Skip to content

GPU Inventory

Every GPU currently in the LocoLab project, organised by machine. Cards move between machines as experiments require — this is the current assignment.

For specifications, acquisition guidance, and generation-level analysis see the Nvidia GPU Reference.


The RTX-era benchmark platform. An 8-GPU enclosed mining rig running LocoBench. Cards here are chosen as floor representatives — the worst card per VRAM tier, so results are honest baselines. Matched trios give repeat-measurement discipline at their tier.

CardVRAMBandwidthTier Role
GTX 1060 6 GB (x3)6 GB192 GB/sFloor of 6 GB tier (Pascal, no Tensor Cores); bridges into Tortuga’s pre-RTX coverage
RTX 2060 Super (x3)8 GB448 GB/sFloor of 8 GB Turing tier (Tensor Cores). Three cards for result-consistency validation
RTX 4060 Ti16 GB288 GB/sFloor of 16 GB consumer tier — documents the memory-bus penalty
RTX 50508 GB320 GB/sBlackwell floor of 8 GB tier (GDDR6, FP4 support). Isolates architecture effect vs bandwidth deficit

Swappable bench cards. Cards rotate through Tortuga to fill out LocoBench tier coverage, particularly the older Maxwell and Pascal generations.

CardVRAMBandwidthTier Role
GTX 9502 GB105 GB/sFloor of 2 GB tier (Maxwell). TinyLlama 1.1B only. Quality cliff reference point
GTX 9604 GB112 GB/sFloor of 4 GB tier (Maxwell, Compute 5.2, Ollama only)
GTX 980 Ti6 GB336 GB/sFloor of 6 GB Maxwell tier (Ollama only)
GTX Titan X12 GB336 GB/sFloor of 12 GB Maxwell tier (Ollama only)
GTX 750 Ti2 GB86 GB/sFloor of 2 GB Maxwell tier. TinyLlama 1.1B only. Pre-Pascal baseline
GTX 1050 Ti4 GB112 GB/sFloor of 4 GB tier (Pascal, no Tensor Cores)
GTX 1060 3 GB3 GB192 GB/sFloor of 3 GB tier (Pascal). Quality cliff reference point
GTX 1060 6 GB6 GB192 GB/sAdditional 6 GB Pascal card

Full-bandwidth multi-GPU research, server GPU benchmarking, and GPU onboarding. X99 dual-Xeon platform with 4x PCIe x16 slots in an open-frame chassis for rapid card swaps. The server GPUs are the primary LocoBench payload at their native VRAM tiers; consumer cards rotate through for onboarding and fill-in benchmarks.

Server GPUs:

CardVRAMBandwidthTier Role
Tesla V10016 GB900 GB/s16 GB server tier (Volta, HBM2, Tensor Cores)
Tesla P10016 GB732 GB/s16 GB server tier (Pascal, HBM2, no Tensor Cores)
Tesla M4024 GB288 GB/s24 GB server floor (Maxwell, CC 5.2, Ollama only). Incoming
Tesla P4024 GB346 GB/s24 GB server tier (Pascal, full modern stack). Incoming
Tesla M48 GB88 GB/s8 GB server floor (Maxwell, CC 5.2, Ollama only). Incoming
Tesla P48 GB192 GB/s8 GB server tier (Pascal, CC 6.1, full modern stack). Incoming

Consumer rotation (onboarding + small-card benchmarks):

CardVRAMBandwidthTier Role
GTX 10708 GB256 GB/s8 GB Pascal (no Tensor Cores). Bandwidth comparison against 2060 Super
RTX 30508 GB224 GB/s8 GB Ampere. Lowest bandwidth Tensor Core card in the 8 GB tier
RTX 3060 AORUS Elite12 GB360 GB/sFloor of 12 GB tier (Ampere, Tensor Cores)

Student-facing host. Ryzen 5 2600 desktop running the “closing the gap” minimal LocoPuente PoC and most LocoEnsayo rehearsal chatbots on a single card.

CardVRAMBandwidthTier Role
RTX 309024 GB936 GB/sSole card for LocoPuente + LocoEnsayo (LLM, image gen, voice, chatbots)

Dedicated AI inference and LocoLLM adapter training. Single-Xeon X99 workstation with 32 GB DDR4. One card, one job — patient and nocturnal.

CardVRAMBandwidthTier Role
Tesla V10032 GB900 GB/s32 GB server tier (Volta, HBM2, Tensor Cores). Primary adapter-training card; single-card inference at 32 GB

Low-profile / office deployment.

CardVRAMBandwidthTier Role
GTX 1050 Ti LP4 GB112 GB/sLow-profile 4 GB card for constrained chassis

CardVRAMBandwidthNotes
GTX 1650 OC LP4 GB128 GB/sTuring, no Tensor Cores. Low-profile. No current assignment
RTX 2060 Super (x2)8 GB448 GB/sAwaiting assignment
RTX 5070 Ti16 GB896 GB/sBlackwell, GDDR7. Awaiting assignment

MachineCardsPrimary Role
Colmena8RTX-era LocoBench consumer tier benchmarking
Tortuga8Swappable tier cards for pre-RTX LocoBench coverage
Hidra5LocoConvoy multi-GPU + server GPU benchmarking + onboarding (2 server cards installed, 4 incoming)
Puente1LocoPuente PoC + LocoEnsayo chatbots (RTX 3090 24 GB)
Búho1Dedicated LocoLLM adapter training and single-card inference
Hormiga1SFF floor node / office deployment
Unassigned4
Total installed28(+ 4 incoming server cards for Hidra)

RTX 5060 Ti: not confirmed in fleet — omitted pending verification.


For GPU generation details and acquisition guidance see Nvidia GPU Reference. For benchmarking methodology see the loco-bench documentation.