←
QuBlock141 posts
QuBlock avatar

QuBlock

@0xQuBlock

A crypto miner from Vietnam

◷ Joined August 2025▧ 9 Photos and videos
QuBlock avatar
QuBlock@0xQuBlock·12h

🚀 Pearl FP8 Update (Nov 17) • PR #503: FP8 security audit • PR #508: Batch verification • PR #509: Avoid redundant certificate verification

QuBlock avatar
QuBlock@0xQuBlock·12h

🚀 Big Pearl FP8 milestone! PR #515 merged: FP8 testnet activation is now set at block 67,370, with v1.6.0-fp8-testnet1. PR #514 also merged, improving H100 proof verification. PR #311 now has 63 commits but remains open. Bullish! Next: valid FP8 blocks and mainnet activation

QuBlock avatar
QuBlock@0xQuBlock·Nov 16

$Pearl FP8 update 16-11-2026 🚀 PR #311 is now at 57 commits, up from 49. Development continues, but it’s still open. PR #488 (FP8 quantization & jackpot security) has merged into fp8-scheme. No mainnet activation date yet. Bullish on progress, waiting for production.

QuBlock avatar
QuBlock@0xQuBlock·Nov 12

The benefit depends on load. When memory fills up, compression reduces cache eviction and repeated calculations. At light load, FP8 was slightly faster.

QuBlock avatar
QuBlock@0xQuBlock·Nov 12

Smaller caches help conversations stay on the GPU. At 128 simultaneous sessions, Pearl reports: • p90 time to first token: 144s → 7s • Peak throughput: +24% Tests used GLM-5.3 on 8× NVIDIA B200 GPUs.

QuBlock avatar
QuBlock@0xQuBlock·Nov 12

For GLM-5.3 with 100,000 tokens of context: • FP8 baseline: 5.1 GB of cache • PearlQuant: 1.8 GB Around 65% less cache memory, with no statistically significant drop across four benchmarks.

QuBlock avatar
QuBlock@0xQuBlock·Nov 12

PearlQuant compresses groups of 8 numbers together instead of rounding each number separately. This reduces storage while keeping the reconstructed numbers close to the originals.

QuBlock avatar
QuBlock@0xQuBlock·Nov 12

An AI saves calculations about earlier text so it can reuse them when generating new words. This temporary memory is called the KV cache. Longer conversations = larger cache = less space for other users.

QuBlock avatar
QuBlock@0xQuBlock·Nov 12

Long AI conversations use growing amounts of GPU memory. Pearl’s new research, PearlQuant, compresses that conversation memory so more sessions can fit on the same hardware. Here’s the idea in simple English 🧵

Each inference session has its own KV cache, so a smaller cache means more concurrent users per GPU. Introducing PearlQuant, our new KV cache quantization technique: 65% less storage than FP8, better throughput and faster serving, same quality. GLM-5.3 on 8×B200, 128 concurrent sessions: ⚡ p90 TTFT: 144s → 7s 📈 24% higher peak throughput 🎯 Quality on par with FP8 pearlresearch.ai/research/bl…

QuBlock avatar
QuBlock@0xQuBlock·Nov 11

7/7 📈 My take: Cautiously bullish on engineering progress. PR #311 remains open, PIP-3 is still Draft, and no FP8 activation height has been confirmed. The real catalysts remain: • FP8 mainnet • Stable GPU support • Working inference marketplace • Real compute demand

QuBlock avatar
QuBlock@0xQuBlock·Nov 11

6/7 🖥️ What about RTX 50-series? The latest work includes B200-specific changes, but SM120 support remains unconfirmed for production FP8 mining. RTX 5070 Ti, 5080, and 5090 miners still need reliable compatibility and performance benchmarks.

QuBlock avatar
QuBlock@0xQuBlock·Nov 11

5/7 🛠️ PR #493 and #494 #493 adds digest validation for V4 verifier caches to prevent incorrect cache reuse. #494 proposes removing a redundant verifier work limit. Small changes, but important for reliability and simplifying the verification pipeline.

QuBlock avatar
QuBlock@0xQuBlock·Nov 11

4/7 ⚡ PR #491 — Faster V4 prover New optimizations target: • FFT operations • Witness generation • Merkle commitments • Blake3 hashing Lower proof overhead could improve mining efficiency. But benchmarks and production validation are still needed.

QuBlock avatar
QuBlock@0xQuBlock·Nov 11

3/7 🔐 Why does #488 matter? FP8 isn’t just about faster matmul. Miners and verifiers must agree on the exact computation. Even small rounding differences could invalidate proofs. Getting quantization and jackpot security right is critical before mainnet activation.