Josie Herzog (@Cute_SC2) profile picture
Germany
Joined January 2019
Photo by Josie Herzog (@Cute_SC2)Photo by Josie Herzog (@Cute_SC2)Photo by Josie Herzog (@Cute_SC2)Photo by Josie Herzog (@Cute_SC2)Photo by Josie Herzog (@Cute_SC2)Photo by Josie Herzog (@Cute_SC2)Photo by Josie Herzog (@Cute_SC2)Photo by Josie Herzog (@Cute_SC2)Photo by Josie Herzog (@Cute_SC2)Photo by Josie Herzog (@Cute_SC2)
Josie Herzog retweeted
I've been working on GLM 5.3 for some time now. Gotta say, every time I use it, I appreciate GLM 5.3 Flash even more. Is the full GLM 5.3 better? Yes ofc. It's about 15% smarter, but much slower than GLM 5.3 Flash, around 1/3 of the speed. Is it worth it? Probably not for most tasks. Just shows you how good GLM 5.3 Flash is. Can't wait for GLM 5.5, especially the Flash.
47
15
4
550
29,194
Josie Herzog retweeted
Over the past week, I’ve been overwhelmed by the love and support of friends, industry colleagues, and people I've never met. Thank you to everyone who has shown kindness to my family and me. It means the world to me. I came to the United States 53 years ago, and becoming an American citizen is one of the greatest gifts and proudest moments of my life. NVIDIA has been my life’s work. Building it into the company it is today alongside my colleagues has been a privilege and an honor. Nowhere but in America could a company like NVIDIA be possible. I love our country, and I love NVIDIA. I did not live up to the standards expected of everyone at NVIDIA, and I have made serious mistakes in my personal life. I created problems for the Board that I deeply regret. I own my mistakes completely, and I am committed to cooperating with the Board’s independent investigation. Since my departure, I’ve seen an outpouring of calls from NVIDIA’s employees, investors, business partners, and customers for me to return as CEO. Many continue to reach out to me directly as the company navigates complex technology, market, competitive, and geopolitical dynamics. NVIDIA is a once-in-a-generation company. If the Board of Directors decides that my return is in the best interest of our shareholders, our employees, the company, and America, I would be honored to answer these calls and return as CEO. Thank you again to everyone who has shared their good wishes. I am profoundly humbled by your support. To my family: I love you, and I will spend the rest of my life being worthy of yours. Jensen
60,710
121,647
24,466
1,172,949
58,239,173
Josie Herzog retweeted
New CXL 2.0 card lets DDR5 servers reuse eight DDR4 memory modules 🔗videocardz.com/newz/new-cxl-…
24
248
94
1,274
176,767
Josie Herzog retweeted
🪙It's like finding gold. An 8GB NVIDIA mining GPU just became an 80GB AI accelerator. And it's outperforming a retail A100 in an FP16 compute benchmark! Hardware researcher Kate discovered how to unlock disabled hardware on NVIDIA's CMP 170HX cryptocurrency mining card. So what's the secret? The 8GB CMP 170HX uses NVIDIA's GA100 silicon, the same GPU architecture as the A100. And it physically contains much more memory than NVIDIA originally exposed. 👀Look at this ... Stock CMP 170HX → Hardware-unlocked CMP 170HX 💾 VRAM: 8GB → 80 GiB 🧠 Streaming multiprocessors: 70 → ~120–126 🔗 NVLink: Disabled → Enabled ⚡ PCIe: Gen1 → Gen4 Now look at the reported FP16 compute benchmark: 🟢 Retail NVIDIA A100 → 72.49 TFLOPS 🔥 Unlocked CMP 170HX → 85.86 TFLOPS According to Kate, of 10x 8GB mining cards examined ✅ 9x had a functional 80 GiB of memory. ✅ 1x had 64 GiB. Kate's new method uses hardware voltage glitching and permanent fuse modifications to unlock additional memory and compute resources. Imagine 80GB of HBM on 1x GPU. Enough raw memory capacity to accommodate a 70B model at 8-bit or a much larger model at 4-bit, depending on runtime overhead. And Kate has demonstrated 2x unlocked cards connected via NVLink. From two cards originally sold as 8GB cryptocurrency mining accelerators! 🔗 Research and follow-up details in threads
34
52
21
637
66,330
Well, my 20 cards are now worth more than 200k,I can buy a fucking house with it.
If this is real... > be me > buy 8 old NVIDIA mining GPUs, 8GB each > turns out they unlock to 64GB PCIe 2 > turns out they actually unlock to 80GB PCIe 4 > turns out they're faster than A100s > my $12,000 purchase is worth $80,000 > 640GB of data centre GPUs now sit in my living room blog.kate.land/posts/using-v…
1
1
5
180
Try using t3 code, it could fix that issue and is neat.
I pay for Claude Ultra to write code. Instead, I'm spending my weekend using ChatGPT to investigate where Claude Code keeps burning my quota. Anthropic confirmed the runaway bug in the quoted post. Now on v2.1.319, I caught a 781,235 token cache rewrite just 30 seconds after a successful cache hit. Right after ToolSearch. Diagnostic: tools_changed. Posted the evidence on GitHub. I miss debugging my own code.
47
I pay for Claude Ultra to write code. Instead, I'm spending my weekend using ChatGPT to investigate where Claude Code keeps burning my quota. Anthropic confirmed the runaway bug in the quoted post. Now on v2.1.319, I caught a 781,235 token cache rewrite just 30 seconds after a successful cache hit. Right after ToolSearch. Diagnostic: tools_changed. Posted the evidence on GitHub. I miss debugging my own code.
You cannot make this up. Paying for Anthropic's top-tier Ultra plan, losing my entire weekly quota in 2 hours to a known, runaway Claude Code bug, and getting auto-rejected for a quota reset by... an AI support agent. Still ghosted by human support. Earlier this week, my weekly limit was wiped out in 2 hours. Thanks to logs, we found 2 background subagents got stuck in an infinite read loop: • 8,308 file reads (50+ per image) • 2,399 model requests • 217M tokens written to cache (image reads repeatedly invalidating prompt cache, matching open GitHub issues #106556 & #90716) I packaged all of this up, including exact timestamps, machine specs, and matching GitHub issues, and sent it to support. After days of silence, my request for a usage reset was instantly denied by a support AI bot. No human review, just an automated "no" for a known, verified platform defect. The irony of paying top dollar for AI, getting locked out of your workflow by an AI bug, and getting dismissed by a support AI is wild. @ClaudeDevs @bcherny this is how you alienate power users. Can we get actual human eyes on support tickets when platform-side bugs break paid access?
32
10
2
242
19,328
8x a100 80gb > 8x cmp 170hx > 4 DGX sparks > m5 ultra 512GB > 2x dgx sparks > m5 Ultra 256GB > MI50 32GB
8 170hx > 4 dgx sparks > m5 ultra 512gb > 2 dgx sparks > m5 ultra 256gb
2
2
292
Josie Herzog retweeted
8 170hx > 4 dgx sparks > m5 ultra 512gb > 2 dgx sparks > m5 ultra 256gb
4 DGX Sparks > M5 ultra 512Gb > 2 DGX sparks > M5 ultra 256Gb
5
1
1
19
2,613
Josie Herzog retweeted
Introducing Kimi K3.1 Pro: our most capable 3T-class open-weight model. 🔹 1M-token context, 2.8T total / 104B active params (MoE), Multimodal. 🔹 Full-project coding: games in Unreal, Unity & Godot. 🔹 Excels at parametric 3D modeling in Blender & CAD. 🔹 More reliable at finance work: Accurate modeling with hallucination down by 65%. 🔹 Matching the strongest proprietary models across coding, agentic and knowledge benchmarks. Kimi K3.1 Pro and K3.1 Pro Highspeed (2–3× faster) are now live on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API. Open weights by Nov 4, 2026. 🔗 API: platform.kimi.ai 🔗 Tech blog: kimi.ai/blog/kimi-k3-1-pro
1,302
3,242
2,385
30,306
4,310,475
Josie Herzog retweeted
Dear @NVIDIAAI @JensenHuang: the CMP 170HX is a GA100, the same chip as the A100, with ~1.6 TB/s of HBM2e. It's deliberately handicapped in firmware. We're in the middle of a compute shortage. Flip the switch, unlock them officially, and thousands of cards go back to work. Do it for the people.
Dear @nvidia, a friendly ask from a guy running 4 of your 3090s and 1 of 3080 Ti. The 30 series is 6 years old. You dont make these cards anymore. Maybe its time to hand them over to the community completely. My post about unlocking the 3090 got 270K views in a day, and the replies are the real story. People asking about the 4090, 3080, 3060 and 2080 Ti. People asking how to do it on Windows. Others listing what the community already patches on these cards: the NVENC session limit, P2P between GPUs. Thousands of people are digging through your old cards with kernel modules, 1 setting at a time, because they love this hardware and want everything it can give. Publish what is limited on consumer Ampere and Turing. Ship an official driver switch for older generations. Or just open it up. I understand the business side. Some people might keep their old cards longer instead of buying new ones. But it would be a huge gesture. You talk about local AI on stage, you ship open models, you sell DGX Spark for home. This would show the care goes all the way down to a used 3090 under someone’s desk. Give them to us. We will do the rest: benchmarks, guides, models tuned for these cards. @NVIDIAAI
7
5
3
59
6,596
Josie Herzog retweeted
My 3090s just got 1.9x FASTER at BF16 tensor math!!! Same cards, same driver, same 300W, one Linux kernel module. bf16 tensor pipe: 82 -> 155 TFLOPS cuBLAS BF16 GEMM 8192: 64 -> 99 TFLOPS TF32: 41 -> 79 TFLOPS NVIDIA ships every 3090 with fp32-accumulate tensor math at half speed. The driver even reports it: FMLA32 = 1/2. gpu-unlock by Zagitta sets it to full before the NVIDIA driver loads. After a cold boot all 4 of my cards say FMLA32 = full. Real inference got less than I predicted. Strata, Qwen3.8 Flash-Next on 2 cards, prefill: 4K: +12% 10K: +12% 90K: 3205 -> 3536 tok/s, +10% Before installing I said +14 to 19% at 90K, so I was wrong. The GEMM phases got 23% shorter, but they are only about 40% of Strata prefill, and the int8 experts, copies and CUDA core kernels didnt get any faster. Outputs did not change. With the same expert cache size, my log-prob dump over 1996 teacher-forced positions is the same file byte for byte as the stock run before the module. Next is Strata itself, now its everything around the GEMMs that makes prefill slow. How I set it up is in the replies.
62
86
47
766
251,910
Josie Herzog retweeted
Replying to @araseb_
Shrödingers code, it’s both a mess and not a mess until you look at it.
2
3
1
124
9,663
Josie Herzog retweeted
Meet Mistral Large 4, aka Le Chonk. • 1T parameters, natively multimodal. 49B active. It is the best open weights model from US or Europe on aggregated benchmarks. • State-of-the-art on critical workloads, including cyber defense, manufacturing and finance and it surpasses closed frontier models on visual grounding. • Forged in Europe end-to-end and is deployable from Europe via our own Mistral Cloud infrastructure. • Available to all via API today. Working with cybersecurity partners privately. Open weights release end of October.
🤖 Made with AI
2,030
4,552
2,814
40,574
5,776,364
Josie Herzog retweeted
CMP 170HX now has ECC unlocked!
The CMP 170HX - Going even further beyond. 😎 Let's see how much more of this GPU's capabilities can be unlocked in the coming weeks and months. (P.S. PCIe 3 and 4 do not currently reliably work).
2
1
19
1,636
Josie Herzog retweeted
How do you know what hardware to buy for local AI? 🤔 The results I post on 170HXs show that the most mainstream hardware is not the best choice for people willing to tinker to get maximum performance. But some people just want the best possible performance out of the box. That’s fine too. What you need to understand are the tradeoffs you’re making. Choosing hardware for local AI fundamentally requires you to find the best balance FOR YOU of speed (how fast does it run), capacity (how big a model can it fit), convenience (is it easy to set up and maintain) and price (how much do you want to cry after you buy it). I’m going to be posting some really useful resources for people to help them make these decisions for themselves. If there’s any specific hardware, issues or questions you want me to cover, let me know!
6
1
2
18
4,965
Josie Herzog retweeted
Replying to @itsjustmarky
Don’t do it at these prices…
Do not pay $12,000 for 2 x DGX Sparks. For the same price you can buy 4 x 170HXs and a computer to run them in, for FAR better performance. Should I post complete builds at various price points?
1
1
4
254
Josie Herzog retweeted
For the love of... DO NOT BUY A 64GB DGX SPARK. Do yourself a favor, buy 2 used 3090s for half the price and triple the performance.
New @NVIDIAAI DGX Spark 64GB variants spotted! 🤩🤩 Starting Friday, Oct. 23, the new 64GB configuration will be available from Acer, ASUS, Dell, Gigabyte, HP and MSI! These will ship WITH CX7 and NVIDIA Sync Cluster Assistant! blogs.nvidia.com/blog/local-…
21
13
3
247
15,094