←
Matt Virus987 posts
Matt Virus avatar

Matt Virus

@themattvirus

above-average husband/father, engineer @ Armis. Infosec,compute, IoT, farming in my blood. Loud/angry rock music FTW. founder @hackspaceio Yes it's my real name

◷ Joined June 2012▧ 174 Photos and videos
Matt Virus avatar
Matt Virus@themattvirus·Nov 18

Awesome!

Before you chase the 80GB CMP 170HX mod, check whether your card even has the memory for it. There's a free read-only tool for it called gpcprobe. It reads the GA100 fuses with the Nvidia driver still loaded and never writes anything. The screenshot is from a Japanese owner who ran it this week. Both of their cards show 80GB reachable. The quick version: compare OPT_FBP_DISABLE with OPT_FBP_DEFECTIVE. Disabled means switched off at the factory. Defective means actually dead. A typical 8GB card reads 0x852 and 0x840, which means two partitions are off but healthy, so about 80GB is in there. Four defective caps you at 64GB. Or just paste this into your AI agent on the box with the card: You are on a Linux system with an NVIDIA CMP 170HX (GA100) GPU. Check whether this card is viable for VRAM unlock (especially 80 GB) using the read-only gpcprobe tool. Steps: 1. Clone the tool: git clone github.com/matstatman/gpcpro… cd gpcprobe 2. Build and load the kernel module (requires kernel headers and root): make sudo insmod gpcprobe.ko sudo dmesg | grep gpcprobe 3. Run the probe: sudo python3 gpc_probe.py 4. Also capture nvidia-smi info: nvidia-smi -q Analyze the gpcprobe output carefully. Focus on: - OPT_FBP_DISABLE / OPT_FBP_DEFECTIVE - OPT_FBPA_DISABLE / OPT_FBPA_DEFECTIVE - Effective FBPA count and HBM stack layout (full/half stacks, 2 GiB vs 4 GiB dies) - Which partitions are floorswept (disabled but potentially enableable) vs physically defective - GPC/SM status - Any notes about 80 GB / 96 GB potential Summarize in plain English: - Is the card healthy enough for a standard 64 GB unlock? - Is 80 GB realistically possible (enough non-defective FBPAs alive)? - Any red flags (too many defective partitions, bad HBM stacks, etc.)? - Quote the key register values and the HBM configuration line. Do not modify any fuses or attempt writes. This is read-only inspection only. Reading the fuses is safe. The unlock itself burns fuses permanently and can brick a card. If you run it, what does your card show?

Post attachment
Matt Virus avatar
Matt Virus@themattvirus·Nov 16

This is so awesome. Looking forward to modifying my stack of cmp170s

lifcc@mylifccNov 16

这绝对是今年最让人拍案叫绝的硬件逆向神作:安全研究员 Kate 发现英伟达旗舰计算核心 GA100 芯片内部存在一个电压故障注入漏洞(Voltage Fault Injection),并亲手把一张当年矿潮遗留下来的阉割版 CMP 170HX 8GB 矿卡,原地解锁成了满血的 A100-80GB! 当年英伟达为了防止矿工抢占数据中心显卡,拿 GA100 大核心切出了 CMP 170HX 矿卡,不仅在硬件熔丝(Fuses)层面砍掉了显存容量、禁用了 NVLink 高速互联、把 PCIe 降到 Gen1,还锁死了张量计算吞吐。 Kate 顺着 GA100 上电启动读取板载熔丝配置寄存器的微妙时间窗口,用精密控制的电压毛刺(Voltage Glitching)去干扰芯片的熔丝传感放大器——让芯片在自检瞬间把原本烧断的限制熔丝全部误读为 0(未熔断状态),再配合重新植球焊满的高带宽显存,一举解开了英伟达在硅片层面的全部封印:不仅显存认满 81,920 MiB(80GB),连原本被阉割的 PCIe Gen4、NVLink、MIG 切分 也全数复活。 更离谱的是实测性能:因为英伟达在卖零售版 A100 时为了保证晶圆良率,统一锁在了 108 个流多处理器(SM);而这张被毛刺攻击彻底放飞自我的 GA100 核心,居然把晶圆上全部完好的 120 到 126 个 SM 单元 全点亮了,跑出来的真实算力比官方正版 A100 还要快出整整 1.2 倍。

Post attachment
ノメル avatar
⟳ Matt Virus retweeted
ノメル@NolmerPortion·Nov 15

CMP 170HX 80GB VRAM改造成功! キターーーーーー!!!!!!!!!

Post attachment
Matt Virus avatar
Matt Virus@themattvirus·Oct 26

maybe a job for GLM, deepseek flash, and some CMP170 assist action? Pinging @nmatt0 on this to make sure it's crossed his feed.

Has anyone used AI to get root control of a John Deere tractor? don’t really need right to repair if you can jailbreak it 😎

Matt Virus avatar
Matt Virus@themattvirus·Oct 12

No beans here...but finishing hay season. Yields down by 50%. Fertilizer and diesel fuel costs up by 50%. Glad to be done. Hoping for better next season.

Post attachmentPost attachment
Matt Virus avatar
Matt Virus@themattvirus·Sep 8

YES.

And, that, kids, is why your math department should invest in an open-source model running locally on your own secure hardware that isn’t backdoored by Silicon Valley tech bros out to steal your work!

Matt Virus avatar
Matt Virus@themattvirus·Aug 22

Yep. Let's go.

Post attachmentPost attachment

🤯 How about a 64GB GA100 for $1,500? 🔥Nvidia sold this GPU with 8GB of memory. Local AI hackers unlocked 64GB of HBM2e AND its crippled Tensor Cores. Local AI hackers have now unlocked ... 🔥 8GB → 64GB HBM2e 🔥 crippled Tensor/SM compute 🔥 PCIe Gen1 → Gen2 Prices of used CMP 170HX are already shooting 🆙 The NVIDIA CMP 170HX, a 2021 crypto-mining card built from GA100 silicon, was sold by NVIDIA, but they nerfed it! 🔫 🔒 8GB HBM2e 🔒 throttled Tensor/SM compute 🔒 restricted PCIe 🎯The open-source CMP Unlocker can now turn the 8GB version into: 🔥 64GB HBM2e 👀 🔥 full SM compute throughput 🔥 PCIe Gen2 Here is what the author benchmarked... 🧮 Tensor throughput 6.3 → 193 TFLOPS 🦙 Llama 2 7B Q4_0 / llama.cpp pp512 599.6 → 3,468 tok/s That's a 5.8× increase in prompt processing. They dug deeper and found what appears to be a deliberate compute throttle ... 🐌 Locked → 256 cycles per MMA instruction 🚀 Unlocked → 24 cycles 🎉The card suddenly gets roughly95% of A100 per-SM Tensor throughput in their test. And remember ... this is now a 64GB HBM2e Local AI GPU. 👀 Here are the caveats ... ⚠️ Linux + patched NVIDIA open driver ⚠️ Secure Boot off ⚠️ PCIe still only Gen2 x4 without hardware mods ⚠️ unlocked memory has no ECC ⚠️ community hack — definitely enthusiast territory 🔗 Reddit: r/LocalLLM/comments/1vulabl/ 🔗 GitHub: /amoghmunikote/cmpunlocker

Post attachment
Matt Virus avatar
Matt Virus@themattvirus·Apr 12

The off-grid solar setup @_kernelcon_ worked great. Thursday's rain meant we ran on battery, Friday's sun powered everything and charged the 16kwh battery back to 100% and powered the equipment directly. Max wattage from the array was 3124 watts. 30 laptops. 1 local AI server, core network gear + 2x starlink dish setups. Any questions let us know! Happy to post an overview of what we used, what we liked, and what we would do differently. kernelcon.org hackspace.io @hackspaceio

Post attachmentPost attachmentPost attachmentPost attachment