- Home
- Skills
- APIs & Backend
- ai chip arch toolkit
ai chip arch toolkit
Quantitative AI chip architecture budgeting for compute, bandwidth, cost, and power optimization.
$10
Works with the AI Agents you already use
ai chip arch toolkit
Example session with this skill installed
Calculate the inference budget for a 70B parameter model at INT4 precision, 8K sequence length, batch 32, targeting 200ms TTFT and 30ms TPOT. Determine the bottleneck.
- Read your context and instructions
- Compiled the ai chip arch
I have calculated the inference budget for the 70B INT4 workload. The analysis shows that the design is currently memory-bandwidth constrained during the decoding phase.
AI Chip Architecture Sizing Report
| Metric | Value | Unit | Notes |
|---|---|---|---|
| Model Weight Capacity | 35.0 | GB | 70B parameters @ 4-bit precision |
| KV Cache (8K Seq, B32) | 16.4 | GB | Based on 80 layers, 8 KV heads, 128 dim |
| Total Memory Required | 51.4 | GB | Minimum SRAM/HBM footprint |
| Required Bandwidth (TPOT) | 1,166 | GB/s | To meet 30ms TPOT target (100% MBU) |
| Effective TOPS (Prefill) | 1,400 | TOPS | To meet 200ms TTFT (100% MFU) |
| Arithmetic Intensity | 12.8 | ops/byte | Decoding phase intensity |
Bottleneck Analysis
Primary Constraint: Memory Bandwidth. To achieve 30ms TPOT, the system requires ~1.5 TB/s (assuming 80% MBU). Standard LPDDR5 will not suffice; HBM3 is recommended.
Compute Status: The 1,400 TOPS requirement for TTFT is manageable for high-end accelerators, but check SRAM hit rates to avoid falling back to bandwidth limits during prefill.
Next steps
- Run
python3 scripts/chip_sizing.py roofline --tops 1500 --bw-gbs 1500to visualize the balance point. - Review
references/architecture-patterns.mdto evaluate weight-stationary vs. output-stationary dataflows for this intensity. - Populate the
PPA_Table_Templateinreferences/deliverable-templates.mdwith these values.
Connects securely to your tools. The creator never sees your data.
About this skill
Shift AI architecture discussions from intuition to rigorous budgeting. This skill provides the mathematical framework and tooling required to size hardware against specific model workloads. It eliminates guesswork by enforcing hard constraints across compute, bandwidth, and memory capacity.
What it does
- Workload Sizing calculates precise requirements for TOPS, memory bandwidth, and KV cache capacity based on specific model parameters and latency targets.
- Roofline Analysis determines if a design is compute-bound or memory-bound by mapping arithmetic intensity against machine balance points.
- Cost Modeling evaluates chiplet yield gains versus packaging overhead to inform partitioning strategies.
- Energy Budgeting estimates power consumption by tracking energy-per-bit for HBM transfers, SRAM access, and NoC movement.
- Architecture Benchmarking provides quantitative comparison matrices for different dataflow paradigms and hardware patterns.
How it works
- Initialize Requirements by defining the target model, precision, sequence length, and concurrency goals.
- Execute Calculators using the
chip_sizing.pyscripts to generate numerical tables for inference or training budgets. - Compare Patterns against the architecture reference library to select optimal dataflow and storage hierarchies.
- Generate Deliverables by populating spec sheets, PPA tables, and risk checklists using provided templates.
Frameworks & tools
Python-based sizing scripts, Markdown-based architecture templates, and reference math for HBM3, SRAM, and advanced packaging yield models.
Why this beats prompting it yourself
Generic LLMs often hallucinate memory math or confuse bits and bytes, leading to unbuildable hardware specs. This skill uses hardcoded formulas and scripts to ensure every TOPS and GB/s calculation is physically grounded and includes realistic Utilization (MFU/MBU) factors.
Use cases
- Spec Definition for sizing a new NPU to handle Llama 3 70B within a specific power envelope.
- Design Review to verify if a vendor's claimed performance survives bandwidth reality.
- Financial Analysis to decide between monolithic dies or chiplet-based designs based on wafer costs.
Known limitations
Does not provide real-time vendor pricing or proprietary foundry IP data. Results are order-of-magnitude estimates requiring validation against specific process design kits (PDKs).
Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for APIs & Backend.
Requires validation against specific PDKs. Does not include proprietary IP pricing. Numerical outputs are estimates based on user-provided utilization assumptions.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before buying a skill.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 13 days ago