Pinned Tweet
Common Foundry: We unlocked nearly 2× FP8→FP32 Tensor Core throughput on an RTX 4070 Ti SUPER. A matched GEMM in a custom NVIDIA CUTLASS 4.8.1 harness ran 1.95× faster: 158.54 → 309.08 TFLOP/s. Same GPU, binary and 240 W cap. Results:
github.com/JustAResearche…
github.com/JustAResearche…
15
12
143
19,452
Weekend setup check: did you get a Common Foundry wallet, node or miner running?
Share one thing that worked and one thing that was confusing. GPU model and OS help others compare notes.
New here? Setup help is available in Discord:
discord.gg/XGuutqWMWP
Share one thing that worked and one thing that was confusing. GPU model and OS help others compare notes.
New here? Setup help is available in Discord:
discord.gg/XGuutqWMWP
0
0
0
131
After the 30 series was unlocked, then we unlocked the 40 series and the 20 series coming right after that.
Naturally everyone turns their attention to the 50 series.
Who will be the first to unlock its FP8 potential?
I can tell you, its not going to unlock the way 20-40 did
Naturally everyone turns their attention to the 50 series.
Who will be the first to unlock its FP8 potential?
I can tell you, its not going to unlock the way 20-40 did
2
0
9
411
A useful GPU benchmark lets someone else repeat it.
For CMFD, share your setup, test length, accepted work and power measurement alongside speed. Include the settings that changed.
A screenshot is a starting point. Reproducible results help the next operator.
For CMFD, share your setup, test length, accepted work and power measurement alongside speed. Include the settings that changed.
A screenshot is a starting point. Reproducible results help the next operator.
0
1
4
470
Let's go!!! 20 series included too.
Nice work
Nice work
Now I also confirmed that 2080ti unlock is successful with similar approach as the 30-series' method
it reach nearly same speed as 350w 3090(locked) with only 250w in mmapeak and mamf-finder
github.com/KohakuBlueleaf…
1
3
22
1,026
GPU operators and AI builders: what would you need from an inference marketplace before trying it?
Model support? Predictable latency? Clear pricing? Reliable availability?
We're building toward customer-paid inference. Which requirement would decide it for you?
Model support? Predictable latency? Clear pricing? Reliable availability?
We're building toward customer-paid inference. Which requirement would decide it for you?
1
1
8
749
Common Foundry retweeted
Now 🤯 an insane JAILBREAK for RTX 4000 series,
NVIDIA really tightened things down with blackwell :(
Git Repo link : github.com/JustAResearche…
NVIDIA really tightened things down with blackwell :(
Git Repo link : github.com/JustAResearche…
RTX 3000 series can run BF16 almost 2x faster
NVIDIA 🤯 locked it via kernel not hardware limitation
detailed post:
reddit.com/r/StableDiffus…
50
80
1,446
139,876
Common Foundry retweeted
Meet the hero @CommonFoundry1
His work just unlocked 2x hidden Tensor Core performance on Ada GPUs (158 -> 309 TFLOPS on CUTLASS GEMM).
I dont have 40-series cards, but I tested the exact same boost on my RTX 3090s earlier this week. The hardware jump is completely real.
Follow him!
His work just unlocked 2x hidden Tensor Core performance on Ada GPUs (158 -> 309 TFLOPS on CUTLASS GEMM).
I dont have 40-series cards, but I tested the exact same boost on my RTX 3090s earlier this week. The hardware jump is completely real.
Follow him!
Common Foundry: We unlocked nearly 2× FP8→FP32 Tensor Core throughput on an RTX 4070 Ti SUPER. A matched GEMM in a custom NVIDIA CUTLASS 4.8.1 harness ran 1.95× faster: 158.54 → 309.08 TFLOP/s. Same GPU, binary and 240 W cap. Results:
github.com/JustAResearche…
13
13
271
15,881
Common Foundry retweeted
Desktop RTX 4060 Ti 16GB tensor unlock, confirmed on TWO different boards 🎉
MSI (10de:2803) + ASUS ProArt (10de:2805) — same chip, different board variants, both unlocked with the same kernel module.
GEMM 8192³, locked vs unlocked @165W:
TF32: 23.4 → 34.7 TF/s (+48%)
74.7 (+56%)
FP16/FP32-acc: 46.4 → 71.2 (+53%)
FP8 e4m3: 93 → 126 (+36%)
INT8 & FP32 CUDA: unchanged (control ✓)
No vBIOS flash, no driver patch. A kernel module runs before the NVIDIA driver binds and opens the FP32-accumulate tensor paths. Fully reversible.
@CommonFoundry1 your AD106M payload extended to desktop AD106 (2803/2805) just works — thanks! Also kudos to Zagitta for gpu-unlock.
RTX 3090 (GA102) on the same box: BF16 78→129 TF/s, FP16 78→125 @420W.
MSI (10de:2803) + ASUS ProArt (10de:2805) — same chip, different board variants, both unlocked with the same kernel module.
GEMM 8192³, locked vs unlocked @165W:
TF32: 23.4 → 34.7 TF/s (+48%)
74.7 (+56%)
FP16/FP32-acc: 46.4 → 71.2 (+53%)
FP8 e4m3: 93 → 126 (+36%)
INT8 & FP32 CUDA: unchanged (control ✓)
No vBIOS flash, no driver patch. A kernel module runs before the NVIDIA driver binds and opens the FP32-accumulate tensor paths. Fully reversible.
@CommonFoundry1 your AD106M payload extended to desktop AD106 (2803/2805) just works — thanks! Also kudos to Zagitta for gpu-unlock.
RTX 3090 (GA102) on the same box: BF16 78→129 TF/s, FP16 78→125 @420W.
1
1
4
476
Common Foundry retweeted
4000 series unlock is here! Get 2x speed increase in tensor core!
Common Foundry: We unlocked nearly 2× FP8→FP32 Tensor Core throughput on an RTX 4070 Ti SUPER. A matched GEMM in a custom NVIDIA CUTLASS 4.8.1 harness ran 1.95× faster: 158.54 → 309.08 TFLOP/s. Same GPU, binary and 240 W cap. Results:
github.com/JustAResearche…
0
1
1
77
Common Foundry retweeted
Common Foundry: We unlocked nearly 2× FP8→FP32 Tensor Core throughput on an RTX 4070 Ti SUPER. A matched GEMM in a custom NVIDIA CUTLASS 4.8.1 harness ran 1.95× faster: 158.54 → 309.08 TFLOP/s. Same GPU, binary and 240 W cap. Results:
github.com/JustAResearche…
1
1
3
508
NVIDIA GeForce RTX 30 GPUs get up to 2X faster AI compute performance with new GPU unlock mod, real-world gains of up to 60%. 🔗wccf.tech/1lfle
1
0
10
917
Common Foundry retweeted
🔥 中午刚说 40 系还没支持，今晚 4090 到 4060Ti 全解锁了，连 FP8 都翻倍
40 系不能直接套 30 系的模块，是因为它换了另一套签名引导固件，多锁了四道权限门。今天研究者 TOTORONG 放出 AD10x 载荷，在驱动绑定前一次捅开四道门，而整个 40 系共用同一份引导固件，所以从 4090 到 4060 Ti 一下全通了——同样不刷 vBIOS、不改驱动。
1️⃣ RTX 4090：FP16/BF16 指令吞吐从 165.6 涨到 331.6 TFLOPS；500W 下跑 cuBLAS 8192³ 矩阵乘，TF32 从 77.5 涨到 165.9 TFLOPS（2.14 倍），3 张非公卡已复现
2️⃣ RTX 4080 SUPER：PyTorch 矩阵乘 FP16 从 110.6 涨到 202.7 TFLOPS（1.83 倍）；mmapeak 实测连 40 系独有的 FP8（FP32 累加）都从 201.8 涨到 376.1 TFLOPS
3️⃣ RTX 4070 Ti SUPER：240W 下 FP16/BF16 从 95 涨到 189.9 TFLOPS；跑 CUTLASS 4.8.1 原生 FP8 快累加矩阵乘从 158.5 涨到 309.1 TFLOPS（1.95 倍），322 万个输出单元零比特差异
4️⃣ RTX 4060 Ti 16GB 与 Win11：三卡同机 A/B 实测，182W 功耗墙下 BF16/FP16 仍涨 38–39%；Windows 11 也出了免核显、单卡直接亮机的 UEFI 引导版
💡 Linux 端装下方仓库（4080S 看原仓 PR #24）。40 系别用 targets=all，先跑 python3 tools/print_boot_option.py 拿到单卡参数加进内核启动项，关掉 Secure Boot，关机拔电源等几秒「冷启动」才生效；Windows 端去 Hugging Face 搜 TOTORONG/AD103_win_unlocker。
github.com/JustAResearche…
中午还在等 40 系的各位，今晚敢拔电源试一把吗？
40 系不能直接套 30 系的模块，是因为它换了另一套签名引导固件，多锁了四道权限门。今天研究者 TOTORONG 放出 AD10x 载荷，在驱动绑定前一次捅开四道门，而整个 40 系共用同一份引导固件，所以从 4090 到 4060 Ti 一下全通了——同样不刷 vBIOS、不改驱动。
1️⃣ RTX 4090：FP16/BF16 指令吞吐从 165.6 涨到 331.6 TFLOPS；500W 下跑 cuBLAS 8192³ 矩阵乘，TF32 从 77.5 涨到 165.9 TFLOPS（2.14 倍），3 张非公卡已复现
2️⃣ RTX 4080 SUPER：PyTorch 矩阵乘 FP16 从 110.6 涨到 202.7 TFLOPS（1.83 倍）；mmapeak 实测连 40 系独有的 FP8（FP32 累加）都从 201.8 涨到 376.1 TFLOPS
3️⃣ RTX 4070 Ti SUPER：240W 下 FP16/BF16 从 95 涨到 189.9 TFLOPS；跑 CUTLASS 4.8.1 原生 FP8 快累加矩阵乘从 158.5 涨到 309.1 TFLOPS（1.95 倍），322 万个输出单元零比特差异
4️⃣ RTX 4060 Ti 16GB 与 Win11：三卡同机 A/B 实测，182W 功耗墙下 BF16/FP16 仍涨 38–39%；Windows 11 也出了免核显、单卡直接亮机的 UEFI 引导版
💡 Linux 端装下方仓库（4080S 看原仓 PR #24）。40 系别用 targets=all，先跑 python3 tools/print_boot_option.py 拿到单卡参数加进内核启动项，关掉 Secure Boot，关机拔电源等几秒「冷启动」才生效；Windows 端去 Hugging Face 搜 TOTORONG/AD103_win_unlocker。
github.com/JustAResearche…
中午还在等 40 系的各位，今晚敢拔电源试一把吗？
162
330
2,770
296,526
Hey we made that! 😄
Have fun with it.
Have fun with it.
🔥 中午刚说 40 系还没支持，今晚 4090 到 4060Ti 全解锁了，连 FP8 都翻倍
40 系不能直接套 30 系的模块，是因为它换了另一套签名引导固件，多锁了四道权限门。今天研究者 TOTORONG 放出 AD10x 载荷，在驱动绑定前一次捅开四道门，而整个 40 系共用同一份引导固件，所以从 4090 到 4060 Ti 一下全通了——同样不刷 vBIOS、不改驱动。
1️⃣ RTX 4090：FP16/BF16 指令吞吐从 165.6 涨到 331.6 TFLOPS；500W 下跑 cuBLAS 8192³ 矩阵乘，TF32 从 77.5 涨到 165.9 TFLOPS（2.14 倍），3 张非公卡已复现
2️⃣ RTX 4080 SUPER：PyTorch 矩阵乘 FP16 从 110.6 涨到 202.7 TFLOPS（1.83 倍）；mmapeak 实测连 40 系独有的 FP8（FP32 累加）都从 201.8 涨到 376.1 TFLOPS
3️⃣ RTX 4070 Ti SUPER：240W 下 FP16/BF16 从 95 涨到 189.9 TFLOPS；跑 CUTLASS 4.8.1 原生 FP8 快累加矩阵乘从 158.5 涨到 309.1 TFLOPS（1.95 倍），322 万个输出单元零比特差异
4️⃣ RTX 4060 Ti 16GB 与 Win11：三卡同机 A/B 实测，182W 功耗墙下 BF16/FP16 仍涨 38–39%；Windows 11 也出了免核显、单卡直接亮机的 UEFI 引导版
💡 Linux 端装下方仓库（4080S 看原仓 PR #24）。40 系别用 targets=all，先跑 python3 tools/print_boot_option.py 拿到单卡参数加进内核启动项，关掉 Secure Boot，关机拔电源等几秒「冷启动」才生效；Windows 端去 Hugging Face 搜 TOTORONG/AD103_win_unlocker。
github.com/JustAResearche…
中午还在等 40 系的各位，今晚敢拔电源试一把吗？
0
1
7
685
Common Foundry today: a live proof-of-work chain, GPU mining, wallets and a public explorer.
The next direction: customer-paid AI inference, still in development.
Mining and customer inference are separate jobs. Here's the project overview:
commonfoundry.ai
The next direction: customer-paid AI inference, still in development.
Mining and customer inference are separate jobs. Here's the project overview:
commonfoundry.ai
0
1
6
399