VOOZH about

URL: https://willitrunai.com/gpus/a16-64gb

โ‡ฑ AI Models for NVIDIA A16 64GB โ€” What Runs on 64GB VRAM


NVIDIA

NVIDIA A16 64GB

Ampere DatacenterDatacenterAmperePCIe 4CUDA
64GB
VRAM
600GB/s
Bandwidth
78TFLOPS
FP16 Compute
624TOPS
INT8 Inference
$6,500 MSRP
NVIDIA A16 64GBCategory AvgMac Studio M3 Ultra 96GB

Operating mode

Choose the operating mode for this hardware

Use this to bias workload recommendations toward responsiveness, background autonomy, lighter serving, or multi-GPU scale-out.

Current mode

Balanced

Balanced for general local use. Keeps the ranking neutral across personal and serving workflows.

See Full AI Tier List for NVIDIA A16 64GB โ†’

About this GPU for AI

The NVIDIA A16 is a virtualization-oriented Ampere GPU that packages four GA107 GPU dies on a single card with a combined 64 GB of GDDR6 memory โ€” 16 GB per die. It was designed for virtual workstation and virtual desktop infrastructure (VDI) use cases, but its aggregate 64 GB VRAM makes it usable for large-model inference when software supports multi-die configurations. Each GPU die is a modest chip, so raw FP16 compute (78 TFLOPS combined) is lower than dedicated inference GPUs. For teams already running A16 infrastructure for VDI, it can double as an AI inference host.

Beyond LLMs

AI Capability Matrix

What AI tasks this GPU can handle โ€” from text generation to image and video creation.

CapabilityStatusRepresentative ModelDetail
LLM Chat (7B)Runs nativelyLlama 3.1 8B Q4โ€”
LLM Coding (30B)Runs nativelyQwen 3 30B Q4โ€”
LLM Large (70B)Runs nativelyLlama 3.1 70B Q4โ€”
Image Gen (SDXL)Runs nativelySDXL 1.0 FP16~~4.1s per image
Image Gen (Flux)Runs nativelyFlux.1 Dev FP16~~18.4s per image
Image Gen (SD 3.5)Runs nativelySD 3.5 Large FP16~~22.5s per image
Video Short (25f)Runs nativelyLTX Video 2B~~3.6s/frame
Video Long (100f)Very constrainedWan Video 14B~~18s/frame
large-vrammulti-dievirtualization-focusedenterprise-grade

Specifications

Compute
FP1678 TFLOPS
INT8624 TOPS
ArchitectureAmpere
Memory
VRAM64 GB
Bandwidth600 GB/s
General
FamilyAmpere Datacenter
SegmentDatacenter
InterconnectPCIe 4
Compute PlatformCUDA
MSRP$6,500

Key Features

64 GB GDDR6 total (4ร— 16 GB dies on one card)600 GB/s combined memory bandwidth78 TFLOPS FP16 combined / 624 INT8 TOPS4ร— NVIDIA GA107 GPU dies on a single PCIe cardPCIe 4.0, 250W TDPMIG-like isolation via virtual GPU partitioning per die

For AI Workloads

Strengths
  • 64 GB aggregate VRAM allows 30B models at Q4 on a single card
  • Four independent GPU dies provide natural multi-tenant isolation for virtualized deployments
  • Lower cost than A40 (48 GB) in the secondary market despite higher total VRAM
  • Suitable for organizations already running NVIDIA vGPU infrastructure
Considerations
  • Multi-die architecture complicates single-model inference across all 64 GB โ€” not a unified memory pool
  • Combined 600 GB/s bandwidth is low for 64 GB capacity โ€” bandwidth per GB is poor
  • Each individual GA107 die only has 16 GB, limiting what can run on a single compute unit
  • Primarily a VDI product โ€” AI inference is a secondary use case, and framework support can vary

Architecture

Ampere

Ampere is NVIDIA's second-generation RTX architecture, built on Samsung's 8nm process. It introduced 3rd-generation Tensor Cores with support for sparsity-accelerated INT8 operations and improved FP16 throughput over Turing.

AI Relevance

Sparsity-aware Tensor Cores can effectively double throughput for structured sparse workloads. However, the lack of FP8 support means quantized inference is less efficient than Ada Lovelace or Blackwell.

Process: Samsung 8nmPlatform: CUDATensor Cores: Gen 3Precisions: FP32, FP16, BF16, INT8, INT4

Buying advice

Should you buy NVIDIA A16 64GB for local AI?

Excellent choice for local AI

Runs 29 of 50 top models well โ€” a strong all-rounder for local inference.

64.0 GB

VRAM

$6,500

MSRP

$102/GB

Cost per GB VRAM

Best models for this GPU

What will limit you first

This setup is broadly balanced for this model.

No major red flags

This recommendation has enough memory headroom and acceptable estimated speed for the selected workload.

Best upgrade itinerary

Unlocks 1 additional models that do not fit on the current setup.

Want more headroom? Mac Studio M3 Ultra 96GB (96.0 GB unified memory) is the next step up.

Recommendations by Workload

Chat

S

Qwen 3.5 27B

This model is a direct match for chat. It belongs to a current frontier family for local AI. It fits natively with comfortable headroom. Known channels: huggingface, ollama, lm-studio.

Decode 30.7 tok/s ยท 131K ctx ยท llama.cppEST.
25.4 GB / 64.0 GB VRAM

Coding

S

Qwen3-Coder-Next

This model is a direct match for coding. It belongs to a current frontier family for local AI. It should run, but memory headroom will be limited. Known channels: huggingface, ollama, lm-studio.

Decode 31.6 tok/s ยท 86K ctx ยท llama.cppEST.
57.6 GB / 64.0 GB VRAM

Agentic Coding

S

Qwen3-Coder-Next

This model is still usable for agentic-coding, but it is not the most specialized pick. It belongs to a current frontier family for local AI. It should run, but memory headroom will be limited. Known channels: huggingface, ollama, lm-studio.

Decode 31.6 tok/s ยท 86K ctx ยท llama.cppEST.
59.0 GB / 64.0 GB VRAM

Reasoning

S

Qwen 3 32B

This model is a direct match for reasoning. It belongs to a current frontier family for local AI. It fits natively with comfortable headroom. Known channels: huggingface, ollama, lm-studio.

Decode 26.1 tok/s ยท 131K ctx ยท llama.cppEST.
30.7 GB / 64.0 GB VRAM

RAG

S

Qwen 3.5 27B

This model is a direct match for rag. It belongs to a current frontier family for local AI. It fits natively with comfortable headroom. Known channels: huggingface, ollama, lm-studio.

Decode 30.7 tok/s ยท 131K ctx ยท llama.cppEST.
30.1 GB / 64.0 GB VRAM

Full Model Compatibility

๐Ÿ‘ Alibaba
Qwen 3.6 35B A3B
S94
35B32.8 GB60 tok/s138K ctx
+1moe
๐Ÿ‘ Alibaba
Qwen3-Coder 30B A3B Instruct
S93
30.5B27.4 GB71 tok/s256K ctx
moe
๐Ÿ‘ Alibaba
Qwen3-VL 30B A3B Instruct
S92
30B27.1 GB73 tok/s256K ctx
moe
๐Ÿ‘ Alibaba
Qwen 3.5 35B A3B
S92
35B30.1 GB65 tok/s131K ctx
moe
๐Ÿ‘ Alibaba
Qwen 3 30B A3B
S91
30.5B27.4 GB71 tok/s131K ctx
moe
๐Ÿ‘ Alibaba
Qwen3-Coder-Next
S90
80B57.6 GB32 tok/s86K ctx
moe
๐Ÿ‘ Alibaba
Qwen 3.5 27B
S90
27B26.9 GB31 tok/s131K ctx
dense
S89
32B30.7 GB26 tok/s131K ctx
dense
S88
9B15.0 GB92 tok/s131K ctx
dense
๐Ÿ‘ NVIDIA
Nemotron Cascade 2 30B A3B
S88
30B28.5 GB72 tok/s210K ctx
moe
๐Ÿ‘ Mistral
Magistral Small 2507
S88
24B24.4 GB34 tok/s131K ctx
dense
๐Ÿ‘ Mistral
Devstral Small 2 24B Instruct
S88
24B24.4 GB34 tok/s256K ctx
dense
๐Ÿ‘ Alibaba
Qwen 3.6 27B
S88
27B24.7 GB23 tok/s262K ctx
+1dense
๐Ÿ‘ OpenAI
GPT-OSS 20B
S88
21B22.6 GB90 tok/s128K ctx
moe
๐Ÿ‘ Alibaba
Qwen 2.5 VL 72B
S88
72B56.1 GB12 tok/s33K ctx
dense
๐Ÿ‘ NVIDIA
Nemotron 3 Nano 30B
S87
30B28.0 GB28 tok/s131K ctx
dense
๐Ÿ‘ Google
Gemma 4 31B
S87
30.7B40.7 GB20 tok/s41K ctx
dense
S87
14B18.3 GB59 tok/s131K ctx
dense
S87
8B14.4 GB103 tok/s131K ctx
dense
๐Ÿ‘ Microsoft
Phi-4-reasoning-plus 14B
S86
14.7B19.3 GB56 tok/s33K ctx
dense
๐Ÿ‘ Mistral
Devstral Small 1.1
S86
24B24.4 GB34 tok/s131K ctx
dense
๐Ÿ‘ Google
Gemma 4 26B A4B
S85
25.2B26.3 GB76 tok/s181K ctx
moe
A84
4B11.9 GB56 tok/s131K ctx
dense
๐Ÿ‘ LG AI
EXAONE 4.0 32B
A83
32B30.7 GB26 tok/s131K ctx
dense
๐Ÿ‘ NVIDIA
Nemotron Nano 8B
A82
8B14.1 GB103 tok/s131K ctx
dense
๐Ÿ‘ Mistral
Ministral 3 14B
A82
14B18.3 GB59 tok/s262K ctx
multimodal
๐Ÿ‘ Microsoft
Phi-4 Mini Reasoning 4B
A81
3.8B11.1 GB53 tok/s131K ctx
dense
๐Ÿ‘ Jina AI
Jina Embeddings v3
A74
0.57B10.4 GB8 tok/s8K ctx
dense
0.57B9.6 GB8 tok/s8K ctx
dense
๐Ÿ‘ Alibaba
Qwen 3.5 397B A17B
F0
397B252.3 GB2 tok/s4K ctx
moe
๐Ÿ‘ Mistral
Devstral 2 123B Instruct
F0
123B87.7 GB3 tok/s4K ctx
dense
1000B624.7 GB2 tok/s4K ctx
moe
1000B624.7 GB2 tok/s4K ctx
+1moe
๐Ÿ‘ DeepSeek
DeepSeek V4 Pro
F0
1600B871.2 GB2 tok/s4K ctx
moe
๐Ÿ‘ Alibaba
Qwen 3.5 122B A10B
F0
122B84.2 GB8 tok/s4K ctx
moe
๐Ÿ‘ DeepSeek
DeepSeek V4 Flash
F0
284B166.6 GB2 tok/s4K ctx
moe
๐Ÿ‘ Mistral
Mistral Small 4 119B
F0
119B85.3 GB8 tok/s4K ctx
moe
๐Ÿ‘ Cohere
Command A 111B
F0
111B78.9 GB4 tok/s4K ctx
dense
๐Ÿ‘ OpenAI
GPT-OSS 120B
F0
117B83.6 GB3 tok/s4K ctx
dense
754B486.3 GB2 tok/s4K ctx
moe
๐Ÿ‘ Mistral AI
Pixtral Large 124B
F0
124B88.3 GB3 tok/s4K ctx
dense
744B480.2 GB2 tok/s4K ctx
moe
๐Ÿ‘ DeepSeek
DeepSeek V3.2
F0
671B417.1 GB2 tok/s4K ctx
moe
๐Ÿ‘ Alibaba
Qwen 3 235B A22B
F0
235B153.5 GB2 tok/s4K ctx
moe
๐Ÿ‘ Alibaba
Qwen3-Coder 480B A35B Instruct
F0
480B303.0 GB2 tok/s4K ctx
moe
MiniMax M2.7
F0
230B151.4 GB2 tok/s4K ctx
moe
๐Ÿ‘ Mistral
Leanstral 119B A6B
F0
119B88.7 GB7 tok/s4K ctx
moe
๐Ÿ‘ DeepSeek
DeepSeek Coder V2 236B
F0
236B209.9 GB2 tok/s4K ctx
moe
๐Ÿ‘ DeepSeek
DeepSeek R1 671B
F0
671B476.2 GB2 tok/s4K ctx
moe
๐Ÿ‘ DeepSeek
DeepSeek V3.1 671B
F0
671B476.2 GB2 tok/s4K ctx
moe

Just out of reach

Models you could run with an upgrade

High-quality models that need a bit more memory

Image & Video Generation

Diffusion Model Compatibility

51 of 52 models can generate images or video on your NVIDIA A16 64GB

ModelMax ResolutionGen TimeGrade
SD TurboImage512ร—512500msS
Stable Diffusion 1.5Image512ร—768~1sS
Realistic Vision v5.1Image512ร—768~1sS
DreamShaper 8Image512ร—768~1sS
LCM DreamShaper v7Image512ร—768300msS
PixArt-SigmaImage1024ร—1024~4.1sS
FramePack I2VVideo1280ร—720~7.5s/frameS
SDXL TurboImage512ร—512500msS
SDXL LightningImage1024ร—1024~1.5sS
Stable Diffusion XL 1.0Image1024ร—1024~4.1sS
Playground v2.5Image1024ร—1024~6.1sS
RealVisXL v5.0Image1024ร—1024~4.6sS
DreamShaper XLImage1024ร—1024~4.6sS
Juggernaut XL v9Image1024ร—1024~4.6sS
Animagine XL 3.1Image1024ร—1024~4.6sS
Pony Diffusion V6 XLImage1024ร—1024~4.6sS
Animagine XL 4.0Image1024ร—1024~4.6sS
Illustrious XLImage1024ร—1024~4.6sS
Wan Video 2.1 1.3BVideo480ร—832~3s/frameS
Stable Diffusion 3.5 MediumImage1024ร—1024~7.2sS
Flux.2 Klein 4BImage1024ร—1024~1.2sS
LTX Video 2BVideo1280ร—720~3.6s/frameS
KolorsImage1024ร—1024~8.2sS
Stable CascadeImage1024ร—1024~10.2sS
AuraFlow v0.3Image1536ร—1536~18.4sS
Stable Diffusion 3.5 LargeImage1024ร—1024~22.5sS
Stable Diffusion 3.5 Large TurboImage1024ร—1024~4.1sS
CogVideoX 2BVideo720ร—480~3.6s/frameS
HunyuanVideoVideo720ร—1280~7.5s/frameS
ChromaImage1024ร—1024~4.1sS
Z-Image TurboImage1536ร—1536~4.2sS
Flux.1 DevImage1024ร—1024~18.4sS
Flux.1 SchnellImage1024ร—1024~3.6sS
LTX Video 13BVideo1280ร—720~7.5s/frameS
Flux.1 Kontext DevImage1024ร—1024~20.5sS
AnimateDiff v1.5.3Video512ร—768~1.9s/frameS
Cosmos Diffusion 7BVideo1024ร—576~5.9s/frameS
CogVideoX 5BVideo720ร—480~5.1s/frameS
Wan2.2 TI2V 5BVideo832ร—480~5.1s/frameS
Flux.2 Klein 9BImage1024ร—1024~2sS
Flux.1 Fill DevImage1024ร—1024~17.4sS
Mochi 1 PreviewVideo848ร—480~6.8s/frameS
HunyuanVideo 1.5Video720ร—1280~6.3s/frameS
Helios 14BVideo1280ร—720~7.7s/frameS
SkyReels V2 14BVideo1280ร—720~7.7s/frameS
Wan Video 2.1 14BVideo480ร—832~7.7s/frameA
Wan Video 2.2 14BVideo480ร—832~7.7s/frameA
Qwen ImageImage1024ร—1024~6.9sB
Qwen Image EditImage1024ร—1024~6.9sB
Flux.2 DevImage256ร—256~3m 14sB
MAGI-1Video256ร—256~15.6s/frameB
HunyuanImage 3.0Image256ร—256~12.1sF

Image models estimated at 1024ร—1024 (28 steps, FP16). Video models estimated at 768ร—512 (25 frames, 30 steps, FP16). Actual performance varies with runtime and system load.

Upgrade paths

Upgrade from NVIDIA A16 64GB

See what you unlock with more powerful hardware

Upgrade options

Upgrade options

Mac Studio M3 Ultra 96GBNext step up
96 GB Unified (+32)819 GB/s (+219)
A
Unlocks 1 additional models that do not fit on the current setup.Unlocks Command A 111B+9% faster avg

Unlocks 1 additional models that do not fit on the current setup.

~$3,999 MSRP

๐Ÿ‘ NVIDIA
NVIDIA H100 80GBNVIDIA upgrade
80 GB VRAM (+16)3350 GB/s (+2750)
A
Unlocks 7 additional models that do not fit on the current setup.Unlocks Devstral 2 123B Instruct, Qwen 3.5 122B A10B, Mistral Small 4 119B+4 more ยท +98% faster avg

Unlocks 7 additional models that do not fit on the current setup.

Lifts average decode speed across fitting models by about 98%.

~$40,000 MSRP

MacBook Pro M3 Max 128GBBest value
128 GB Unified (+64)
B
Unlocks 8 additional models that do not fit on the current setup.Unlocks Devstral 2 123B Instruct, Qwen 3.5 122B A10B, Mistral Small 4 119B+5 more

Unlocks 8 additional models that do not fit on the current setup.

~$2,499 MSRP

AMD Instinct MI350X 288GBBiggest leap
288 GB VRAM (+224)8000 GB/s (+7400)
B
Unlocks 21 additional models that do not fit on the current setup.Unlocks Qwen 3.5 397B A17B, Devstral 2 123B Instruct, Qwen 3.5 122B A10B+18 more ยท +164% faster avg

Unlocks 21 additional models that do not fit on the current setup.

Lifts average decode speed across fitting models by about 164%.

~$8,000 MSRP

Frequently Asked Questions

Compare with similar

NVIDIA A16 64GB vs AMD Instinct MI210 64GBNVIDIA A16 64GB vs RTX 6000 Ada 48GBNVIDIA A16 64GB vs RTX PRO 5000 Blackwell 48GB
Compare this GPUCompare with another GPU โ†’