LLM Configurator – Comino Labs
All specs verified from NVIDIA datasheets · 2026

Run Any AI Model
On Your Own Hardware

Liquid-cooled PCIe GPU servers with up to 8 GPUs per node. From RTX 5090 to H200 NVL. Latest models: Llama 4, Qwen 3.5, Gemma 4, Mistral Small 4, DeepSeek R2. Every precision format. Real benchmarks.

8
GPUs per Server
1,128 GB
Max VRAM (8x H200 NVL)
FP4
Native on Blackwell
100%
Boost Clock (Liquid Cooled)

Six GPUs. Every Precision Format.

All TFLOPS numbers shown as Dense / With Sparsity. Precision support verified from NVIDIA official datasheets. Each GPU available in 1x to 8x with Comino liquid cooling.

Blackwell · 5th Gen Tensor

RTX 5090

GB202 · 21,760 CUDA · 680 Tensor Cores
32 GB GDDR7
Bandwidth1,792 GB/s
TDP575W
PCIeGen 5 x16
ECCNo
NVLink / MIGNo / No
TFLOPS (Dense / Sparse)
FP32104.8
FP16/BF16 Tensor209.5 / 419
FP8 Tensor419 / 838*
FP4 Tensor~1,676 / ~3,352
FP4FP6FP8BF16FP16INT8INT4
Best for: Budget multi-GPU inference, local fine-tuning. ~150 tok/s on 8B (Ollama Q4), ~213 tok/s (llama.cpp). *FP8 with FP32 accum runs at half rate on GeForce.
Blackwell · 5th Gen Tensor

RTX PRO 6000

GB202 · 24,064 CUDA · 752 Tensor Cores
96 GB GDDR7 ECC
Bandwidth1,792 GB/s
TDP600W
PCIeGen 5 x16
ECCYes
NVLink / MIGNo / Yes (4 inst)
TFLOPS (Dense / Sparse)
FP32125
FP16/BF16 Tensor~500 / ~1,000
FP8 Tensor~1,000 / ~2,000
FP4 Tensor~2,000 / ~4,000
FP4FP6FP8BF16FP16INT8INT4
Best for: 96GB = run 70B at FP16 on one GPU. No FP8 accumulator restriction (unlike RTX 5090). 12,109 tok/s on 8B FP8 with 8x vLLM (StorageReview). MIG: 4x 24GB instances.
Ada Lovelace · 4th Gen Tensor

RTX 6000 Ada

AD102 · 18,176 CUDA · 568 Tensor Cores
48 GB GDDR6 ECC
Bandwidth960 GB/s
TDP300W
PCIeGen 4 x16
ECCYes
NVLink / MIGNo / No
TFLOPS (Dense / Sparse)
FP3291.1
FP16/BF16 Tensor364 / 728
FP8 Tensor729 / 1,457
FP4 TensorNot supported
FP8BF16FP16INT8INT4 No FP4No FP6
Best for: Power-constrained – only 300W for 48GB ECC. Lowest watts/GB. 4th gen Tensor Cores with FP8 via Transformer Engine.
Ada Lovelace · 4th Gen Tensor

L40S

AD102 · 18,176 CUDA · 568 Tensor Cores
48 GB GDDR6 ECC
Bandwidth864 GB/s
TDP350W
PCIeGen 4 x16
ECCYes
NVLink / MIGNo / No
TFLOPS (Dense / Sparse)
FP3291.6
FP16/BF16 Tensor362 / 733
FP8 Tensor733 / 1,466
FP4 TensorNot supported
FP8BF16FP16INT8INT4 No FP4
Best for: Passive data center cooling. 3x NVENC/NVDEC with AV1. ~114 tok/s 8B Q4 (llama.cpp). Dense rack inference serving.
Hopper · 4th Gen Tensor

H200 NVL

GH100 · 16,896 CUDA · 528 Tensor Cores
141 GB HBM3e ECC
Bandwidth4,800 GB/s
TDP600W
PCIeGen 5 x16
ECCYes (HBM)
NVLink / MIG900 GB/s / Yes (7 inst)
TFLOPS (Dense / Sparse) — NVL variant
FP3260
FP16/BF16 Tensor836 / 1,671
FP8 Tensor1,570 / 3,341
FP4 TensorNot supported
FP8BF16FP16INT8 No FP4No INT4
Best for: 141GB HBM3e + 4.8 TB/s = 70B at FP16 on one card. NVLink bridge (900 GB/s). 42% faster than H100. MIG: 7x ~16.5GB. NVL is ~15% lower clocks than SXM (600W vs 700W).
Ampere · 3rd Gen Tensor

A100 PCIe 80GB

GA100 · 6,912 CUDA · 432 Tensor Cores
80 GB HBM2e ECC
Bandwidth1,935 GB/s
TDP300W
PCIeGen 4 x16
ECCYes (HBM)
NVLink / MIG600 GB/s / Yes (7 inst)
TFLOPS (Dense / Sparse)
FP3219.5
TF32 Tensor156 / 312
FP16/BF16 Tensor312 / 624
INT8 Tensor624 / 1,248 TOPS
FP8 / FP4Not supported
BF16FP16INT8INT4 No FP8No FP4
Best for: Budget data center, multi-tenant MIG (7 instances). Proven ecosystem. ~138 tok/s 8B Q4 (llama.cpp). PCIe variant = 1,935 GB/s (not 2,039 like SXM).

Full Specification Comparison

All numbers from NVIDIA official datasheets. TFLOPS shown as Dense / With Sparsity.

Specification RTX 5090 RTX PRO 6000 RTX 6000 Ada L40S H200 NVL A100 PCIe
ArchitectureBlackwellBlackwellAda LovelaceAda LovelaceHopperAmpere
DieGB202GB202AD102AD102GH100GA100
Tensor Core Gen5th5th4th4th4th3rd
CUDA Cores21,76024,06418,17618,17616,8966,912
Tensor Cores680752568568528432
VRAM32 GB GDDR796 GB GDDR748 GB GDDR648 GB GDDR6141 GB HBM3e80 GB HBM2e
ECC MemoryNoYesYesYesYesYes
Memory BW1,792 GB/s1,792 GB/s960 GB/s864 GB/s4,800 GB/s1,935 GB/s
TDP575W600W300W350W600W300W
PCIe Gen5.05.04.04.05.04.0
NVLinkNoNoNoNo900 GB/s600 GB/s
MIGNo4 instNoNo7 inst7 inst
FP4 NativeNVFP4 NVFP4
FP6 Native
FP8 Native2nd gen2nd gen1st gen1st gen1st gen
FP32 TFLOPS104.812591.191.66019.5
FP16/BF16 Tensor209 / 419~500 / ~1,000364 / 728362 / 733836 / 1,671312 / 624
FP8 Tensor419 / 838~1,000 / ~2,000729 / 1,457733 / 1,4661,570 / 3,341N/A
FP4 Tensor~1,676 / ~3,352~2,000 / ~4,000N/AN/AN/AN/A

Measured Inference Speed

Real tok/s from Ollama, llama.cpp, vLLM, and TensorRT-LLM. Batch 1 = interactive, High batch = production API serving.

8B Model · Single GPU · Interactive (Batch 1)

RTX 5090 Q4
~150 tok/s
Ollama Q4
A100 80GB Q4
~138 tok/s
llama.cpp Q4
L40S Q4
~114 tok/s
llama.cpp Q4
L40S FP16
~44 tok/s
vLLM FP16
Sources: DatabaseMart Ollama benchmark, XiongjieDai GPU LLM benchmark (GitHub), Fluence L40S benchmark.

8B Model · Single GPU · High Batch (vLLM)

H200 NVL
~7,954 tok/s
vLLM offline
A100 80GB
~3,071 tok/s
vLLM offline
H200 online
~2,284 tok/s
vLLM online
A100 online
~1,074 tok/s
vLLM online
Source: E2E Networks benchmark (Qwen3-8B, vLLM). H200's 4.8 TB/s HBM bandwidth gives 2.6x throughput over A100.

Medium Models · RTX 5090 (Ollama Q4)

Gemma 3 12B
~70 tok/s
Ollama Q4
DeepSeek-R1 14B
~89 tok/s
Ollama Q4
Qwen 2.5 32B
~45 tok/s
Ollama Q4
Gemma 3 27B
~47 tok/s
Ollama Q4
Source: DatabaseMart RTX 5090 Ollama benchmark (March 2025). All Q4 quantization, single GPU.

70B+ Models · Multi-GPU

2x 5090 70B Q4
~27 tok/s
Ollama
H200 70B FP8
~341 tok/s
TRT-LLM b64
8xH200 405B
~400 tok/s
TRT-LLM FP8
PRO6000 MoE 120B
~182 tok/s
vLLM single
Sources: DatabaseMart 2x RTX 5090 benchmark, NVIDIA TRT-LLM H200 blog, CloudRift PRO 6000 benchmark.

Time to First Token (TTFT)

Measured with vLLM. Target <200ms for interactive UX.

21-48
milliseconds (mean)
Qwen3-8B (8B params)
H200: 20.7ms · A100: 47.8ms
P99: H200 105ms · A100 192ms
66-188
milliseconds (mean)
Qwen3-32B (32B params)
H200: 66ms · A100: 188ms
P99: H200 378ms · A100 677ms
200-1200
milliseconds
Frontier (405B-671B)
8xH200: ~200ms (405B) · ~563ms (671B)
Scales linearly with prompt length
Source: E2E Networks GPU comparison benchmark (vLLM, Qwen3 models, April 2025)
Independent Third-Party Review

8x RTX PRO 6000 · Measured.

StorageReview independently tested the Comino Grando with 8x NVIDIA RTX PRO 6000 GPUs (768 GB VRAM) in their lab. Real vLLM benchmarks, real thermal data, real noise measurements.

8x GPUs
RTX PRO 6000 (768 GB)
4U
Chassis · 55 kg
8,000W
4x 2000W PSU (80+ Platinum)
EPYC 9474F
AMD 48C/96T · 7x PCIe 5 x16

vLLM Production Benchmarks (Batch 256)

Measured by StorageReview on 8x RTX PRO 6000 Comino Grando. vLLM, batch size 256.

StorageReview Lab · 8x PRO 6000 · vLLM Batch 256

Llama 3.1 8B FP8
12,109 tok/s
8 GPU tensor
GPT-OSS 120B
11,726 tok/s
8 GPU pipeline
MiniMax M2.5 230B
5,753 tok/s
8 GPU tensor

Real-World: Claude Code Sessions on MiniMax M2.5 230B

67.3
tok/s · 1 session
38.7
tok/s · 8 concurrent sessions
309.6
tok/s total · 8 sessions combined

Measured by StorageReview. 230B parameter model, all 768 GB VRAM utilized. 8 users simultaneously coding with near-real-time response.

Measured Thermal & Acoustic Data (StorageReview Lab)

6.5 kW
Cooling Capacity
39 dB
Idle Noise
70 dB
Full Load Noise
450 ml
Coolant Reservoir

3x 140mm radiators, direct-to-chip liquid cooling on all 8 GPUs + CPU. 100% boost clock sustained under full load.

What Fits On Your Server?

Select a GPU, choose count (1-8x), see which latest models (April 2026) fit at every quantization level.

RTX 5090 (32 GB)
PRO 6000 (96 GB)
6000 Ada (48 GB)
L40S (48 GB)
H200 NVL (141 GB)
A100 (80 GB)
1x RTX 5090
Total VRAM: 32 GB
1 GPU2345678 GPUs
Model (April 2026)ParamsFP16/BF16FP8INT4 (AWQ)GGUF Q4_K_MGGUF Q8_0Status
Llama 4 ScoutMoE109B MoE
Apr 2025
220 GB120 GB65 GB61 GB115 GB— More VRAM
Llama 4 MaverickMoE400B MoE
Apr 2025
800 GB420 GB—224 GB230 GB420 GB— More VRAM
Llama 3.3 70B70B
Dec 2024
148 GB—80 GB—42 GB—43 GB—75 GB—— More VRAM
Llama 3.1 405B405B
Jul 2024
810 GB486 GB243 GB250 GB430 GB— More VRAM
Qwen 3.5 397BNEWMoE397B MoE
Feb 2026
794 GB420 GB—230 GB—260 GB—420 GB—— More VRAM
Qwen 3.5 32BNEW32B
Feb 2026
68 GB—36 GB20 GB20 GB36 GB Q4_K_M
Qwen 3 8B8B
Apr 2025
18 GB10 GB6 GB5.5 GB9.5 GB FP16 Full
Qwen 3 72B72B
Apr 2025
152 GB80 GB 44 GB 45 GB 78 GB More VRAM
Gemma 4 31BNEW31B
Apr 2026
66 GB35 GB 19 GB 20 GB 35 GB Q4_K_M
Gemma 4 26B MoENEWMoE26B MoE
Apr 2026
55 GB 30 GB 16 GB “17 GB “28 GB “ FP8
Gemma 3 27B27B
Mar 2025
58 GB 32 GB 17 GB 17 GB 30 GB “ FP8
Mistral Small 4NEWMoE119B MoE
Mar 2026
238 GB 130 GB 70 GB 78 GB 130 GB More VRAM
Mistral Large 3MoE675B MoE
Dec 2025
1350 GB 700 GB 390 GB 400 GB 700 GB More VRAM
DeepSeek R2NEWMoE~1.2T MoE
Mar 2026
2400 GB 1250 GB 700 GB 700 GB 1250 GB More VRAM
DeepSeek V3.2MoE685B MoE
Late 2025
1370 GB 700 GB 400 GB 400 GB 710 GB More VRAM
Phi-4 Reasoning 15BNEW15B
Mar 2026
32 GB 18 GB “10 GB “10 GB “17 GB ““ FP16 Full
Command A111B
Mar 2025
222 GB 120 GB 64 GB 66 GB 118 GB More VRAM
Mixtral 8x22BMoE176B MoE
Apr 2024
264 GB 180 GB 66 GB 80 GB 185 GB More VRAM

Precision Matters.
Choose Wisely.

Quantization compresses model weights from 16-bit to 4-bit. Blackwell GPUs (RTX 5090, PRO 6000) unlock native FP4 – a precision unavailable on Ada/Hopper/Ampere.

16
FP16 / BF16
2.0 bytes/param
All GPUs support this
~100% quality
8
FP8 / INT8
1.0 byte/param · 2x savings
Ada + Hopper + Blackwell only
A100 has no FP8!
97-99% quality
4
FP4 (NVFP4)
0.5 bytes/param · 4x savings
Blackwell ONLY
RTX 5090 + PRO 6000
95-98% quality
~4.5
GGUF Q4_K_M
0.57 bytes/param
Software quant (any GPU)
Most popular GGUF variant
95-97% quality

Which GPU Supports Which Precision?

PrecisionRTX 5090PRO 60006000 AdaL40SH200 NVLA100
FP4 (NVFP4)✓ Native✓ Native
FP6✓ Native✓ Native
FP8✓ 2nd gen✓ 2nd gen✓ 1st gen✓ 1st gen✓ 1st gen
FP16 / BF16
INT8
INT4
GPTQ / AWQ (software)
GGUF (llama.cpp)

Quantization Methods

GPTQ Safetensors

Post-training calibration-based quantization. Groups weights into blocks.

vLLM (Marlin) · TGI · TensorRT-LLM

AWQ Safetensors

Activation-Aware. Protects important weights. Best INT4 quality.

vLLM (Marlin) · TGI · TensorRT-LLM

GGUF .gguf

Universal format. CPU offloading. Q2_K through Q8_0. Runs on any hardware.

Ollama · llama.cpp · LM Studio

EXL2 Custom

Variable bits per layer. Fastest interactive inference on GPU.

ExLlamaV2 · TabbyAPI

NVFP4 Blackwell

NVIDIA native FP4 via TensorRT-LLM. Hardware-accelerated on 5th gen Tensor Cores.

TensorRT-LLM · vLLM · RTX 5090 / PRO 6000 ONLY

HQQ Safetensors

Zero calibration data needed. Quantize in minutes.

vLLM · HuggingFace

Built for Every AI Workload

LLM Inference Serving

Self-hosted OpenAI-compatible APIs with vLLM or TGI. 40-200x lower cost than cloud APIs at 30M+ tokens/day.

7-14B: 1x RTX 5090 (~150 tok/s)
32-72B: 2-4x RTX 5090 or 1-2x PRO 6000
70B FP8: 1x H200 NVL (~341 tok/s batched)
405B/671B: 4-8x H200 NVL

Fine-Tuning (LoRA / QLoRA)

QLoRA makes 70B fine-tuning accessible on a single GPU. PCIe Gen5 bandwidth is adequate.

QLoRA 7-14B: 1x RTX 5090 (32 GB)
QLoRA 70B: 1x A100 80GB or 2x RTX 5090
LoRA 70B: 4x PRO 6000 or 4x A100
Full 7B: 2x A100 or 1x H200 NVL

RAG Pipelines

Embedding + reranker + LLM on one server. Vector DB on CPU/RAM.

Embed + Rerank + 70B Q4: ~50 GB
Recommended: 1-2x L40S or 1x PRO 6000

Image & Video Generation

SDXL, Flux, video models. L40S has 3x NVENC hardware encoders.

Flux.1: 1x RTX 5090 (32 GB)
Multi-user: 2-4x L40S
Professional: 1x PRO 6000 (96 GB)

Speech (ASR + TTS)

Whisper + LLM + TTS on a single GPU.

Full pipeline FP16: ~33-40 GB
Recommended: 1x RTX 5090 or 1x L40S

Code Generation

Self-hosted Copilot with Qwen 3.5, Mistral Small 4, DeepSeek Coder. 8 concurrent Claude Code sessions on 230B model (StorageReview verified).

Single dev: 1x RTX 5090 + 32B Q4
Team (5-20): 2x L40S + 70B model
Enterprise (8 sessions): 8x PRO 6000 + 230B

Multi-Model Serving

Chat + code + image + speech simultaneously. One 8-GPU server replaces 5+ cloud subscriptions.

8x RTX 5090 (256 GB):
GPU 1-2: 72B Q4 ~45 GB
GPU 3-4: Mistral Small 4 ~78 GB
GPU 5-6: Flux + Video ~48 GB
GPU 7: Whisper + Embed ~10 GB
GPU 8: Spare / fine-tuning

Embeddings at Scale

5,000-10,000 docs/sec per L40S with BGE-large.

BGE-large: ~2 GB VRAM
GTE-Qwen2-7B: ~16 GB
Recommended: 1-2x L40S or A100

Liquid Cooling Changes
Everything.

8x 600W GPUs = 4,800W in one server. Air cooling can't handle it. Comino direct-to-chip liquid cooling (6.5 kW capacity) keeps 100% boost clock 24/7.

Air Cooled Industry Standard

GPU Temp85-95 C
Thermal ThrottlingFrequent
Sustained Clock60-80% boost
Noise75-85 dB
Max GPUs / 4U4 (thermal limit)
GPU Lifespan3-4 years

Comino Liquid Direct-to-Chip · Measured

GPU Temp45-55 C
Thermal ThrottlingNever
Sustained Clock100% boost
Noise (idle)39 dB *
Noise (full load)70 dB *
Max GPUs / 4U8 GPUs
Cooling Capacity6.5 kW
Coolant System3x 140mm rad, 450ml
GPU Lifespan5-7 years
* Measured by StorageReview.com in independent lab review of 8x RTX PRO 6000 Comino Grando

Air-cooled: 60-80% sustained †’ Comino liquid: 100% = 25-40% more actual throughput from the same GPUs.

Find Your Config

ScenarioBest ConfigVRAMWhy
Startup, 7-32B2x RTX 509064 GBBest price/perf, native FP4, ~150 tok/s per GPU
70B comfortably2x PRO 6000192 GB70B FP16 + room for context. ECC, MIG
Dense rack inference4-8x L40S192-384 GBPassive cooling, data center grade
Max single-GPU1x H200 NVL141 GB70B FP16 on 1 card, 4.8 TB/s, 21ms TTFT (8B)
DeepSeek R2 / 405B6-8x H200 NVL846-1,128 GB~400 tok/s on 405B (TRT-LLM FP8)
QLoRA fine-tuning 70B1x A100 80GB80 GBProven, 300W, MIG for multi-tenant
Full AI stack8x PRO 6000768 GBChat + Code + Image + Speech. 12,109 tok/s on 8B (StorageReview verified)
Power-constrained8x 6000 Ada384 GBOnly 2,400W total. Lowest W/GB ratio
8-dev code team (230B)8x PRO 6000768 GB38.7 tok/s per session, 8 concurrent (StorageReview)

MoE Models: VRAM Trap

MoE models load ALL expert weights into VRAM. Only active params compute per token → fast inference, but huge VRAM footprint.

Mistral Small 4: 119B total / 6.5B active → ~70 GB INT4 | Llama 4 Scout: 109B / 17B active → ~65 GB INT4 | Qwen 3.5 397B: 397B / 17B active → ~230 GB Q4 | DeepSeek R2: ~1.2T / ~78B active → ~700 GB Q4

Ready to Build Your
AI Infrastructure?

Tell us your workload, models, and budget. Our engineers design the optimal Grando configuration – GPU, quantization strategy, and cooling included. Up to 8 GPUs per server, verified by StorageReview.