[ INFERENCE FIT-CHECK ]

Will your model run on SmartTec?

Pick a model and a GPU. Get VRAM fit, throughput, max concurrent sequences, and a ready-to-run vLLM command.

Run the fit-checkSee pricing
1 · Pick your model
▼

Strong general-purpose. ~140 GB at FP16.

2 · Pick your GPU
▼

8 GPUs per node via NVLink + IB · 80 GB VRAM · 495 BF16 TFLOPS · 3,350 GB/s BW

3 · Quantization
▼

Half the VRAM. ~99% quality retention.

4 · Sequence length & batch
VRAM fit
✓ Fits on 1 GPU
75.0 GB used (71.3 GB at 1 seq) of 80 GB total · 73.6 GB usable
0 GB74 GB usable80 GB (H100)
Memory breakdown
Model weights (FP8)70.0 GB
KV cache (4 × 4k)5.0 GB
Total75.0 GB
Est. throughput
57
tokens / sec
Max concurrent
2
sequences @ 4k ctx
Auto-config · best way to run this model
1× H200[✓ BEST VALUE]
$2,336/mo · ~82 tok/s · 130 GB capacity
2× H100
$3,504/mo · ~115 tok/s · 147 GB capacity
1× B200
$3,504/mo · ~137 tok/s · 177 GB capacity
1× GB200[FASTEST]
$252,288/mo · ~9,874 tok/s · 12718 GB capacity
Counts rounded to deployable tensor-parallel sizes (2/4/8 per node, then whole nodes). GB200 NVL72 shown for scale reference (Phase-1 fleet: B200 / H-class + CS-3). Cerebras CS-3 billed per token. AMD MI455X selectable above for Phase 2/3 fit exploration — priced configs publish at OEM availability.
Deploy command
# vLLM — Llama 3.1 70B on 1× H100
# Pricing: ~$2.40/hr · ~$1752/mo 24/7
docker run --gpus all --shm-size 1g \
  -p 8000:8000 \
  -v ~/.cache/huggingface:/root/.cache/huggingface \
  vllm/vllm-openai:latest \
  --model llama-3.1-70b \
  --tensor-parallel-size 1 \
  --max-model-len 4096 \
  --quantization fp8 \
  --gpu-memory-utilization 0.92

Tensor-parallel = 1. Adjust --max-model-len and --gpu-memory-utilization for your batch / latency tradeoff.

[ Export this fit-check ]
[ FLEET BUILDER · MULTI-SERVER COST ]

Stack as many servers as you need.

Pick H100s, H200s, B200s, GB200 racks, and Cerebras CS-3 systems in any combination. See the run cost on SmartTec, what you could rent it for, and your gross margin — all live. Phase 2/3 cost-alternative track: AMD Instinct MI455X (432 GB HBM4, 19.6 TB/s — 2.4× the memory bandwidth of B200) is now selectable in the fit-check for planning; priced configurations publish once OEM availability and ROCm serving mature.

[ FLEET BUILDER · MULTI-SERVER COST ]
Hours / month730
1 hr burst24/7 (730)
GPU
Quantity
VRAM total
640 GB
tok/s
14.4K
Run cost · 730h
$10.5K
GPU
Quantity
VRAM total
1,128 GB
tok/s
22.4K
Run cost · 730h
$14.0K
Cost to run on SmartTec
$24.5K
730h × 16 GPUs · SmartTec launch
You could rent it for
$32.7K
On-demand retail · 96710.40M tok/mo
MARGIN
Gross margin
$8,176
25.0% margin · 11 kW
Total GPUs
16
Total VRAM
1,768 GB
Peak power
11 kW
Tokens / sec
36.8K
What this fleet fits
Training jobs
✓ Small / fine-tuning workloads
Concurrent users
1,226 sustained users~30 tok/s per user
Power
11 kW · behind-the-meterSub-10ms failover · no grid queue
Design-partner pricing locks launch rates 12 months · Q4 2026 power-on
Get fixed quote →
[ THE HARDWARE ]

Five systems. One fabric.

What you see below is what SmartTec actually runs. Three NVIDIA systems + one wafer-scale engine, on the same power fabric.

[ NVIDIA STACK · HOPPER + BLACKWELL ]
NVIDIA H200 SXM5 · 141 GB
[ NVIDIA ]

H200 SXM5 · 141 GB

BF16 TFLOPS
989
MEM BW
4.8 TB/s
VRAM
141 GB

The Hopper refresh. Same compute, 76% more memory bandwidth than H100. The sweet spot for 70B / 8B-class models at 8k–128k context.

NVIDIA HGX B200 · 8× Blackwell
[ NVIDIA ]

HGX B200 · 8× Blackwell

BF16 TFLOPS
18,000
PER-GPU VRAM
192 GB
BASEBOARD VRAM
1.5 TB

The Blackwell baseboard. 8× B200 SXM, 1.5 TB total HBM3e per board. Where most frontier inference starts.

NVIDIA GB200 NVL72 · exascale rack
[ NVIDIA ]

GB200 NVL72 · exascale rack

BF16 TFLOPS
1.4 exaFLOPS
GRACE + BLACKWELL
36 + 72
NVLINK DOMAIN
130 TB/s

A single rack. 36 Grace CPUs, 72 Blackwell GPUs, one NVLink domain. What trillion-parameter inference runs on.

[ CEREBRAS STACK · WAFER-SCALE ]
Cerebras CS-3 system · 1× WSE-3
[ CEREBRAS ]

CS-3 system · 1× WSE-3

BF16 TFLOPS
125,000
ON-CHIP SRAM
44 GB
MEM BW
21 PB/s

One chip, one wafer, one system. The largest silicon ever built. Per-1K-token billing, not per-hour.

Cerebras CS-3 in production
[ CEREBRAS ]

CS-3 in production

POWER DRAW
25 kW
CORES
900,000
TRANSISTORS
4 trillion

What a CS-3 actually looks like in a rack. Pink fiber = wafer-scale fabric. Front mesh = the cooling intake.

Scale WSE-3 vs B200 · at-scale
[ SCALE ]

WSE-3 vs B200 · at-scale

WSE-3 AREA
46,225 mm²
B200 DIE
~ 800 mm²
RATIO
56× larger

The wafer on the left holds one die. The wafer on the right is the entire NVIDIA B200 chip, scaled up for visibility.

[ IMAGES · SOURCE ]

NVIDIA hardware photography sourced from nvidia.com/data-center (NVIDIA press kit, public marketing assets). Cerebras hardware photography sourced from cerebras.ai (CS-3 product page, public marketing assets). The wafer-vs-B200 scale comparison is from the Cerebras WSE-3 announcement. These exact systems deploy at our Mead, OK facility at Q4 2026 power-on.

[ HOW THE MATH WORKS ]

Three numbers, all open.

1 · Weights
params × bytes-per-param

Llama 3 70B at FP16 = 70B × 2 bytes = 140 GB. Quantized to FP8 = 70 GB. We add a 5% buffer for activations and CUDA kernels.

2 · KV cache
2 × L × Hk × d × ℓ × B

Per sequence at seq_len ℓ and batch B. Llama 3 70B at 8k context, batch 8 = ~10 GB. Linear in batch size.

3 · Throughput
roofline of memory & compute

Decode is memory-bound: tokens/s ≈ BW / (params × bytes). Prefill is compute-bound: tokens/s ≈ TFLOPS / (2 × params). We take the minimum.

Sources & assumptions
  • Model architecture (layers, KV heads, head dim) sourced from each model's published config.json on HuggingFace.
  • GPU specs (VRAM, BF16 TFLOPS, memory bandwidth) from NVIDIA / Cerebras datasheets — Blackwell numbers per vendor release notes.
  • Pricing reflects June 2026 on-demand retail market rates (CoreWeave, Lambda, AWS p5, GCP a3 ranges). Subject to change.
  • Throughput is a roofline estimate — actual achieved tokens/sec depends on scheduling, batching policy, attention impl (FlashAttention-2 vs 3), and KV cache compression.
  • Recommendation: always benchmark your specific workload. This calculator tells you what fits; how fast is the next 10% engineering.
[ NEXT STEP ]

Fit confirmed. Now scope the build.

Send this fit-check to partners@smarttec.dev. We'll send back a deployment plan, Q4 2026 power-on target, and design-partner pricing locked for 12 months.

Reserve compute