Choose my GPU
Answer nine questions. We rank the GPUs in our catalog by fit for local AI on your stack — top picks, alternates, and what to avoid. Hand-written rationale per card, honest caveats, and a one-click handoff into the custom build engine.
We don’t fake tok/s numbers. Every recommendation cites a model class and a workload-realistic range. Cards over your budget appear last with explicit framing. Recommendations are rule-based scoring, not measured benchmarks.
Configure your build
Tell us about your build
URL updates as you change fields. Showing the balanced default — change any field to refine.
USD ceiling. We treat over-budget cards as 'avoid' unless they'd otherwise have been a top pick — those land in 'Worth saving up for'.
Hard cut for vendor compatibility. NVIDIA + macOS = avoid (CUDA dropped in 10.13). Apple Silicon + Windows = impossible. AMD + macOS = abandoned.
Sets the VRAM target. 'Coding' wants 24GB minimum, 'chat' is fine at 8GB. Different workloads have different sweet-spot VRAM tiers.
Beginners get steered toward NVIDIA + Apple — the only paths where Ollama / LM Studio 'just work' with no driver wrangling.
Maximum catalog power for one GPU. Cards above this limit are excluded from recommendations. Allow additional PSU capacity for the CPU, other components and transient loads.
Thermal headroom proxy. High-power cards in small cases get loud first. Apple Silicon scores high for 'silent' because it actually is.
Workstation cards (RTX A6000, H100) have weak gaming drivers. Apple Silicon doesn't game. Consumer Ada/Blackwell does both well.
Crude perf-per-watt proxy via bandwidth-per-watt. Apple Silicon dominates 'strict' because the M-series is the perf-per-watt king for inference.
Display currency only. USD is the source-of-truth. FX rates are mid-market, dated, and approximate — actual buying rates may carry a 3-5% spread.
Price versus performance
Top GPU picks for your build
NVIDIA GeForce RTX 3090
Top pick for your setup. With your $1,500 budget on Linux for coding agents, the NVIDIA GeForce RTX 3090 ranks here because 24 GB hits the workable band for coding agents — fits at sensible quants without becoming the bottleneck.
- No benchmarks on file for this hardware.
- Dual RTX 3090 workstation stack — 70B-class on $1,800 of used GPUs — Workstation · GPUs (2× 24GB used, the cheapest path to 48 GB total)
- Quad RTX 3090 workstation stack — the prosumer 100B-class ceiling — Homelab · GPUs (4× 24GB used; the prosumer-ceiling stack)
How we scored this card▸
Each dimension is a 0-100 score. The card's position in the ranking is the weighted sum — but we surface tiers, not raw numbers. Bars are sorted by weight (most-influential first).
- VRAM × workloadweight 22%70Good
- Budget fitweight 18%95Excellent
- OS compatibilityweight 16%100Excellent
- Skill matchweight 10%95Excellent
- Power headroomweight 8%80Strong
- Multi-GPU pathweight 8%80Strong
- Thermal / noiseweight 6%95Excellent
- Gaming alignmentweight 6%95Excellent
- Perf-per-wattweight 6%85Strong
Tier mapping: top ≥ 75 composite · alternate 60-74 · acceptable 40-59 · avoid < 40 or over-budget / incompatible.
- •Used-market only — fan/thermal-pad inspection required; new MSRP from launch is no longer the relevant price.
- Dual RTX 3090 workstation stack — 70B-class on $1,800 of used GPUs — Workstation tier · GPUs (2× 24GB used, the cheapest path to 48 GB total)
- Quad RTX 3090 workstation stack — the prosumer 100B-class ceiling — Homelab tier · GPUs (4× 24GB used; the prosumer-ceiling stack)
NVIDIA RTX PRO 4000 Blackwell
Top pick for your setup. With your $1,500 budget on Linux for coding agents, the NVIDIA RTX PRO 4000 Blackwell ranks here because 24 GB hits the workable band for coding agents — fits at sensible quants without becoming the bottleneck.
- No benchmarks on file for this hardware.
How we scored this card▸
Each dimension is a 0-100 score. The card's position in the ranking is the weighted sum — but we surface tiers, not raw numbers. Bars are sorted by weight (most-influential first).
- VRAM × workloadweight 22%70Good
- Budget fitweight 18%85Strong
- OS compatibilityweight 16%100Excellent
- Skill matchweight 10%95Excellent
- Power headroomweight 8%95Excellent
- Multi-GPU pathweight 8%80Strong
- Thermal / noiseweight 6%95Excellent
- Gaming alignmentweight 6%85Strong
- Perf-per-wattweight 6%95Excellent
Tier mapping: top ≥ 75 composite · alternate 60-74 · acceptable 40-59 · avoid < 40 or over-budget / incompatible.
NVIDIA GeForce RTX 5090 Mobile
Top pick for your setup. With your $1,500 budget on Linux for coding agents, the NVIDIA GeForce RTX 5090 Mobile ranks here because 24 GB hits the workable band for coding agents — fits at sensible quants without becoming the bottleneck.
- No benchmarks on file for this hardware.
How we scored this card▸
Each dimension is a 0-100 score. The card's position in the ranking is the weighted sum — but we surface tiers, not raw numbers. Bars are sorted by weight (most-influential first).
- VRAM × workloadweight 22%70Good
- Budget fitweight 18%40Acceptable
- OS compatibilityweight 16%100Excellent
- Skill matchweight 10%95Excellent
- Power headroomweight 8%95Excellent
- Multi-GPU pathweight 8%80Strong
- Thermal / noiseweight 6%95Excellent
- Gaming alignmentweight 6%95Excellent
- Perf-per-wattweight 6%95Excellent
Tier mapping: top ≥ 75 composite · alternate 60-74 · acceptable 40-59 · avoid < 40 or over-budget / incompatible.
NVIDIA GeForce RTX 4080
Top pick for your setup. With your $1,500 budget on Linux for coding agents, the NVIDIA GeForce RTX 4080 sits in this tier on a balance of capability, OS compat, power, and budget fit.
- No benchmarks on file for this hardware.
How we scored this card▸
Each dimension is a 0-100 score. The card's position in the ranking is the weighted sum — but we surface tiers, not raw numbers. Bars are sorted by weight (most-influential first).
- VRAM × workloadweight 22%33Weak
- Budget fitweight 18%95Excellent
- OS compatibilityweight 16%100Excellent
- Skill matchweight 10%95Excellent
- Power headroomweight 8%80Strong
- Multi-GPU pathweight 8%80Strong
- Thermal / noiseweight 6%95Excellent
- Gaming alignmentweight 6%90Excellent
- Perf-per-wattweight 6%85Strong
Tier mapping: top ≥ 75 composite · alternate 60-74 · acceptable 40-59 · avoid < 40 or over-budget / incompatible.
- •16 GB is below the comfortable VRAM minimum for coding agents — expect quant downgrades or very tight context windows.
NVIDIA GeForce RTX 4080 Super
Top pick for your setup. With your $1,500 budget on Linux for coding agents, the NVIDIA GeForce RTX 4080 Super sits in this tier on a balance of capability, OS compat, power, and budget fit.
- No benchmarks on file for this hardware.
How we scored this card▸
Each dimension is a 0-100 score. The card's position in the ranking is the weighted sum — but we surface tiers, not raw numbers. Bars are sorted by weight (most-influential first).
- VRAM × workloadweight 22%33Weak
- Budget fitweight 18%95Excellent
- OS compatibilityweight 16%100Excellent
- Skill matchweight 10%95Excellent
- Power headroomweight 8%80Strong
- Multi-GPU pathweight 8%80Strong
- Thermal / noiseweight 6%95Excellent
- Gaming alignmentweight 6%90Excellent
- Perf-per-wattweight 6%85Strong
Tier mapping: top ≥ 75 composite · alternate 60-74 · acceptable 40-59 · avoid < 40 or over-budget / incompatible.
- •16 GB is below the comfortable VRAM minimum for coding agents — expect quant downgrades or very tight context windows.
NVIDIA GeForce RTX 5070 Ti
Top pick for your setup. With your $1,500 budget on Linux for coding agents, the NVIDIA GeForce RTX 5070 Ti sits in this tier on a balance of capability, OS compat, power, and budget fit.
- No benchmarks on file for this hardware.
How we scored this card▸
Each dimension is a 0-100 score. The card's position in the ranking is the weighted sum — but we surface tiers, not raw numbers. Bars are sorted by weight (most-influential first).
- VRAM × workloadweight 22%33Weak
- Budget fitweight 18%95Excellent
- OS compatibilityweight 16%100Excellent
- Skill matchweight 10%95Excellent
- Power headroomweight 8%80Strong
- Multi-GPU pathweight 8%80Strong
- Thermal / noiseweight 6%95Excellent
- Gaming alignmentweight 6%90Excellent
- Perf-per-wattweight 6%85Strong
Tier mapping: top ≥ 75 composite · alternate 60-74 · acceptable 40-59 · avoid < 40 or over-budget / incompatible.
- •16 GB is below the comfortable VRAM minimum for coding agents — expect quant downgrades or very tight context windows.
NVIDIA GeForce RTX 4070 Ti Super
Top pick for your setup. With your $1,500 budget on Linux for coding agents, the NVIDIA GeForce RTX 4070 Ti Super sits in this tier on a balance of capability, OS compat, power, and budget fit.
- No benchmarks on file for this hardware.
How we scored this card▸
Each dimension is a 0-100 score. The card's position in the ranking is the weighted sum — but we surface tiers, not raw numbers. Bars are sorted by weight (most-influential first).
- VRAM × workloadweight 22%33Weak
- Budget fitweight 18%95Excellent
- OS compatibilityweight 16%100Excellent
- Skill matchweight 10%95Excellent
- Power headroomweight 8%80Strong
- Multi-GPU pathweight 8%80Strong
- Thermal / noiseweight 6%95Excellent
- Gaming alignmentweight 6%90Excellent
- Perf-per-wattweight 6%85Strong
Tier mapping: top ≥ 75 composite · alternate 60-74 · acceptable 40-59 · avoid < 40 or over-budget / incompatible.
- •16 GB is below the comfortable VRAM minimum for coding agents — expect quant downgrades or very tight context windows.
NVIDIA GeForce RTX 5060 Ti 16GB
Top pick for your setup. With your $1,500 budget on Linux for coding agents, the NVIDIA GeForce RTX 5060 Ti 16GB sits in this tier on a balance of capability, OS compat, power, and budget fit.
- No benchmarks on file for this hardware.
How we scored this card▸
Each dimension is a 0-100 score. The card's position in the ranking is the weighted sum — but we surface tiers, not raw numbers. Bars are sorted by weight (most-influential first).
- VRAM × workloadweight 22%33Weak
- Budget fitweight 18%80Strong
- OS compatibilityweight 16%100Excellent
- Skill matchweight 10%95Excellent
- Power headroomweight 8%95Excellent
- Multi-GPU pathweight 8%80Strong
- Thermal / noiseweight 6%95Excellent
- Gaming alignmentweight 6%70Good
- Perf-per-wattweight 6%95Excellent
Tier mapping: top ≥ 75 composite · alternate 60-74 · acceptable 40-59 · avoid < 40 or over-budget / incompatible.
- •16 GB is below the comfortable VRAM minimum for coding agents — expect quant downgrades or very tight context windows.
Intel Arc Pro B60 24GB
Top pick for your setup. With your $1,500 budget on Linux for coding agents, the Intel Arc Pro B60 24GB sits in this tier on a balance of capability, OS compat, power, and budget fit.
- No benchmarks on file for this hardware.
How we scored this card▸
Each dimension is a 0-100 score. The card's position in the ranking is the weighted sum — but we surface tiers, not raw numbers. Bars are sorted by weight (most-influential first).
- VRAM × workloadweight 22%65Good
- Budget fitweight 18%80Strong
- OS compatibilityweight 16%70Good
- Skill matchweight 10%55Acceptable
- Power headroomweight 8%95Excellent
- Multi-GPU pathweight 8%80Strong
- Thermal / noiseweight 6%95Excellent
- Gaming alignmentweight 6%85Strong
- Perf-per-wattweight 6%95Excellent
Tier mapping: top ≥ 75 composite · alternate 60-74 · acceptable 40-59 · avoid < 40 or over-budget / incompatible.
- •Intel discrete GPU AI tooling lags NVIDIA — runtime support exists (IPEX-LLM) but documentation is thinner.
GPUs we ruled out
- NVIDIA GeForce RTX 3090 Ti — The catalog GPU power of 450 W exceeds your 350 W GPU limit.
- NVIDIA RTX PRO 4500 Blackwell — Out of budget for this query.~$2,600
- NVIDIA H100 PCIe — Out of budget for this query.~$25,000
Where to go from here
Multi-store, multi-region prices for every card here. US/EU/UK/CA/AU — see what these cards actually cost in your region before you buy.
One step further: this card + runtime + 1-3 models + cost rollup + ready-to-paste install script. Eight inputs → full rig.
Once you’ve picked a card, model the full build (CPU, RAM, runtime) for which models fit comfortably.
The long-form essay version: VRAM tiers, MoE math, NVLink truth, used-market price discipline.
Curated multi-GPU and Apple-cluster setups with effective-VRAM math you can trust.
How the trust layer behind these recommendations actually works — every dimension, every formula, the honest limits.
Where the intelligence graph has signal vs which model × hardware × quant cohorts are still underpowered.
Help tip a cohort across the 5-row threshold for outlier detection — the most operator-impactful contribution.