← All GPUs

NVIDIA · Workstation · Ada Lovelace (AD102)

NVIDIA RTX 6000 Ada Generation

No NVLink, unlike the A6000 it replaced.

VRAM

48 GB

GDDR6 ECC

Memory bandwidth

960 GB/s

FP16 compute

364.2 TFLOPS

dense tensor

Price

No list price

What fits at 4-bit

The 4-bit file plus 8,192 tokens of context and 1.5 GB of overhead, against the VRAM of one card. Speeds are measured medians where someone has published a run at that card count; a blank speed means no run, not a slow one. Machine pages carry estimates for the gaps.

ModelNeeds1 card (48 GB)
Qwen3.5 4B4B · Q4_K_M4.3 GBfits
Llama 3.1 8B Instruct8B · Q4_K_M7.1 GB121 tok/smedian of 1 run
Qwen3 8B8.2B · Q4_K_M7.3 GB134 tok/smedian of 1 run
Gemma 4 E4B IT8B · Q4_K_M6.5 GBfits
Granite 4.2 8B8.8B · Q4_K_M7.9 GBfits
Qwen3.5 9B9B · Q4_K_M7.0 GBfits
Gemma 4 12B IT11.95B · Q4_K_M8.2 GBfits
Ministral 3 14B Reasoning (2512)13.9B · Q4_K_M10.4 GBfits
Phi-414.7B · Q4_K_M11.3 GBfits
gpt-oss-20b21B, 3.6B active · MXFP4 (native)12.5 GB161 tok/smedian of 1 run
Devstral Small 2 24B (2512)24B · Q4_K_M16.1 GBfits
Qwen3.6 27B27B · Q4_K_M17.7 GBfits
Gemma 4 26B-A4B IT25.2B, 3.8B active · Q4_K_M17.4 GBfits
Qwen3.8 27B27B · Q4_K_M18.2 GBfits
Granite 4.2 30B29.3B · Q4_K_M20.3 GBfits
Gemma 4 31B IT30.7B · Q4_K_M18.9 GBfits
Qwen3 Coder 30B-A3B Instruct30.5B, 3.3B active · Q4_K_M19.5 GBfits
Ornith 1.5 35B-A3B35B, 3B active · Q4_K_M21.9 GBfits
Qwen3.6 35B-A3B35B, 3B active · Q4_K_M22.4 GBfits
Nemotron 3.5 Lightning 30B-A3B30B, 3B active · Q4_K_M25.3 GBfits
Llama 3.3 70B Instruct70B · Q4_K_M43.6 GB18 tok/smedian of 1 run
Qwen3 Coder Next80B, 3B active · Q4_K_M46.9 GBfits

Measured speeds

One card

Single-user runs on 1 card
ModelQuantBackendGenerationPromptContextSource
gpt-oss-20bMXFP4llama.cpp (build not stated)161 tok/s6,496 tok/s4,096-token prompt4,096review2026-03
Llama 3.1 8B InstructQ4_K_MLocalScore (llamafile; version not stated on leaderboard)121 tok/s6,808 tok/s0LocalScore2026-10
Llama 3.3 70B InstructQ4_Kllama.cpp (build not stated)18 tok/s678 tok/s4,096-token prompt4,096review2026-03
Qwen3 8BQ4_Kllama.cpp (build not stated)134 tok/s4,933 tok/s4,096-token prompt4,096review2026-03

Speculative decoding, several users, or long context

Kept apart because each measures a different workload from one person chatting near the start of a conversation.

Other runs on this card
ModelQuantBackendGenerationPromptContextMeasured onSource
gpt-oss-20bMXFP4llama.cpp (build not stated)137 tok/s5,351 tok/s16,384-token prompt16,384RTX 6000 Ada Generationreview2026-03
gpt-oss-20bMXFP4llama.cpp (build not stated)116 tok/s3,819 tok/s32,768-token prompt32,768RTX 6000 Ada Generationreview2026-03
Llama 3.3 70B InstructQ4_Kllama.cpp (build not stated)14 tok/s526 tok/s16,384-token prompt16,384RTX 6000 Ada Generationreview2026-03
Qwen3 8BQ4_Kllama.cpp (build not stated)99 tok/s4,096 tok/s16,384-token prompt16,384RTX 6000 Ada Generationreview2026-03
Qwen3 8BQ4_Kllama.cpp (build not stated)69 tok/s2,219 tok/s32,768-token prompt32,768RTX 6000 Ada Generationreview2026-03

Specifications

Memory
48 GB GDDR6 ECC
FP16 basis
Dense FP16/BF16 tensor with FP32 accumulate (Ada pro whitepaper); 728.4 sparse. Product page quotes 1,457 AI TOPS (FP8 sparse).
Board power
300 W
Slots
2
Length
267 mm
PCIe
Gen 4 x16
NVLink
No
Cooling
Blower
Software
CUDA, Vulkan
Released
2022
Availability
sold new
Price basis
No NVIDIA list price found; NVIDIA's store page did not load from our tooling.
Maker page
nvidia.com

Sources

  • nvidia.com: 48 GB GDDR6 ECC, 300 W, PCIe Gen 4 x16, 4.4 x 10.5 in dual slot, active; no NVLink row
  • images.nvidia.com: NVIDIA Ada professional whitepaper: RTX A6000 vs RTX 6000 Ada bandwidth, TGP, FP16 tensor TFLOPS
  • nvidia.com: datasheet: 18,176 CUDA cores, 48 GB ECC, PCIe Gen 4