Skip to content

OnpremBench.aiby Understand.tech

Find my machine

Loading the finder…

Tools / Model fit finder

Plan your local AI workload.

Choose a model, context and simultaneous requests. Compare memory estimates, save your plan, then validate on a real machine.

Your model and workload

Model

Weight format

64 layers · 8 KV heads · head size 128. Model card. Scales and higher-precision layers make 4-bit formats cost about 4.3 to 4.9 bits per parameter; 4.6 is used.

Context per request

KV cache precision

Simultaneous requests 4Requests being generated at the same moment, not registered users. The capacity planner on each machine page turns users into this number.

Constraints

Hardware budget in US dollars · 0 = no limit

Checked against supplier list prices, converted to dollars at an indicative rate where the offer is in euros or pounds. Machines without a public price stay in the list.

NVIDIA CUDA required

Use all GPUs on multi-GPU towers (tensor parallel)

Workstations only (hide edge boards, accelerator chassis, prototypes)

Keep this workload

Save every input privately, or share the model, context and concurrency. Shared links exclude your budget.

24 of 24 machines fit

Lowest supplier list price first; machines without a public price come last. Speed figures are bandwidth ceilings for decode at this context, not measurements; real throughput is lower and depends on the runtime and kernels.

Framework Desktop · Ryzen AI Max+

Framework · AMD Ryzen AI Max

Cheapest that fits

Desktop · Ryzen AI Max+

47 GB of 128 GB · 69 GB headroom

Decode ceiling, one stream: bandwidth not published
All 4 streams batched: bandwidth not published
Supplier list price: ≈ $3.4k · Supplier list price · United States · Sept 2026
Evidence on this platform: 2 model configurations

Models on this machine Cost of ownership

MINISFORUM MS-S1 MAX

MINISFORUM · AMD Ryzen AI Max

MS-S1 MAX

47 GB of 128 GB · 69 GB headroom

Decode ceiling, one stream: bandwidth not published
All 4 streams batched: bandwidth not published
Supplier list price: ≈ $3.8k · Supplier list price · United States · Sept 2026
Evidence on this platform: 2 model configurations

Models on this machine Cost of ownership

NVIDIA DGX Spark

NVIDIA · GB10

DGX Spark

47 GB of 128 GB · 69 GB headroom

Decode ceiling, one stream: ≈ 11.8 tok/s
All 4 streams batched: ≈ 30.3 tok/s
Supplier list price: ≈ $4.7k · Supplier list price · United States · Sept 2026
Evidence on this platform: 6 model configurations

Models on this machine Cost of ownership

ASUS Ascent GX10

ASUS · GB10

Ascent GX10

47 GB of 128 GB · 69 GB headroom

Decode ceiling, one stream: ≈ 11.8 tok/s
All 4 streams batched: ≈ 30.3 tok/s
Supplier list price: ≈ $5.3k · Supplier list price · France · LDLC · Sept 2026 · converted from EUR
Evidence on this platform: 6 model configurations

Models on this machine Cost of ownership

Acer Veriton GN100

Acer · GB10

Veriton GN100

47 GB of 128 GB · 69 GB headroom

Decode ceiling, one stream: ≈ 11.8 tok/s
All 4 streams batched: ≈ 30.3 tok/s
Supplier list price: ≈ $5.5k · Supplier list price · United States · Sept 2026
Evidence on this platform: 6 model configurations

Models on this machine Cost of ownership

MSI EdgeXpert

MSI · GB10

EdgeXpert

47 GB of 128 GB · 69 GB headroom

Decode ceiling, one stream: ≈ 11.8 tok/s
All 4 streams batched: ≈ 30.3 tok/s
Supplier list price: ≈ $6k · Supplier list price · United States · Sept 2026
Evidence on this platform: 6 model configurations

Models on this machine Cost of ownership

Showing 6 of 24 matching machines. Exports include every result.

Compare the first 4

MEMORY DEMAND

37GB + reserve

Weights · 32.8B × 4.6 bits: 19.4 GB
KV cache · 256 KB/token × 16K × 4: 17.2 GB
Runtime + OS reserve: 6 GB discrete · 10 GB unified
Fits when: demand ≤ 90% of the memory tier

GB300 machines are screened on their 252 GB GPU tier; RTX PRO towers on one 96 GB card unless multi-GPU is allowed. No CPU offload is assumed.

HOW THE SPEED CEILING IS ESTIMATED

Generating a token reads the active weights plus the request’s KV cache from memory. Ceiling = bandwidth ÷ bytes read per token. With 32.8B active parameters at 4.6 bits and 16K context, one stream reads about 23.2 GB per token. Batching several streams re-uses the weights, which is why servers get far more total tokens per second than one user sees. Prefill and real kernels reduce these ceilings.

Architecture figures come from the model cards linked above. Memory needs also depend on the runtime’s allocator, kernels and any speculative decoding. Test before you buy.

Read the full method