AI Hardware Calculator Singapore
Free tool Β· Singapore pricing in SGD
Which AI server or workstation runs your model?
Work out which AI server or workstation actually runs your model β and whether buying it beats paying for cloud API calls. Tell us the model and your current spend; we compare 26 systems from NVIDIA, AMD and Apple on memory, tokens per second, power draw and total cost, priced in Singapore dollars at the current SP Group tariff.
- 26systems compared, from a mini PC to a GB300 NVL72 rack
- 9vendors β NVIDIA, AMD, Apple, HP, Dell, Lenovo, ASUS, Supermicro, Gigabyte
- SGDpriced with GST, duty cycle and the live SP Group tariff
- No formyou see the number before you talk to us
Sizing AI hardware comes down to three questions. Will the model fit in memory? Will it run fast enough to be useful? And does owning the machine cost less than renting the equivalent capacity from a cloud API provider? Most vendor pages answer the first and skip the other two.
This calculator answers all three. It covers desktop systems like the NVIDIA DGX Spark and GB10 class, Apple Mac Studio and AMD Ryzen AI Max+ 395; configurable tower workstations from HP, Dell and Lenovo; rack GPU servers from ASUS, Supermicro and Gigabyte; and datacentre platforms including AMD Instinct MI300X/MI325X/MI355X nodes and the NVIDIA GB300 NVL72. Every price is indicative Singapore retail and every calculation is documented below, so you can check the working.
What do you want to run?
Pick the model and how it will be used. Everything else is calculated.
What are you spending today?
This is what the hardware has to beat.
Which systems will run it
| Compare | System | Memory needed | Est. speed | Daily capacity | Verdict | Power | Singapore price |
|---|
Side by side
Own it or keep renting?
Cumulative cost over 36 months
View the numbers as a table
| Month | Cloud API cumulative | On-prem cumulative | Difference |
|---|
Before we quote datacentre-class hardware
A server node or rack isn't a cart purchase. Four things decide whether your site can host one β worth checking before we talk price.
Power
Budget roughly 120β150Β kW per rack. That is more than most Singapore commercial floors deliver to a single rack position without an upgrade.
Cooling
Direct liquid cooling is required. You'll need a CDU, secondary loop and facility water β air cooling is not an option at this density.
Floor loading
A populated rack lands around 1,400β1,600Β kg. Raised floors and delivery routes need checking, including lift capacity.
Networking
Scale-out beyond one rack needs InfiniBand or Spectrum-X Ethernet fabric planned in from the start, not retrofitted.
How these numbers are calculated
Every figure here is an estimate you can check. Nothing is hidden.
Memory required
- Weights = total parameters Γ bits-per-weight Γ· 8. Mixture-of-experts models need all experts resident, so total parameters is what counts, not active.
- KV cache = 2 Γ layers Γ KV heads Γ head dimension Γ context Γ concurrent sessions, at FP16. Models using multi-head latent attention (DeepSeek) use their compressed KV size instead.
- Overhead β 5% on weights for activations, plus ~1.5Β GB runtime.
- Usable memory is below nameplate: the OS and driver reserve a slice.
Token generation speed
Single-stream decode is bound by memory bandwidth, not compute. Each token requires reading the active weights plus the KV cache:
tokens/sec β (bandwidth Γ efficiency) Γ· (active weight bytes + KV read)
- Efficiency: 0.75 for dense models, 0.55 for MoE (expert routing overhead).
- Multi-accelerator systems shard the model across GPUs, so bandwidth scales with GPU count β discounted for the interconnect. We assume 70% efficiency over NVLink or Infinity Fabric and 50% over PCIe, which is why a four-card workstation does not deliver four times one card.
- On tiered-memory systems, a model that exceeds the fast HBM tier spills to slower memory and effective bandwidth becomes the harmonic blend of both.
- Above ~250 tokens/sec a single stream stops being bandwidth-bound and hits kernel-launch limits, so we cap the headline figure β the surplus is real, it just shows up as concurrency rather than one faster stream.
- Daily capacity assumes a 60% duty cycle β real deployments are not saturated 24 hours a day.
These are estimates, not benchmarks. Actual throughput depends on your runtime, batching and prompt shape. We can benchmark your model on a live unit before you commit.
Costs and break-even
- Electricity β 31.91Β Β’/kWh before GST, the SP Group regulated tariff for 1 Julyβ30 September 2026. Editable below.
- GST β 9%, applied to hardware and power.
- Power draw β system nameplate at the assumed duty cycle, not idle.
- Break-even is the month when cumulative cloud spend passes hardware capex plus cumulative running cost.
- Prices marked indicative are street estimates for planning only; prices without a marker are current SourceIT list. Either way, a written quotation is the only firm price β allocation and stock move weekly at this end of the market.
- Not costed here: installation, networking, racking, UPS, cooling upgrades, extended warranty or support contracts. For anything above workstation class those are a real share of the project, and we'll itemise them in the quotation.
How much memory does each model need?
Memory is the first gate: if the weights plus the KV cache don't fit, the model simply won't load. These are the requirements at 8K context for a single session β the calculator above adjusts them for your context length and concurrency.
| Model | 4-bit (Q4_K_M) | 8-bit (Q8_0) | FP16 | Smallest system that holds it at 4-bit |
|---|---|---|---|---|
| Llama 3.1 8B | 7 GB | 11 GB | 18 GB | AMD Ryzen AI Max+ 395 |
| GPT-OSS 20B | 14 GB | 24 GB | 43 GB | AMD Ryzen AI Max+ 395 |
| Mistral Small 24B | 17 GB | 28 GB | 50 GB | AMD Ryzen AI Max+ 395 |
| Qwen3 32B | 22 GB | 37 GB | 66 GB | AMD Ryzen AI Max+ 395 |
| Llama 3.3 70B | 45 GB | 77 GB | 141 GB | AMD Ryzen AI Max+ 395 |
| GPT-OSS 120B | 71 GB | 124 GB | 231 GB | NVIDIA GB10 class |
| Qwen3 235B-A22B | 142 GB | 247 GB | 463 GB | Mac Studio M3 Ultra |
| Llama 3.1 405B | 245 GB | 426 GB | 798 GB | 4 Γ RTX PRO 6000 workstation |
| DeepSeek V3 / R1 671B | 398 GB | 699 GB | 1.31 TB | NVIDIA DGX Station GB300 |
Figures include a 5% activation allowance and 1.5Β GB of runtime overhead. Mixture-of-experts models such as GPT-OSS 120B and DeepSeek R1 need every expert resident in memory even though only a fraction runs per token β which is why a 117B MoE model needs as much memory as a 117B dense one, but runs far faster.
AI hardware prices in Singapore
Indicative Singapore pricing across the range we supply, from a desktop system to a liquid-cooled rack. Prices exclude GST unless noted and move with NVIDIA and AMD allocation β a written quotation is the only firm price.
| System | Class | Memory | Indicative SGD |
|---|---|---|---|
| AMD Ryzen AI Max+ 395 (Strix Halo) | Mini PC | 128 GB unified | From S$3,200 |
| Apple Mac Studio M4 Max | Desktop | 128 GB unified | From S$4,500 |
| NVIDIA DGX Spark & GB10 systems | Desktop AI appliance | 128 GB unified | S$6,895 β S$10,385 |
| Apple Mac Studio M3 Ultra | Desktop | 256 GB unified | From S$10,000 |
| HP Z6 G5 / Dell Precision 7875 / Lenovo P8 | Tower workstation | Up to 3 GPUs | From S$18,300 configured |
| HP Z8 Fury G5 / Dell 7960 / Lenovo PX | Tower workstation | Up to 4 GPUs | From S$24,800 configured |
| ASUS ESC8000A-E13P / Supermicro SYS-522GA-NRT | 4Uβ5U GPU server | Up to 8 GPUs | From S$48,800 configured |
| NVIDIA DGX Station GB300 | Deskside | 748 GB coherent | On application |
| AMD Instinct MI300X / MI325X / MI355X Γ8 | Server node | 1.5 β 2.3 TB HBM | On application |
| NVIDIA GB300 NVL72 | Rack scale | 20 TB HBM3e | On application |
How to choose AI hardware in Singapore
Start with memory, not TFLOPS
Inference speed for a single user is governed by memory bandwidth, not raw compute. A machine with enormous FP4 throughput and not enough memory will not run your model at all, while a modest system with plenty of unified memory will run it slowly but usefully. Size for capacity first, then check the speed is tolerable.
Match the speed target to the job
Interactive chat and coding copilots need roughly 20 tokens per second to feel responsive β around reading speed. Agentic and batch workloads that run unattended are perfectly usable from about 8 tokens per second. Prototyping tolerates less. Buying for 100 tokens per second when 10 will do is the most common way to overspend.
Know the difference between the 300W and 600W RTX PRO 6000
This is the single most misquoted specification in the category. Most current tower workstations β including the Dell Precision 7960 and Lenovo ThinkStation PX β accept only the 300W Max-Q version of the RTX PRO 6000 Blackwell, not the 600W Workstation Edition. Memory bandwidth is identical on both, so inference speed is effectively the same; the 600W card wins on compute-bound work such as fine-tuning and prompt processing. Specifying the wrong one is a returns problem. The calculator above filters the list correctly when you switch cards.
CUDA is a real constraint
AMD Instinct on ROCm and Apple Silicon on Metal both run mainstream open-weight models well, and often at better capacity per dollar. But if your team depends on CUDA-only libraries, custom kernels or a vendor's NVIDIA-specific stack, that advantage evaporates. Decide this before comparing prices, not after.
Check the room before the rack
An 8-GPU server draws 10β14 kW and a GB300 NVL72 rack draws 120β150 kW with mandatory liquid cooling. Most Singapore commercial floors do not deliver that to a single rack position without an upgrade, and a populated rack lands around 1,400β1,600 kg. We run a site readiness review covering power, cooling, floor loading and delivery routes before quoting anything at that scale.
Data residency often decides it
Break-even is the financial argument, but it is rarely the deciding one. On-premises inference means prompts and documents never leave your building. For PDPA-regulated personal data, MAS-supervised financial workloads, legal privilege or client confidentiality, that is frequently the whole case β and it appears nowhere in the cost comparison.
Frequently asked questions
How much VRAM do I need to run Llama 70B?
About 45Β GB at 4-bit quantization with an 8K context window, 77Β GB at 8-bit, or 141Β GB at full FP16 precision. A 128Β GB unified-memory system such as the NVIDIA DGX Spark, Mac Studio M4 Max or AMD Ryzen AI Max+ 395 will hold the 4-bit version comfortably.
How much does an NVIDIA DGX Spark cost in Singapore?
The NVIDIA DGX Spark 4TB Founders Edition is S$8,800 at SourceIT. Equivalent GB10 systems from GIGABYTE, Lenovo, Acer and Dell range from S$6,895 to S$10,385 depending on storage and brand. All are in stock locally with local warranty and GST invoicing.
Is it cheaper to buy AI hardware or use cloud APIs?
It depends almost entirely on volume. At around S$4,000 a month of API spend, a desktop-class system typically pays for itself within one to three months. Below roughly S$500 a month, renting is usually cheaper and we will tell you so. The calculator above works out your specific break-even point using your actual spend, Singapore electricity tariffs and GST.
Can I run DeepSeek R1 671B locally?
Yes, but not on a desktop. At 4-bit it needs roughly 398Β GB, which rules out every 128Β GB system. The practical options are an NVIDIA DGX Station GB300 with 748Β GB of coherent memory, a multi-GPU server, or an AMD Instinct node. Because DeepSeek uses multi-head latent attention its KV cache is unusually small for its size, so long contexts cost less memory than you would expect.
What is the difference between the RTX PRO 6000 Blackwell Max-Q and Workstation Edition?
Both carry 96Β GB of GDDR7 at 1,792Β GB/s. The Max-Q is capped at 300W and fits most current tower workstations; the Workstation Edition runs at 600W and needs a chassis rated for it, such as the HP Z8 Fury G6i, Dell Pro Max Precision 9 T6 or Lenovo ThinkStation P7/P8. Since inference speed is bound by memory bandwidth, the two perform almost identically for running LLMs β the 600W card pulls ahead on training and prompt processing.
Which AI workstation is best for LLM inference?
For models up to about 90Β GB, a single RTX PRO 6000 in any supported tower gives the best single-stream speed per dollar. Beyond that you need either multiple GPUs β an HP Z8 Fury G5, Dell Precision 7960 or Lenovo ThinkStation PX each take four β or a large unified-memory system. The right answer depends on your model, so use the calculator rather than a generic recommendation.
Do you supply AMD Instinct in Singapore?
Yes. We supply MI300X, MI325X and MI355X in 8-GPU node configurations, giving 1.5Β TB, 2.05Β TB and 2.3Β TB of HBM respectively. They offer strong capacity per dollar against comparable NVIDIA nodes, with the trade-off that they run ROCm rather than CUDA.
What power and cooling does a GB300 NVL72 need?
Budget 120β150Β kW per rack with direct liquid cooling β a CDU, secondary loop and facility water. A populated rack weighs approximately 1,400β1,600Β kg, so raised floors, lift capacity and delivery routes all need checking. We carry out a site readiness review before quoting.
How fast will my model actually run?
Single-stream decode speed is roughly memory bandwidth divided by the bytes read per token. A 70B model at 4-bit on a 273Β GB/s system gives about 5 tokens per second; the same model on a 1,792Β GB/s GPU gives roughly 33. The calculator estimates this for every system, and we can benchmark your exact model on a live unit before you commit.
Do you sell to Singapore government agencies?
Yes. SourceIT is GeBIZ registered and transacts through Vendors@Gov and SAP Ariba, with GST invoicing and purchase order support. Public-sector procurement for AI hardware follows the same process as our AV and IT business.
Is the pricing on this page final?
No. Prices shown are indicative Singapore retail for planning purposes and move with vendor allocation and stock. Items marked "indicative" are street estimates; items marked "on application" depend on configuration and allocation. Request a quotation for firm pricing, lead time and warranty terms.
Can I try before I buy?
We can benchmark your specific model and quantization on a unit we hold in stock and send you the measured tokens per second before you commit. Contact sales@sourceit.com.sg or call +65 6978 3502.
SourceIT Pte Ltd Β· 178 Paya Lebar Road, Singapore Β· +65 6978 3502 Β· sales@sourceit.com.sg
Estimates are indicative and provided for planning purposes. Specifications follow NVIDIA published figures. Prices exclude delivery and installation unless quoted.