 | 24 GB | ~346 GB/s | Pascal | 6.1 | installed | LocoBanco | 24 GB server tier; full modern stack |
| Tesla M4 8 GB Arrived; awaiting installation | Unassigned | 1 | 1 | 8 GB | ~88 GB/s | Maxwell | 5.2 | installed | LocoBanco | 8 GB server floor; Ollama only |
| Tesla P4 8 GB Arrived; awaiting installation | Unassigned | 1 | 1 | 8 GB | ~192 GB/s | Pascal | 6.1 | installed | LocoBanco | 8 GB server tier; full modern stack |
| GTX 1070 | Hidra | 1 | 1 | 8 GB | ~256 GB/s | Pascal | 6.1 | installed | LocoBanco | 8 GB Pascal bandwidth comparison |
| RTX 3050 | Hidra | 1 | 1 | 8 GB | ~224 GB/s | Ampere | 8.6 | installed | LocoBanco | 8 GB Ampere; lowest-bandwidth Tensor Core card |
| RTX 3060 AORUS Elite | Hidra | 1 | 1 | 12 GB | ~360 GB/s | Ampere | 8.6 | installed | LocoBanco | 12 GB tier floor |
| RTX 3090 | Puente | 1 | 1 | 24 GB | ~936 GB/s | Ampere | 8.6 | installed | LocoBanco + LocoPuente | LocoPuente PoC and LocoEnsayo chatbots |
| Tesla V100 32 GB | Búho | 1 | 1 | 32 GB | ~900 GB/s | Volta | 7.0 | installed | LocoBanco + LocoLLM | Adapter training and 32 GB inference |
| GTX 1050 Ti LP | Hormiga | 1 | 1 | 4 GB | ~112 GB/s | Pascal | 6.1 | installed | LocoBanco | Low-profile SFF floor node |
| GTX 1650 OC LP | Unassigned | 1 | 1 | 4 GB | ~128 GB/s | Turing | 7.5 | installed | LocoBanco | Low-profile; no current assignment |
| RTX 2060 Super | Unassigned | 2 | 2 | 8 GB | ~448 GB/s | Turing | 7.5 | installed | LocoBanco | Awaiting assignment |
| RTX 5070 Ti | Unassigned | 1 | 1 | 16 GB | ~896 GB/s | Blackwell | — | installed | LocoBanco | Awaiting assignment |
| GeForce GTX 770 2 or 4 GB variant; confirm | Relic shelf | 1 | 1 | 2 GB | ~224 GB/s | Kepler | 3.0 | installed | LocoFósil | Kepler 3.0 experiment |
| Tesla K10 Per GPU; two cards, four devices | Relic shelf | 2 | 4 | 4 GB / GPU | ~160 GB/s / GPU | Kepler | 3.0 | installed | LocoFósil | Dual-GPU Kepler experiment |
| Quadro K5200 | Relic shelf | 1 | 1 | 8 GB | ~192 GB/s | Kepler | 3.0 | installed | LocoFósil | 27B on a Kepler card |
| Quadro K6000 | Relic shelf | 1 | 1 | 12 GB | ~288 GB/s | Kepler | 3.0 | installed | LocoFósil | 27B capacity and bandwidth |
| Quadro K4200 | Relic shelf | 1 | 1 | 4 GB | ~173 GB/s | Kepler | 3.0 | installed | LocoFósil | 27B on a Kepler card |
| Quadro K420 | Relic shelf | 1 | 1 | 2 GB | ~29 GB/s | Kepler | 3.0 | installed | LocoFósil | Low-power 2B, limited by DDR3 |
| GeForce GT 710 | Relic shelf | 1 | 1 | 2 GB | ~14.4 GB/s | Kepler | 3.5 | installed | LocoFósil | Lowest-power 2B |
| Tesla K80 Per GPU; one card, two devices | Relic shelf | 1 | 2 | 12 GB / GPU | ~240 GB/s / GPU | Kepler | 3.7 | installed | LocoFósil | Dual-GPU experiment; cooling challenge |
| Quadro K620 | Relic shelf | 2 | 2 | 2 GB | ~28.8 GB/s | Maxwell | 5.0 | installed | LocoFósil | Low-power 2B, limited by DDR3 |
| Quadro K1200 | Relic shelf | 2 | 2 | 4 GB | ~80 GB/s | Maxwell | 5.0 | installed | LocoFósil | 2B and 9B efficiency; 27B in pair |
| GeForce GTX 960 Recorded under Tortuga and the relic shelf; counted once | Tortuga + Relic shelf | 1 | 1 | 4 GB | ~112 GB/s | Maxwell | 5.2 | installed | LocoBanco + LocoFósil | 4 GB Maxwell floor |
| Quadro M4000 | Relic shelf | 1 | 1 | 8 GB | ~192 GB/s | Maxwell | 5.2 | installed | LocoFósil | 27B on a Maxwell card |
| Quadro P600 | Relic shelf | 1 | 1 | 2 GB | ~40 GB/s | Pascal | 6.1 | installed | LocoFósil | Low-power 2B |
| Quadro P1000 | Relic shelf | 1 | 1 | 4 GB | ~82 GB/s | Pascal | 6.1 | installed | LocoFósil | 2B and 9B |
| Quadro P2000 | Relic shelf | 1 | 1 | 5 GB | ~140 GB/s | Pascal | 6.1 | installed | LocoFósil | 9B with headroom |
| Quadro P4000 | Relic shelf | 1 | 1 | 8 GB | ~243 GB/s | Pascal | 6.1 | installed | LocoFósil | 27B efficiency candidate |
Which list is canonical
Section titled “Which list is canonical”| Question | Canonical source |
|---|---|
| What do we own, and where is it? | This page |
| Which card is in which benchmark machine? | LocoBanco GPU Inventory |
| Old-card specs, CUDA levels, drivers, devices | LocoFósil Hardware |
| What should I buy? | LocoFósil build tiers for old cards; Nvidia GPU Reference for modern ones |
The project pages remain the detailed views and are the source of truth for their own fields. This registry is the single place to check ownership, count, machine, and status. When a card moves or arrives, update loco-theme/src/gpus.json first.
Counts
Section titled “Counts”- Cards are physical cards. GPU devices count each GPU on a dual-GPU card separately.
- The Tesla K10 is two cards and four devices; the Tesla K80 is one card and two devices.
- Unassigned cards are owned but not installed in a machine.
- The GTX 960 appears in both LocoBanco (Tortuga) and LocoFósil (relic shelf). It is counted once until the two records are confirmed as separate cards.