🔄 Авто-синхронизация: из Discussion #870 каждые 6 часов.
Gonka AI Testnet¶
Автор: @Alert17 · Категория: 
Proposals · Создано: 2026-03-06 17:14 UTC · Обновлено: 2026-05-01 15:45 UTC
📝 Описание¶
Gonka Testnet - Permanent Parallel Network for Consumer GPU Participation¶
Motivation¶
Gonka mainnet requires 320 GB VRAM per ML Node (e.g. 4×H100 80 GB), limiting participation to datacenter-grade multi-GPU setups. Consumer and prosumer GPUs - RTX 3090, RTX 4090, A6000 - are excluded entirely despite being capable compute devices.
This creates three concrete problems:
- Lost participants. Hosts who invested in hardware and contributed to the network were locked out as VRAM requirements increased with PoC V2 and the transition to Qwen3-235B.
- No onboarding path. New users cannot try operating a Gonka node without committing $100K+ in datacenter hardware. There is no way to learn the protocol, test configurations, or evaluate the economics before going all-in.
- No staging environment. Protocol upgrades, model changes, and parameter adjustments are deployed directly to mainnet with no parallel network for validation.
High-Level Solution¶
A permanent parallel Cosmos chain (gonka-testnet-1) running the exact same Gonka protocol with a lighter inference model, opening participation to GPUs with 24 GB+ VRAM.
Design Principle: Identical Protocol, Different Scale¶
The testnet runs the same blockchain as mainnet. No protocol modifications, no alternative reward systems, no different collateral mechanics. The chain code is a direct fork of the mainnet codebase - any upgrade tested on testnet deploys to mainnet via Cosmovisor without modification.
| Parameter | Mainnet | Testnet |
|---|---|---|
| chain-id | gonka-mainnet |
gonka-testnet-1 |
| Token denom | gnk |
tgnk |
| Total supply | 1,000,000,000 GNK | 1,000,000,000 tGNK |
| Miner allocation | 800,000,000 GNK (80%) | 800,000,000 tGNK (80%) |
| Founder allocation | 200,000,000 GNK (20%) | 200,000,000 tGNK (20%) |
| Min VRAM per MLNode | 320 GB (multi-GPU) | 24 GB (single GPU) |
| Inference model | Qwen3-235B-FP8 | Qwen2.5-14B-Instruct (recommended) |
| Emission formula | 323,000 × e^(-0.000475 × epoch) |
Same formula, same decay |
| Collateral | Required (after 180-epoch grace) | Same mechanism |
| PoC V2 | 1-10% random verification | Same mechanism |
| Sprint PoW | Continuous nonce generation | Same mechanism |
| Governance (x/gov) | Active | Same mechanism |
Hardware Requirements¶
Minimum VRAM: 24 GB. This covers RTX 3090 and RTX 4090 as entry-level, while excluding cards that cannot run the target model class (7B-14B parameters).
| GPU | VRAM | Category | Expected Performance |
|---|---|---|---|
| RTX 3090 | 24 GB | Minimum | Runs model with limited headroom; lower nonce rate |
| RTX 4090 | 24 GB | Minimum | Higher compute; faster nonces |
| A5000 | 24 GB | Prosumer | Datacenter-class reliability |
| A6000 / L40 | 48 GB | Optimal | Comfortable VRAM margin; high nonce rate |
| A100 (40 GB) | 40 GB | Optimal | Datacenter performance on lighter model |
| 2×RTX 4090 | 48 GB | Optimal (multi-GPU) | Combined VRAM; tensor parallelism possible |
More powerful GPUs naturally generate more nonces per epoch due to higher throughput - no artificial weight multipliers. Same Sprint mechanism as mainnet.
Model Selection¶
The testnet model must be: (1) small enough for 24 GB VRAM, and (2) large enough that users cannot trivially run it without the Gonka infrastructure.
| Model | Params | FP16 VRAM | FP8 VRAM | 24 GB Fit? | 48 GB Fit? |
|---|---|---|---|---|---|
| Qwen2.5-7B-Instruct | 7.6B | ~15 GB | ~8 GB | ✅ FP16 | ✅ |
| Qwen2.5-14B-Instruct | 14.8B | ~30 GB | ~15 GB | ⚠️ FP8 only | ✅ FP16 |
| Llama-3.1-8B-Instruct | 8.0B | ~16 GB | ~8 GB | ✅ FP16 | ✅ |
| Mistral-Small-24B | 24B | ~48 GB | ~24 GB | ⚠️ FP8 tight | ✅ FP16 |
| Qwen2.5-32B-Instruct | 32.5B | ~65 GB | ~33 GB | ❌ | ⚠️ FP8 tight |
Recommendation: Qwen2.5-14B-Instruct. Fits on 24 GB only with FP8 quantization (tight, not trivial), runs comfortably on 48 GB in FP16. Well-supported by vLLM, strong benchmarks. Alternative: Qwen2.5-7B as safer launch option, upgradeable to 14B via Cosmovisor.
Economic Model¶
Emission¶
Identical to mainnet: R(epoch) = 323,000 × e^(-0.000475 × epoch)
At testnet genesis (epoch 0): 323,000 tGNK/epoch. Halving every ~1,459 epochs (~4.16 years). Total emission converges to ~680,000,000 tGNK.
| Milestone | tGNK Mined | Epoch | Time | Epoch Emission |
|---|---|---|---|---|
| 25% mined | 170,000,000 | ~606 | ~1.7 years | ~242,000 tGNK |
| 50% mined | 340,000,000 | ~1,459 | ~4.2 years | ~161,500 tGNK |
| 75% mined | 510,000,000 | ~2,918 | ~8.3 years | ~80,750 tGNK |
| 90% mined | 612,000,000 | ~4,847 | ~13.8 years | ~32,300 tGNK |
| 99% mined | 673,000,000 | ~9,694 | ~27.7 years | ~3,230 tGNK |
tGNK → GNK Conversion¶
This is the core economic link between testnet and mainnet.
A pool of 1,000,000 GNK (proposed) is allocated from the Community Pool (120M GNK, ~0.83%) via governance. The conversion uses a fixed rate derived from total emission cap:
Expanded: GNK = (Your tGNK / 680,000,000) × 1,000,000
How this works:
- Your tGNK - the tGNK you burn for conversion (earned by mining)
- 680,000,000 - total tGNK that will ever be mined via emission (fixed constant from decay formula, does not change over time)
- 1,000,000 GNK - mainnet pool backing all conversions
- Rate: 0.00147 GNK per tGNK - fixed, same for early and late miners. No rush incentive, no timing games
Worked Examples¶
Since the rate is fixed, the only variable is how much tGNK a miner earns (depends on network size and time):
Scenario A: 100 GPUs, 3 months (~90 epochs). RTX 4090 earning ~1% share → ~285,000 tGNK mined.
Scenario B: 500 GPUs, 6 months (~180 epochs). RTX 4090 earning ~0.2% share → ~111,600 tGNK.
Scenario C: 1,000 GPUs, 12 months (~350 epochs). RTX 4090 earning ~0.1% share → ~104,000 tGNK.
Monthly Income Estimates (500 GPU network)¶
| GPU | VRAM | tGNK (6 months) | GNK | USD (~$0.84) | Per Month |
|---|---|---|---|---|---|
| RTX 3090 | 24 GB | ~67,000 | ~98 | ~$82 | ~$14/mo |
| RTX 4090 | 24 GB | ~112,000 | ~164 | ~$138 | ~$23/mo |
| A6000 | 48 GB | ~179,000 | ~263 | ~$221 | ~$37/mo |
| 2×4090 | 48 GB | ~223,000 | ~328 | ~$276 | ~$46/mo |
Electricity for RTX 4090 (~150W) is ~$15-20/mo. Income exceeds operating costs while remaining modest enough to avoid speculative farming.
Self-Balancing Properties¶
- More miners → smaller share per GPU → lower GNK income → rental unprofitable → only hardware owners remain
- Fewer miners → larger share → higher income → attracts new participants → network grows
- Fixed rate = no timing games. A tGNK mined in month 1 is worth the same as one mined in month 12
- Pool depletion is predictable: after 50% of tGNK is mined (~4.2 years), 50% of pool is consumed
Post-Emission: Transition to Work Coins¶
As emission decays, miners rely increasingly on Work Coins - direct payments from developers for inference:
| Phase | Reward Coins (Emission) | Work Coins (Inference) |
|---|---|---|
| Year 0-2 | ~90% of income | ~10% |
| Year 2-5 | ~50% | ~50% |
| Year 5+ | ~20% and declining | ~80% |
| Year 10+ | Negligible | ~100% |
Critical question at each phase: is developer demand growing fast enough to replace declining emission?
Pool Sustainability¶
When the pool depletes, the community decides:
- Replenish - new governance proposal for additional GNK from Community Pool
- Work Coins only - if inference marketplace is active, pool may not be needed
- Adjust terms - governance can modify rate, pool size, or limits
- Close conversion - long-term scenario (year 10+) if testnet operates independently
IBC Bridge (tGNK → GNK)¶
Both chains are Cosmos SDK, making IBC the natural choice. Trustless, validator-independent, cryptographic verification - no multi-sig oracles or trusted third parties.
Bridge flow:
- Miner initiates tGNK burn on testnet via IBC transfer to mainnet channel
- Testnet relayer submits proof to mainnet
- Mainnet verifies proof against testnet’s light client
- Mainnet conversion module calculates GNK:
tGNK_amount × (pool_size / 680,000,000) - GNK released from conversion pool to miner’s mainnet address
- Received tGNK burned
The mainnet requires a conversion module (or x/inference extension) to handle this logic.
Anti-Fraud Measures¶
Since testnet tokens have real mainnet value, anti-fraud is mandatory at launch:
| Measure | Description | Purpose |
|---|---|---|
| GNK Deposit on Mainnet | Lock 50-100 GNK (proposed) to be eligible for conversion | Raises Sybil cost; mainnet economic tie |
| Per-Host Cap | Max % of epoch rewards per node | Anti-whale; community distribution |
| PoC V2 Verification | Same 1-10% random verification as mainnet | Catches fake compute |
| Nonce Rate Verification | Rate must match claimed hardware | Detects hardware spoofing |
| Conversion Rate Limit | Max GNK per wallet per period (daily/weekly) | Prevents pool drain attacks |
Open Questions¶
These require community input and stakeholder alignment:
- Model selection - Qwen2.5-14B (recommended) vs Qwen2.5-7B (safer) vs other? Tradeoff between accessibility and differentiation from consumer setups
- Pool size - 1,000,000 GNK proposed (~0.83% of Community Pool). Right amount?
- GNK deposit amount - 50-100 GNK proposed for bridge access. What threshold filters Sybils without excluding small participants?
- Per-host cap - 1-5% of epoch reward per node. What balances decentralization vs operational overhead?
- Conversion rate limit - Daily or weekly cap per wallet?
- Post-emission strategy - When emission becomes negligible (year 5+), replenish pool or transition to inference-only economics?
Budget¶
Requested: 1,000,000 GNK from the Community Pool (one-time, for the conversion pool).
This does not fund development - it backs the tGNK → GNK conversion mechanism. Development is handled by the proposing team.
On-chain governance vote (Community Pool spend) will be submitted as a separate transaction once community feedback is incorporated.
References¶
- Gonka Tokenomics - emission formula, halving schedule, supply distribution. Source: docs/tokenomics.md
- Gonka Whitepaper - Sprint PoW, PoC V2, inference architecture. Source: gonka.ai/introduction
- Gonka Repository - Chain Node, ML Node, API Node, allowlist. Source: github.com/gonka-ai/gonka
- Network Statistics - ~11,000 GPUs, 448+ hosts, 2,200+ developers, epoch ~170. Source: network dashboard
- Model Specs - VRAM requirements. Source: Hugging Face model cards, vLLM docs
- Cosmos IBC - bridge protocol. Source: docs.cosmos.network, github.com/cosmos/ibc-go
💬 Комментарии (1)¶
Комментарий 1 — @akamitch¶
2026-05-01 15:45 UTC
This reads more like a second mainnet than a testnet. A testnet needs to run the same hardware and models as mainnet — otherwise it won't catch the real inference-layer bugs (FlashInfer, expert-parallel, FP8 MoE, multi-GPU memory pressure, etc.), which are exactly the class of issues that bite us in production.