GPU Lab

Browse 120+ GPUs ranked by performance tier (consumer entry through national datacenter scale), then estimate which LLM sizes fit your GPU for inference and fine-tuning.

Read the GPU Lab article →

GPUs cataloged

123

Flopper.io spec sheets

Performance tiers

Consumer → national AI scale

Performance tier ladder

Each GPU gets a tier ID from consumer entry (T1) up to hyperscale / country-datacenter class (T6). Click a tier to filter the catalog below.

GPU catalog

Tier	GPU	VRAM	FP32	Score
T2 Mid	NVIDIA Tesla P4 8GB Pascal · —	8 GB	5.5	18
T2 Mid	NVIDIA GeForce RTX 3070 8GB Ampere · —	8 GB	20.3	26
T2 Mid	NVIDIA GeForce RTX 3080 10GB Ampere · —	10 GB	29.8	34
T2 Mid	NVIDIA GeForce RTX 3080 Ti 12GB Ampere · —	12 GB	34.1	40
T3 Enthusiast	Apple M1 Apple M1 · —	16 GB	2.6	33
T3 Enthusiast	NVIDIA P100 PCIe 16GB Pascal · —	16 GB	9.3	37
T3 Enthusiast	Moore Threads S80 MUSA ChunXiao · —	16 GB	14.4	38
T3 Enthusiast	NVIDIA P100 SXM2 16GB Pascal · —	16 GB	10.6	38
T3 Enthusiast	NVIDIA T4 16GB Turing · —	16 GB	8.1	40
T3 Enthusiast	NVIDIA V100 PCIe 16GB Volta · —	16 GB	14	47
T3 Enthusiast	NVIDIA V100 SXM2 16GB Volta · —	16 GB	15.7	48
T3 Enthusiast	Apple M2 Apple M2 · —	24 GB	3.6	50
T3 Enthusiast	Apple M3 Apple M3 · —	24 GB	3.5	50
T3 Enthusiast	NVIDIA Tesla K80 24GB Kepler · —	24 GB	5.6	50
T3 Enthusiast	NVIDIA GeForce RTX 4070 Ti 12GB Ada Lovelace · —	12 GB	40.1	53
T3 Enthusiast	NVIDIA Tesla P40 24GB Pascal · —	24 GB	12	53
T3 Enthusiast	Enflame Yunsui i20 GCU / DTU 2.5 · —	16 GB	32	55
T3 Enthusiast	Google TPU v3 32GB TPU · —	32 GB	—	64
T3 Enthusiast	Google TPU v4 32GB TPU · —	32 GB	—	64
T3 Enthusiast	Google TPU v6e 32GB TPU · —	32 GB	—	64
T3 Enthusiast	NVIDIA A30 24GB Ampere · —	24 GB	10.3	65
T3 Enthusiast	NVIDIA L2 24GB Ada Lovelace · —	24 GB	24.1	65
T3 Enthusiast	Apple M4 Apple M4 · —	32 GB	4.3	66
T3 Enthusiast	Apple M5 Apple M5 · —	32 GB	4.2	66
T3 Enthusiast	Apple M1 Pro Apple M1 · —	32 GB	5.3	67
T3 Enthusiast	NVIDIA GeForce RTX 4080 16GB Ada Lovelace · —	16 GB	48.7	67
T3 Enthusiast	Apple M2 Pro Apple M2 · —	32 GB	6.8	68
T3 Enthusiast	AMD Radeon RX 7900 XT 20GB RDNA 3 · —	20 GB	51.6	69
T3 Enthusiast	Moore Threads S3000 32GB MUSA ChunXiao · —	32 GB	15.2	70
T3 Enthusiast	NVIDIA A10 24GB Ampere · —	24 GB	31.2	70
T3 Enthusiast	NVIDIA GeForce RTX 4080 Super 16GB Ada Lovelace · —	16 GB	52.2	70
T3 Enthusiast	Apple M3 Pro Apple M3 · —	36 GB	7.4	76
T3 Enthusiast	Iluvatar CoreX BI-V100 Tiangai 100 CoreX GPGPU · —	32 GB	—	76
T3 Enthusiast	NVIDIA V100 PCIe 32GB Volta · —	32 GB	14	79
T3 Enthusiast	NVIDIA V100 SXM2 32GB Volta · —	32 GB	15.7	80
T3 Enthusiast	NVIDIA V100S PCIe 32GB Volta · —	32 GB	16.4	81
T3 Enthusiast	Enflame Yunsui T20 GCU / DTU 2.0 · —	32 GB	33.6	88
T3 Enthusiast	NVIDIA GeForce RTX 5080 16GB Blackwell · —	16 GB	56.3	91
T3 Enthusiast	Moore Threads MTT S4000 MUSA 3rd Gen · —	48 GB	25	114
T3 Enthusiast	Apple M4 Pro Apple M4 · —	64 GB	9.2	133
T3 Enthusiast	Apple M5 Pro Apple M5 · —	64 GB	8.3	133
T3 Enthusiast	Apple M1 Max Apple M1 · —	64 GB	10.4	134
T3 Enthusiast	Intel Data Center GPU Max 1100 48GB Xe-HPC · —	48 GB	22.2	138
T3 Enthusiast	MetaX N260 MXN (Xisi) · —	64 GB	—	139
T3 Enthusiast	MetaX C500 XCORE GPGPU · —	64 GB	18	158
T3 Enthusiast	Apple M2 Max Apple M2 · —	96 GB	13.6	200
T3 Enthusiast	Apple M3 Max Apple M3 · —	128 GB	16.4	265
T3 Enthusiast	Apple M5 Max Apple M5 · —	128 GB	16.6	265
T3 Enthusiast	Apple M4 Max Apple M4 · —	128 GB	18.4	266
T3 Enthusiast	Apple M1 Ultra Apple M1 · —	128 GB	21.2	268
T4 Pro	Intel Data Center GPU Flex Series 140 12GB Xe-HPG · —	12 GB	8	27
T4 Pro	NVIDIA RTX A2000 12GB Ampere · —	12 GB	8	28
T4 Pro	Intel Data Center GPU Flex Series 170 16GB Xe HPG · —	16 GB	16	38
T4 Pro	NVIDIA RTX 2000 Ada 16GB Ada Lovelace · —	16 GB	12	41
T4 Pro	NVIDIA RTX A4000 16GB Ampere · —	16 GB	19.2	41
T4 Pro	Intel Arc A770 16GB Alchemist · —	16 GB	19.7	43
T4 Pro	NVIDIA RTX A4500 20GB Ampere · —	20 GB	23.6	51
T4 Pro	NVIDIA RTX A5000 24GB Ampere · —	24 GB	27.8	61
T4 Pro	NVIDIA GeForce RTX 3090 24GB Ampere · —	24 GB	35.6	65
T4 Pro	NVIDIA GeForce RTX 3090 Ti 24GB Ampere · —	24 GB	40	67
T4 Pro	NVIDIA A10G 24GB Ampere · —	24 GB	35	68
T4 Pro	AMD Radeon RX 7900 XTX 24GB RDNA 3 · —	24 GB	61.4	82
T4 Pro	NVIDIA GeForce RTX 4090 24GB Ada Lovelace · —	24 GB	82.6	107
T4 Pro	NVIDIA RTX 5000 Ada 32GB Ada Lovelace · —	32 GB	65.3	111
T4 Pro	NVIDIA RTX A6000 48GB Ampere · —	48 GB	38.7	115
T4 Pro	NVIDIA A40 48GB Ampere · —	48 GB	37.4	123
T4 Pro	NVIDIA GeForce RTX 5090 32GB Blackwell · —	32 GB	104.8	173
T5 DC	Cambricon MLU590 MLUarch05 · —	— GB	—	0
T5 DC	Baidu Kunlun P800 Kunlun III XPU-P (3rd-gen) · —	— GB	—	28
T5 DC	AMD Instinct MI100 32GB CDNA · —	32 GB	23.1	88
T5 DC	Huawei Ascend 910A Da Vinci · —	32 GB	16	91
T5 DC	NVIDIA L4 24GB Ada Lovelace · —	24 GB	30.3	99
T5 DC	Cambricon MLU370-X8 MLUarch03 · —	48 GB	24	113
T5 DC	NVIDIA A100 PCIe 40GB Ampere · —	40 GB	19.5	113
T5 DC	NVIDIA A100 SXM4 40GB Ampere · —	40 GB	19.5	113
T5 DC	NVIDIA L20 48GB Ada Lovelace · —	48 GB	59.8	139
T5 DC	AMD Instinct MI210 PCIe CDNA 2 · —	64 GB	22.6	152
T5 DC	Huawei Ascend 910B Da Vinci v2 · —	64 GB	—	160
T5 DC	NVIDIA L40 48GB Ada Lovelace · —	48 GB	90.5	161
T5 DC	NVIDIA RTX 6000 Ada 48GB Ada Lovelace · —	48 GB	91.1	162
T5 DC	NVIDIA A100 PCIe 80GB Ampere · —	80 GB	19.5	193
T5 DC	NVIDIA A100 SXM4 80GB Ampere · —	80 GB	19.5	193
T5 DC	NVIDIA A800 PCIe 80GB Ampere · —	80 GB	19.5	193
T5 DC	Moore Threads MTT S5000 MUSA 4th Gen · —	80 GB	—	200
T5 DC	NVIDIA H20 96GB Hopper · —	96 GB	44	221
T5 DC	NVIDIA L40S 48GB Ada Lovelace · —	48 GB	91.6	221
T5 DC	Biren BR100 Biren SPC · —	64 GB	256	230
T5 DC	NVIDIA RTX PRO 6000 Blackwell Workstation Edition Blackwell · —	96 GB	125	242
T5 DC	NVIDIA DGX Spark Grace Blackwell · —	128 GB	—	264
T5 DC	Intel Gaudi 2 96GB Gaudi 2 · —	96 GB	45.2	279
T5 DC	NVIDIA H100 PCIe 80GB Hopper · —	80 GB	51	301
T5 DC	NVIDIA H800 PCIe 80GB Hopper · —	80 GB	51	301
T5 DC	AMD Instinct MI250 128GB CDNA 2 · —	128 GB	45.3	303
T5 DC	AMD MI250X 128GB CDNA 2 · —	128 GB	47.9	306
T5 DC	Intel Data Center GPU Max 1550 128GB Xe-HPC · —	128 GB	52	344
T5 DC	NVIDIA H100 SXM5 80GB Hopper · —	80 GB	67	345
T5 DC	NVIDIA H100 NVL 94GB Hopper · —	94 GB	60	346
T5 DC	AMD Instinct MI300A 128GB CDNA 3 · —	128 GB	122.6	462
T5 DC	Intel Gaudi 3 128GB Gaudi 3 · —	128 GB	229	568
T6 Scale	Google TPU v5e 16GB TPU · —	16 GB	—	32
T6 Scale	Google TPU v5p 95GB TPU v5p · —	95 GB	—	208
T6 Scale	NVIDIA GB10 Grace Blackwell Blackwell · —	128 GB	31	308
T6 Scale	NVIDIA H20 141GB HBM3e Hopper · —	141 GB	44	311
T6 Scale	Huawei Ascend 950PR Da Vinci v3 · —	128 GB	—	336
T6 Scale	Huawei Ascend 910C Da Vinci v2 (dual-die) · —	128 GB	50	340
T6 Scale	Huawei Ascend 950DT Da Vinci v3 · —	144 GB	—	368
T6 Scale	Apple M2 Ultra Apple M2 · —	192 GB	27.2	399
T6 Scale	NVIDIA H200 NVL 141GB Hopper · —	141 GB	60	440
T6 Scale	NVIDIA H200 SXM 141GB Hopper · —	141 GB	67	467
T6 Scale	Google TPU v7 192GB TPU · —	192 GB	—	569
T6 Scale	AMD Instinct MI300X 192GB CDNA 3 · —	192 GB	163.4	659
T6 Scale	NVIDIA B100 SXM 192GB Blackwell · —	192 GB	60	688
T6 Scale	Huawei Ascend 960 Da Vinci v4 (chiplet) · —	288 GB	—	736
T6 Scale	NVIDIA B200 SXM 180GB Blackwell · —	180 GB	75	750
T6 Scale	AMD Instinct MI325X OAM CDNA 3 · —	256 GB	163.4	787
T6 Scale	Huawei Ascend 970 Da Vinci v5 · —	288 GB	—	896
T6 Scale	NVIDIA GB300 Grace Blackwell Ultra Superchip 252GB Blackwell Ultra · —	252 GB	80	936
T6 Scale	NVIDIA B300 SXM 288GB Blackwell Ultra · —	288 GB	75	966
T6 Scale	AMD Instinct MI350X OAM CDNA 4 · —	288 GB	144.2	1003
T6 Scale	AMD Instinct MI355X OAM CDNA 4 · —	288 GB	157.3	1042
T6 Scale	NVIDIA GH200 Hopper · —	468 GB	67	1121
T6 Scale	NVIDIA Rubin SXM Rubin · —	288 GB	130	1648
T6 Scale	NVIDIA Vera Rubin Superchip Rubin · —	576 GB	260	3296

Your GPU → LLM fit

Pick a catalog GPU or enter custom VRAM. Estimates assume single-GPU workloads with modest context.

Detect your GPU from terminal

Detected OS: Unknown OS

Run the command below in Terminal (Unknown OS), paste the output, then match to the catalog.

macOS

system_profiler SPDisplaysDataType

If you are on a Mac.

Linux NVIDIA

nvidia-smi --query-gpu=name,memory.total --format=csv,noheader

If you are on Linux with NVIDIA.

Windows

powershell -Command "Get-CimInstance Win32_VideoController | Select-Object Name, AdapterRAM | Format-List"

If you are on Windows.

Paste terminal output

Use custom VRAM instead of catalog GPU Select GPU (weak → strong)

Usable VRAM

22.5 GB

Largest inference (Q4)

CodeLlama / 34B class

Models for inference

11 / 14

Largest QLoRA fine-tune

Llama 2/3 13B

Model	Params	Infer Q4	Infer FP16	QLoRA	Full FT
TinyLlama 1.1B	1.1B	Yes 1.9 GB	Yes 4.8 GB	Yes 5.3 GB	No 23.4 GB
Qwen2.5 1.5B	1.5B	Yes 2.1 GB	Yes 5.7 GB	Yes 5.6 GB	No 29.0 GB
Llama 3.2 3B	3B	Yes 3.1 GB	Yes 8.8 GB	Yes 6.7 GB	No 50.0 GB
Phi-3 Mini 3.8B	3.8B	Yes 3.6 GB	Yes 10.5 GB	Yes 7.2 GB	No 61.2 GB
Mistral 7B	7B	Yes 5.5 GB	Yes 17.2 GB	Yes 9.5 GB	No 106.0 GB
Llama 3.1 8B	8B	Yes 6.2 GB	Yes 19.3 GB	Yes 10.3 GB	No 120.0 GB
Gemma 2 9B	9B	Yes 6.8 GB	Yes 21.4 GB	Yes 11.0 GB	No 134.0 GB
Qwen2.5 14B	14B	Yes 9.9 GB	No 31.9 GB	Yes 14.6 GB	No 204.0 GB
Llama 2/3 13B	13B	Yes 9.3 GB	No 29.8 GB	Yes 13.9 GB	No 190.0 GB
Qwen2.5 32B	32B	Yes 21.0 GB	No 69.7 GB	No 27.5 GB	No 456.0 GB
CodeLlama / 34B class	34B	Yes 22.3 GB	No 73.9 GB	No 29.0 GB	No 484.0 GB
Qwen2.5 72B	72B	No 45.8 GB	No 153.7 GB	No 56.3 GB	No 1016.0 GB
Llama 3.1 70B	70B	No 44.6 GB	No 149.5 GB	No 54.9 GB	No 988.0 GB
Llama 3.1 405B Multi-GPU only	405B	No 252.3 GB	No 853.0 GB	No 296.1 GB	No 5678.0 GB

VRAM estimates: Q4 inference ≈ 0.62× params + overhead; FP16 inference ≈ 2.1× params; QLoRA ≈ 0.72× params + 4.5 GB; full fine-tune ≈ 14× params + 8 GB. Actual needs vary with context length, batch size, and framework.

GPU Lab

Performance tier ladder

Consumer Entry

Consumer Mid

Consumer High

Workstation / Pro

Datacenter GPU

Hyperscale / National AI

GPU catalog

Your GPU → LLM fit

Detect your GPU from terminal