modelrigs

Phi-3.5 MoE 42B-A6.6B on NVIDIA A100 PCIe 40 GB

41.873B (6.6B active) on 40 GB at 1,555 GB/s. Hardware details · Model details

Yes. Phi-3.5 MoE 42B-A6.6B fits on NVIDIA A100 PCIe 40 GB at the recommended Q4_K_M configuration, requiring approximately 27.1 GB at 8K context. Practical context capacity is 64K.
CompatibilityComfortable
Calculated fit
Yes
Calculated Uses under 80% of usable memory. Room for a long context and other work.
Recommended quantization
Q4_K_M
Calculated gguf
Memory required
27.1 GB
Calculated of 36.8 GB usable — 74%
Max practical context
64K
Calculated model supports 128K

Memory budget at 8K context

Model weights ⓘ
24.4 GB Calculated
KV cache ⓘ
1.0 GB Calculated
Runtime overhead ⓘ
1.7 GB Estimated
Total required
27.1 GB Calculated
Headroom ⓘ
9.7 GB Calculated
Highest-precision quantization that leaves headroom: uses 74% of usable memory at 8K context.
Discrete GPU: 40 GB of VRAM, of which we assume 92% is usable after driver and context overhead.
Mixture of experts: all 41.873B parameters must be resident in memory even though only ~6.6B are active per token. Memory follows total parameters; speed follows active parameters.
Performance

Estimated performance · recommended Q4_K_M

Decode, prefill and TTFT below are estimates for Q4_K_M. We do not have a comparable Q4_K_M measurement on this machine.

Estimated decode
—
Estimated Q4_K_M; calculated range
Estimated prefill
—
Estimated Q4_K_M; calculated range
Estimated TTFT at 8K
~1.1–2.2 s
Estimated Q4_K_M; calculated range
Hardware load reference
250 W
Official spec hardware TDP; not this model run
No throughput figures for this configuration. We do not estimate performance for a model that cannot load.
How the estimate is calculated
  1. Decode: reading 6.6B active parameters at 4.85 bits/weight takes 3.30 ms at 1555 GB/s x 78% achieved efficiency.
  2. MoE routing penalty of 15% applied: expert gathers are less bandwidth-efficient than a dense sweep.
  3. Prefill: 312 TFLOPS (FP16) x 2 for native FP8 tensor cores x 32% assumed model-FLOPs utilisation, divided by 2 x 16.6B parameters per token.
  4. Prefill uses 16.6B effective parameters, not the 6.6B active in decode: a batch of hundreds of tokens routes across most of the expert pool.
  • MoE decode depends on how well the engine batches expert gathers; real results vary more than for dense models.
  • Prefill throughput is highly engine-dependent. Flash attention, batch size and quantized KV all move this number substantially, and for sparse mixture-of-experts models it is the least reliable figure we produce.
  • This is a calculated estimate, not a measurement. It assumes a single request, a short prompt, no speculative decoding and a warm model already resident in memory.

Estimator version estimator_v1. Stored with every estimated row so old estimates can be regenerated when the model improves.

Submit a benchmarkSubmissions stay pending until reviewed.
This model at other quantizations on this machine
QuantizationResidentDiskTotal neededUtilisationFitMax context
Q4_K_Mrecommended24.4 GB24.4 GB27.1 GB74%Comfortable64K
Q5_K_M28.6 GB28.6 GB31.4 GB85%Fits32K
Q8_042.3 GB42.3 GB45.5 GB124%Does not fit0
BF1678.0 GB78.0 GB82.3 GB224%Does not fit0
Cost of running Phi-3.5 MoE 42B-A6.6B on this machineUnited States (federal) · C corporation · 8h/day
Monthly economic cost
$1
Calculated after tax
Codex/Claude Code
$100/mo
≈ $100/mo · local is $99 less
Net cash at purchase
$0
Calculated VAT not reclaimable
Total over 5 years
$86
Calculated after tax, after resale
Cost per USD/1M tokens
—
Calculated

The $100 comparison uses the Codex Pro 5x / Claude Max 5x plans. A fixed planning conversion is used for USD. This compares monthly spend only: subscriptions have usage limits, local hardware has different capabilities and constraints, and taxes or regional pricing may change the charged amount. Prices checked 7 September 2026.

Monthly breakdown

Depreciation
0 over 5 years, straight-line to a 0 residual
$0.00
Electricity
13.0 kWh/month at 0.140/kWh
$1.82
Monthly cost before tax$1.82
Electricity tax shield
Running costs are deductible business expenses
−$0.38
Monthly economic cost after tax$1.44

Three different numbers, deliberately

Cash cost
$0
Money that leaves the bank account on day one, net of reclaimable VAT.
Accounting depreciation
$0.00/month
$0 written down over 5 years to a $0 residual.
After-tax economic cost
$1.44/month
Depreciation plus running costs plus cost of capital, less the tax those deductions save. This is the figure to compare between machines.
Assumptions and sources (verified 2026-09-07)
VAT rate
0% Official spec
VAT recoverable
0% Official spec
Federal C-corporation estimate at the flat 21% rate. Pass-through entities should use the sole-proprietor estimate as a rougher proxy.
Effective deduction rate
21.00% Calculated
Headline marginal rate 21.00%.
Depreciation
5 years, straight-line Assumption
Residual value
$0 (15%) Assumption
Two GPU generations later, the card is worth a fraction of its list price.
Electricity
$0.140/kWh Assumption
Average power draw
74 W Calculated
Load 250 W for 20% of powered hours, idle 30 W for the rest — a machine that is on is not generating tokens the whole time.
Investment allowances
§179 100.0% Official spec
Worth $0 in total — first-year expensing that replaces later tax depreciation.
Cost of capital
not applied Assumption
Federal planning estimate only, not tax advice. State and local income tax, sales/use tax and incentives are excluded. It assumes 100% business use, a Section 179 election, enough business income to use it, and that the full annual limit remains available.
Phi-3.5 MoE 42B-A6.6B on other hardware
HardwareDecodePriceFit
Nearest alternatives to the NVIDIA A100 PCIe 40 GB
Phi-3.5 MoE 42B-A6.6B on Dual used RTX 4090 workstation (48 GB)48 GBPhi-3.5 MoE 42B-A6.6B on NVIDIA RTX 6000 Ada 48 GB48 GBPhi-3.5 MoE 42B-A6.6B on NVIDIA RTX A6000 48 GB48 GBPhi-3.5 MoE 42B-A6.6B on NVIDIA RTX 5000 Ada 32 GB32 GBPhi-3.5 MoE 42B-A6.6B on RTX 5090 workstation (1x 32 GB)32 GB
Compare NVIDIA A100 PCIe 40 GB against Dual used RTX 4090 workstation (48 GB) →