Bertrand Schmitt

@ptsi Pleasure, was reading some of your posts on browser control with AI, thanks for the content!

@bschmitt thank you for the follow! 👋☺️

@noguchis Excellent leveraging some ideas from your version, I am getting 50 t/s on RTX Pro 6000

ds4 on DGX Spark — 44× GPU allocator variance eclipses kernel tuning. 13.75 → 24.45 t/s: note.com/ngc_shj/n/n505…

@antirez Yes 96GB of VRAM

@bschmitt 96gb? Likely certain buffers allocation fail and we fall back at a slower path. Likely fixable but I don't have access to the HW.

Not bad. Committed.

@antirez Specs say more is possible: RTX Pro 6000 has 6.5× DGX Spark's mem BW. nvidia-smi dmon shows the GPU's memory bandwidth peaks at only ~33% during gen, while nsys shows ~1,800 small kernel launches per generated token.

Latest DS4 q2-imatrix on RTX Pro 6000 + Linux bare metal: @ ctx 8k → 479 t/s prefill · 41.7 t/s gen (env tweak DS4_CUDA_Q8_F16_CACHE_RESERVE_MB=512 = +33% prefill) vs antirez/ds4 README: DGX Spark 343/14 · M3 Ultra 468/27 · M3 Max 250/22 = 3× DGX gen, 1.4× DGX prefill. @antirez

@antirez @NVIDIAAIDev would be great to support @antirez efforts on DGX Spark and RTX Pro 6000

DS4 running on DGX Spark (GB10 / CUDA), private branch for now. 12 tokens/sec, the memory bandwidth is limited in this system, at 270GB/sec. But prefill is ways more alighed to M3 Max at ~200 t/s. I'll release when more mature, but it is almost sure that it will get merged.

@antirez Excellent news, thanks for the great work, can't wait to test and run it on RTX Pro 6000

@antirez @rikarends 80+ tokens/s is not an unreasonable target

@antirez @rikarends Actually RTX Pro 6000 memory bandwidth is 1,792 GB/s

@llmdevguy @UnslothAI @Kimi_Moonshot better have a 2nd rtx 6000 pro, and very fast RAM with an EPYC CPU

@UnslothAI @Kimi_Moonshot So if I would have rtx 6000 pro + 256gb ram, it should run around 40 t/s?

Kimi K2.6 can now run on CPU, GPU and SSD setups! 🔥 We shrank the 1T model to 340GB via Dynamic GGUFs where important layers are upcasted. Run at >40 tok/s on 350GB RAM/VRAM setups. Run full precision on 610 GB. Guide: unsloth.ai/docs/models/ki… GGUF: huggingface.co/unsloth/Kimi-K…

@Jasonch14221644 @XPlaneOfficial 100%, performance is the biggest issue at this stage for X-Plane, specially for those of us in VR

@XPlaneOfficial At present, I feel that the biggest problem with Xp is frame rate optimization. My current graphics card is 4080S, but many airports cannot even reach a frame rate of 30. I hope the official can consider optimizing the frame rate in the future, such as supporting DLSS.

It’s roadmap time! With 12.2.0 nearly ready to wrap up, we're going to reveal the next 2 releases! Thank you all for the positive reception to 12.2.0, it seems you all enjoy the new cloud rendering! (And yes, there will be more updates on weather in future) The next update will be 12.2.1 - The Gateway Scenery Update. This will be a really short and minor update, and will include ✅ New scenery assets to use in WED ✅ A new heliport After that, will be 12.3.0 - The Weather Update. This will be our next major update and will include. ✅ Weather Radar ✅ WX/WXR API ✅ G1000 Synthetic Vision ✅ A330 Improvements ✅ ExVis Device Improvements ✅ A new enhanced airport Stay tuned for more info! Let's kickstart Summer off with some great flights! 🌞

@AndreasSteno Interesting reply on this exact topic x.com/NickTimiraos/s…

I am not sure that this has received enough attention yet. In case this happens, you don't own enough debasement-bets yet

@XPlaneOfficial Excellent, I remember there was a plan for performance improvements, is it still coming? Triple screen 4K (without 3 computers) and VR still get still very poor framerates, even on the most highend hardware you can get today...

It's out! X-Plane 12.2.0 beta is here, a major graphics-based update packed full with new features. Let's GO!!! 😎 ✅ Dark-cockpit fix ✅ All new cloud rendering/scattering ✅ New tone mapping ✅ Enhanced atmospheric rendering ✅ Stability updates. ✅ Enhancements to brakes, parking brakes & new chocks. ✅ ATC fixes ✅ Airport-specific ground texture enhancements ✅ New gateway cut. Blog post: x-plane.com/2025/04/whats-… Version Notes: x-plane.com/kb/x-plane-12-…

@XPlaneOfficial Please performance improvements for triple screens and VR!

It’s roadmap time! We’re loving the the positive reception to 12.1.4, and the 12.1.X updates – glad to see that everyone is enjoying them. The next update will be 12.2.0. This is a big one – we can’t wait to show what we’ve been cooking up. Some of the features that we’re targeting: ✅ Dark Cockpits will be addressed in this update ✅ Broad Lighting Improvements ✅ Cloud Improvements ✅ Stability Improvements Stay tuned! 🥳

@markankcorn @dhh Framework Desktop has to USB4 ports, compliant with Thunderbolt 4, so can go to 40gbps

@dhh The M4 Ultra will have 4x the memory bandwidth. This tops out at ~200gb/s and the two year old M2 Ultra is already 800gb/s. Cluster will be even more painful since the Macs use Thunderbolt 5 at 120 gbps while this new box stops at 5 gbps

Found my new PC 😍. Specs on the AMD Max+ 395 chip is insane. CPU is a close match for a M4 Pro (3K/20K Geekbench 6), but it can be configured with an insane 128GB of shared RAM for less than $2,000?! That would cost $4,800 and require a Mac Studio. Will be a beast for local AI!

@neilcybart @waitbutwhy My primary reason for not using much my AVP is the lack of comfort. Even 30 min is painful. Let alone 2 hours. In comparison, my Varjo VR4 is way more comfortable, and I will use it longer.

The solution for long-duration Apple Vision Pro usage won’t be that Edward scissorshands contraception on your head. Instead, it’s a smaller and lighter headset. The tech isn’t there yet, hence Apple’s focus on easy to wear hand bands good for up to 2 hours of usage or so at a time.

Allow me to explain this absurd thing on my head (made by a 3D printer, not Apple). I am possibly the only person on Earth who uses the Apple Vision Pro 20+ hours a week. As a writer, it's just such a MASSIVE upgrade to go from sitting in an office looking at a 32-inch monitor to sitting on a mountaintop, totally immersed and undistracted, looking at a 50-foot screen. There's just one problem: the idiotic way Apple designed their headset, all the weight sits your face. This is uncomfortable and was on the way to making me look 80 years old. My friend @jcoon1800, also appalled by Apple, designed a solution: a headset that redistributes the weight from the face to the top and back of the head. The headset is now not touching my face at all—it hovers a millimeter off the skin. This is WORLD'S more comfortable, causes no skin damage, and can be worn for hours with no issue—way better than even the lightest headsets I've worn. This isn't an ad and the headset isn't for sale. I just needed to say: The fact that one dude managed to figure this out and Apple somehow did not boggles my mind. @apple please talk to Jonathan and get your shit together. VR, even with a big heavy headset, does not have to be uncomfortable! It's all in the design. (@mkbhd get in touch, you need to try this.)

Glad to see the complexity of OpenAI offering is going to disappear

@yishan Great analogy!

I think the Deepseek moment is not really the Sputnik moment, but more like the Google moment. If anyone was around in ~2004, you'll know what I mean, but more on that later. I think everyone is over-rotated on this because Deepseek came out of China. Let me try to un-rotate you. Deepseek could have come out of some lab in the US Midwest. Like say some CS lab couldn't afford the latest nVidia chips and had to use older hardware, but they had a great algo and systems department, and they found a bunch of optimizations and trained a model for a few million dollars and lo, the model is roughly on par with o1. Look everyone, we found a new training method and we optimized a bunch of algorithms! Everyone is like OH WOW and starts trying the same thing. Great week for AI advancement! No need for US markets to lose a trillion in market cap. The tech world (and apparently Wall Street) is massively over-rotated on this because it came out of CHINA. I get it. After everyone has been sensitized over the H1BLM uproar, we are conditioned to think of OMG Immigrants China as some kind of Alien Other. As though the Alien-Other Chinese Researchers are doing something special that's out of reach and now China The Empire is somehow uniquely in possession of Super Efficient AI Power and the US companies can't compete. The subtext of "A New Fearsome Power Now Under The Command of the CCP" is what's driving the current sentiment, and it's not really valid. Like, no. These are guys basically working on the same problems we are in the US, and not only that, they wrote a paper about it and open-sourced their model! It is not actually some sort of tectonic geopolitical shift, it is just Some Nerds Over There saying "Hey we figured out some cool shit, here's how we did it, maybe you would like to check it out?" Sputnik showed that the Soviets could do something the US couldn't ("a new fearsome power"). They didn't subsequently publish all the technical details and half the blueprints. They only showed that it could be done. With Deepseek, if I recall correctly, a lab in Berkeley read their paper and duplicated the claimed results on a small scale within a day. That's why I say it's like the Google moment in 2004. Google filed its S-1 in 2004, and revealed to the world that they had built the largest supercomputer cluster by using distributed algorithms to network together commodity computers at the best performance-per-dollar point on the cost curve. This was in contrast to every other tech company, who at that time just bought what were essentially larger and larger mainframes, always at the most expensive leading edge of the cost curve. (To the young people reading this, this will sound incredible to you) I worked at PayPal at the time, and in order to keep pace with the rising transaction volume, the company was forced to buy bigger and bigger database servers from Oracle. We were totally Oracle's bitch. At one point when we ran into scalability issues, the Oracle reps told us we were their biggest installation so they had no other reference point on how to help us overcome our scalability issues. We literally resorted to flipping random config switches and rebooting it. (This heavily influenced me when I was a young manager later at Facebook. I deliberately torpedoed an Oracle salesman's pitch to try and get us to switch from open source MySQL databases to an Oracle contract: of course we had scalability problems, but at least when we had them, we could open up the hood and figure out how to fix it ... assuming we had good enough engineers, and we did. When it's closed-source infra, you're at the mercy of the vendor's support engineers) Back to Google - in their S-1, they described how they were able to leapfrog the scalability limits of mainframes and had been (for years!) running a far more massive networked supercomputer comprised of thousands of commodity machines at the optimal performance-per-dollar price point - i.e. not the more expensive leading edge - all knit together by fault-tolerant distributed algorithms written in-house. Some time later, Google published their MapReduce and BigTable papers, describing the algorithms they'd used to manage and control this massively more cost-effective and powerful supercomputer. Deepseek is MUCH more like the Google moment, because Google essentially described what it did and told everyone else how they could do it too. In Google's case, a fair bit of time elapsed between when they revealed to the world what they were doing and when they published a papers showing everyone how to do it. Deepseek, in contrast, published their paper alongside the model release. Now, I've also written about how I think this is also a demonstration of Deepseek's trajectory, but that's also no different from Google in ~2004 revealing what it was capable of. Competitors will still need to gear up and DO the thing, but they've moved the field forward. But it's not like Sputnik where the Soviets have developed technology unreachable to the US, it's more like Google saying, "Hey, we did this cool thing, here's how we did it." There is no reason to think nVidia and OAI and Meta and Microsoft and Google et al are dead. Sure, Deepseek is a new and formidable upstart, but doesn't that happen every week in the world of AI? I am sure that Sam and Zuck, backed by the power of Satya, can figure something out. Everyone is going to duplicate this feat in a few months and everything just got cheaper. The only real consequence is that AI utopia/doom is now closer than ever. ==== Bonus: This is also a little similar the Ethereum PoS moment, when AI finally has a counterpoint to the environmentalists who say AI uses so much electricity. We just brought down the cost of inference by 97%!

@sc_cath Also deepseek breakthrough optimizations benefit all AI Semiconductor companies, not just Nvidia

I'd love to see a proper IO analysis of the deepseek shock on Nvidia. What we learned, is that you can do much more with an Nvidia GPU than we thought. But it’s not immediately obvious that this is bad news for Nvidia. Imagine Nvidia had just announced that its next-generation chip would be 10 times more powerful and efficient than the previous one—would that be seen as bad news for the stock? Yes, achieving a given computational goal now requires fewer GPUs, but that perspective feels shortsighted. Lower costs—both in price and energy consumption—effectively expand the range of feasible applications, which should, in turn, increase demand. So why the sharp drop in Nvidia’s stock price? One key difference between a sudden leap in next-generation chip performance and the DeepSeek shock is that the latter retroactively upgrades the existing installed base of Nvidia GPUs, making them far more efficient than expected. In a sense, Nvidia is now competing against its own previously sold GPUs. The realization that these existing chips can deliver significantly greater output can be viewed as a massive positive supply shock—one coming from an unexpected "competitor." So maybe the overall effect on the company's value is ambiguous. Afterall, Nvidia, like any monopoly, must ration quantities to maximize profits.

Pretty good list!