SCROLL NEWS / DISCOVERY
Search headlines
Find the stories shaping the conversation.
Results for " rubin " 51 found 🔀 AI-powered Shuffle
Sign in to save shuffle history.
Ubisoft's Mark Rubin Discusses Possibility of XDefiant Returning
XDefiant could end up as the one that got away, and the game's executive producer has opened up on whether the Ubisoft shooter could return.
NSF–DOE Vera C. Rubin Observatory Launches Real-Time Discovery Machine for Monitoring the Night Sky
NSF–DOE Rubin Observatory has issued its first scientific alerts, marking a historic milestone in astrophysics. Expected to increase to seven million alerts per night, these first alerts start a new era of dynamic, real-time observation of the night sky.
A newly spotted asteroid spins faster than any of its size ever seen
Among the first finds from the Vera C. Rubin Observatory, the discovery hints at a population of exceptionally strong asteroids.
Did Lil Wayne Skip Out on Maine Show to Attend Michael Rubin’s White Party?
One day after being a no-show on the opening night of his tour, Wayne attended the annual White Party.
New SCCT president, a veteran cardiologist, highlights importance of listening and teamwork
New SCCT president Ronen Rubinshtein, MD, says the fields of interventional cardiology and cardiac imaging are two of his greatest interests.
Scientists Release Biggest 2D Map of the Universe
Key Takeaways The DESI Legacy Imaging Surveys combined more than 263,000 telescope exposures to make the largest 2D map of the universe in visible and near-infrared light. Astronomers can pair the Legacy Surveys map with their own observations to explore our universe and search for rare phenomena. The 2D map serves as the foundation for the Dark Energy Spectroscopic Instrument survey to measure the universe in 3D and investigate dark energy. Hold on to your telescopes: the DESI Legacy Imaging Surveys team has released the largest-ever 2D color map of the universe. The 5.6-trillion-pixel map contains nearly 4 billion celestial objects, primarily stars and galaxies. The data is available for all to use and publicly viewable through the Legacy Survey Sky Viewer. Astronomers and citizen scientists can explore the map or combine it with their own observations to better understand our universe. Researchers can search for rare phenomena like gravitational lenses, observe fleeting events like supernovae, and investigate two of physics’ biggest mysteries: dark matter, the invisible substance that accounts for most of the mass in our universe, and dark energy, the force driving our universe’s accelerating expansion. The new map builds on earlier versions from the DESI Legacy Imaging Surveys that have already proved invaluable. To date, more than 1,800 science papers that reference the Legacy Surveys data have been published. “It’s part of the fabric of astronomy research now,” said David Schlegel, a co-lead of the Legacy Surveys and scientist at the Department of Energy’s Lawrence Berkeley National Laboratory (Berkeley Lab). “When you’re working with astronomical objects today, you often start by pulling up the Legacy Imaging Viewer to see what you’re looking at.” Covering roughly 75% of the sky in visible and near-infrared light, the updated map provides a deep view of the extragalactic universe not blocked by the dust and stars of our own Milky Way. Researchers expect it will remain the most comprehensive 2D map of our universe for years to come. More than 160 scientists contributed to data collection for the project, and a team of 20 produced the final dataset released today. It was built by combining 263,407 telescope exposures from three ground-based sky surveys: the Dark Energy Camera Legacy Survey (DECaLS) at NSF Cerro Tololo Inter-American Observatory, the Mayall z-band Legacy Survey (MzLS) at NSF Kitt Peak National Observatory, and the Beijing-Arizona Sky Survey (BASS) at the University of Arizona’s Steward Observatory. That was supplemented by years of data from NASA’s Wide-field Infrared Survey Explorer (WISE) satellite mission and additional public data. “For our team, these data are fundamental to our investigation of the expansion history of the universe and the formation of our galaxy,” said Arjun Dey, co-lead of the Legacy Surveys and an astronomer at NSF NOIRLab. “But the skies belong to everyone, and this survey gives everyone the chance to explore the sky and marvel at its wonders.” Here be galaxies The DESI Legacy Imaging Surveys were originally conducted to prepare for the Dark Energy Spectroscopic Instrument (DESI) survey. The Legacy Surveys’ 2D map is essentially a deep photograph of the sky; it records where galaxies and stars appear and how bright they appear. This crucial step enables DESI to select objects and measure their light in different wavelengths to determine their distances, building the largest high-resolution 3D map ever made. Scientists study the way galaxies have clustered at different ages of the universe to track dark energy over time. In April 2026, DESI completed its original five-year survey ahead of schedule and with vastly more objects than expected. The early results have shown surprising hints that dark energy’s impact may be weakening over time — a paradigm shift that could potentially shape the predicted fate of our universe. DESI expects to publish improved results using their first five years of data in 2027 and is continuing observations into 2028. DESI was so efficient at observing galaxies, the Legacy Surveys map needed to expand. Early on, “it became clear we might run out of galaxies to look at and run out of sky, and we better start doing something about that,” said Schlegel, who also works on DESI. The new Legacy Surveys map has been used to select DESI targets since June 2026 and will guide the telescope’s operations over the coming years. Computing the cosmos Merging hundreds of thousands of images taken on 2,285 nights, each with unique atmospheric and telescope conditions, was a massive computational effort. It took about a year to develop the computer code and eight weeks to process all the images at the Perlmutter supercomputer at the National Energy Research Scientific Computing Center (NERSC) at Berkeley Lab. Beyond supporting DESI, the Legacy Surveys will be a foundational reference for the next generation of telescopes. As new observatories like the NSF-DOE Vera C. Rubin Observatory (jointly funded by NSF and DOE’s Office of Science) and NASA’s Nancy Grace Roman Space Telescope come online, researchers can compare their observations with one of the deepest and most comprehensive views of the sky ever assembled. The Legacy Surveys data will also help scientists train artificial intelligence tools to analyze petabytes of astronomical data and accelerate new discoveries. It will be among the datasets used in an astrophysics pilot project within the American Science Cloud, part of the DOE’s Genesis Mission. The DESI Legacy Imaging Surveys are supported by the U.S. Department of Energy’s Office of High Energy Physics; the National Energy Research Scientific Computing Center, a DOE Office of Science user facility; the U.S. National Science Foundation, Division of Astronomical Sciences; and the partner institutions. DESI is supported by the DOE Office of Science and NERSC. Additional support for DESI is provided by the NSF; the Science and Technology Facilities Council of the United Kingdom; the Gordon and Betty Moore Foundation; the Heising-Simons Foundation; the French Alternative Energies and Atomic Energy Commission (CEA); the Secretariat of Science, Humanities, Technology and Innovation (SECIHTI) of Mexico; the Ministry of Science and Innovation of Spain; and by the DESI member institutions. ### Lawrence Berkeley National Laboratory (Berkeley Lab) is committed to groundbreaking research focused on discovery science and solutions for abundant and reliable energy supplies. The lab’s expertise spans materials, chemistry, physics, biology, earth and environmental science, mathematics, and computing. Researchers from around the world rely on the lab’s world-class scientific facilities for their own pioneering research. Founded in 1931 on the belief that the biggest problems are best addressed by teams, Berkeley Lab and its scientists have been recognized with 17 Nobel Prizes. Berkeley Lab is a multiprogram national laboratory managed by the University of California for the U.S. Department of Energy’s Office of Science. DOE’s Office of Science is the single largest supporter of basic research in the physical sciences in the United States, and is working to address some of the most pressing challenges of our time. For more information, please visit energy.gov/science.
Rhodes Business Students Triumph in Investment Research Competition
Rhodes College students Cole McNair ’26, John Mori ’26, Jason Rubinstein ’27, Graham Allen ’26, and Ismail Malik ’26 secured first place at the Mid-South competition of the Chartered Financial Analysts Institute Research Challenge held Feb. 21 in Memphis.
The Real Danger to Women Isn’t Hormone Therapy, It’s Bad Science
Are all hormones created equal? Drs Rachel Rubin and Kelly Casperson unpack the missing details in recent menopause research and explain why nuance matters.
Action! NSF–DOE Vera C. Rubin Observatory Begins Capturing the Greatest Cosmic Movie Ever Made
The wait is over: NSF–DOE Rubin Observatory, funded by the U.S. National Science Foundation and the U.S. Department of Energy’s Office of Science, is now capturing the cosmos in unprecedented detail, transforming the way we study the dynamic Universe.
Fanatics’ Michael Rubin on taking the sports platform to the next level
Michael Rubin has built one of the most influential platforms in sports, evolving Fanatics from a merchandise leader to a multi-pronged, billion-dollar platform. Rubin spoke with CNBC’s Andrew Ross Sorkin at the CNBC Sport x Boardroom Game Plan Summit about the opportunities he sees for Fanatics.
University of Washington astronomers discover a record-breaking asteroid
Astronomers at the University of Washington recently discovered a record-breaking asteroid using early images from the new Vera C. Rubin Observatory in Chile.
Nvidia Vera Rubin, Dow tops 49K, business optimism: 3 Things
Nvidia (NVDA) CEO Jensen Huang unveiled a variety of new projects and innovations at his keynote speech at CES 2026 on Monday, including the chipmaker's Vera Rubin AI platform. US stock futures (ES=F, NQ=F, YM=F) are mixed in Tuesday's premarket trading after the Dow Jones Industrial Average (^DJI) reached a new record high above 49,000 yesterday. A majority of small and medium business owners appear cautiously optimistic for 2026, according to a new study from Bank of America. To watch more expert insights and analysis on the latest market action, check out more Morning Brief.
Vera C. Rubin Observatory discovers enormous, record-breaking asteroid in first 7 nights of observations
In its preliminary data release, taken from just seven nights of observations, the powerful Vera C. Rubin Observatory has discovered an enormous, fast-spinning asteroid that sets a new record.
Fanatics CEO Michael Rubin on FIFA partnership, growth of sports gaming and prediction markets
Fanatics CEO Michael Rubin joins 'Squawk Box' to discuss details of the new collectibles partnership with FIFA, state of the collectibles business, growth of the sports gaming business, his thoughts on prediction markets, GameStop's offer for eBay, and more.
Jon Maples has big lead over Jonathan Rubin in HD 87 GOP Primary, and he’s got room to grow
Even with more than 1/3 of voters undecided, Maples already has majority support.
NVIDIA confirms RTX Spark roadmap with Rubin in 2027 and Rosa Feynman in 2029
RTX Spark gets a multi-generation roadmap NVIDIA confirms RTX Spark roadmap through Rubin and Rosa Feynman NVIDIA has confirmed that RTX Spark and DGX
This conservation nonprofit founder says he’s the GOP’s best shot at retaking HD 87
Jonathan Rubin says local Republican voters have soured on the party’s preferred candidate in the district.
AbCellera women’s health data top ‘most aggressive upside scenario,’ putting Astellas and Bayer on notice
AbCellera has reported phase 2 results “beyond even the most aggressive upside scenario” that its CEO dared to consider, boosting the biotech’s chances of challenging Astellas and Bayer for a women’s health market. The phase 2 clinical trial compared ABCL635, an antibody targeting NK3R, to placebo in patients with hot flashes caused by menopause. Investigators enrolled 92 postmenopausal women experiencing a mean of about 10 moderate or severe hot flashes per day. By Week 4, the number of moderate or severe hot flashes had fallen by a mean of 8.8 on ABCL635 and by 3.5 on placebo. That statistically significant, placebo-adjusted reduction of 5.3 hit the trial’s primary endpoint. The level of efficacy also cleared the bar set by existing drugs. While cross-trial comparisons can be misleading, AbCellera CEO Carl Hansen, Ph.D., had seen enough to start hyping the program. “We now view ABCL635 as a potential best-in-class non-hormonal therapy with strong differentiation against approved NK3R competitors across all important dimensions, including efficacy, safety, dosing and convenience,” Hansen said on a Monday morning conference call with investors to discuss the data. “We believe ABCL635 is now a highly de-risked program.” Last week, Hansen used AbCellera’s earnings call to set expectations for readout, with the CEO defining success as a clean safety profile and efficacy comparable to the two approved small molecules. The data vaulted over that bar. Astellas linked (PDF) one of the small molecules, Veozah, to placebo-adjusted reductions in hot flash frequency of 2.1 and 2.6 at Week 4 of its two phase 3 trials. Bayer reported (PDF) placebo-adjusted reductions of 3.04 and 3.29 at Week 4 of pivotal studies for its small molecule, Lynkuet. At 5.3, the placebo-adjusted reduction in AbCellera’s trial comfortably cleared the standard set by the incumbents. Veozah and Lynkuet came to market based on evidence of their ability to reduce the frequency and severity of hot flashes. AbCellera’s phase 2 study covered both endpoints, with the Canadian biotech reporting a mean reduction in severity of 1.4 points at Week 4 on ABCL635, compared to 0.3 points for placebo. Again, the placebo-adjusted result compares favorably to data on Veozah and Lynkuet. Analysts shared Hansen’s enthusiasm for the data. The phase 2 results “firmly validate ABCL635 as a potentially best-in-class treatment,” BMO Capital Markets analysts said in a note to investors Monday. ABCL635’s apparent efficacy edge “is likely attributable to the higher dosing/target engagement the asset is able to reach vs. the oral agents due to its wider therapeutic index,” the analysts added. The analysts, who model $2.3 billion in unadjusted peak sales for ABCL635, flagged the safety profile as another positive of the phase 2 readout. While Veozah carries a boxed warning for liver damage, nobody on ABCL635 in the phase 2 trial had elevated liver enzymes or bilirubin values. If replicated in phase 3, the lack of a liver signal could allow approval without the monitoring requirements imposed on Veozah. Investors appeared equally enthused, sending AbCellera's stock up 32% to $9.21 in premarket trading Monday from a Friday closing price of $6.93. Next, AbCellera will collect the Week 12 data and talk to its advisors about further development. The results open up “a wide variety of possibilities of how we might take this forward into phase 3 studies to truly maximize the value of the asset and make sure that ABCL635 is able to demonstrate its differentiation potential,” Sarah Noonberg, M.D., Ph.D., AbCellera’s chief medical officer, told investors.
NVIDIA Vera Rubin Delivers World-Class Supercomputers for Science
With 7 Exaflops of AI for Science and 5 Petaflops of Native FP64 Performance, Vera Rubin Packs TOP500 Supercomputing in a Single RackNews Summary: NVIDIA Vera Rubin accelerates the world's ...
Cameron Diaz and Benji Madden welcome third baby: 'Welcome to the world' Nautas
Cameron Diaz and Benji Madden expanded their family via surrogate for the third time, the musician revealed on Monday.
Lil Wayne, Karl-Anthony Towns top attendees at Michael Rubin's 2026 'White Party'
Some of the biggest names in the sports world partied with Fanatics CEO Michael Rubin at his annual "White Party."
Entertainment Notebook: Grow blooms at the Segal Centre; Grand Prix Party fever and more
The Segal Centre for Performing Arts will conclude its 2025-26 season with Grow, a bold new Canadian musical by the originating producer of Come From Away, Michael Rubinoff. The musical
Why CBS Remains Bullish on Syndicated TV Shows
The head of CBS Media Ventures, Bryon Rubin, breaks down why the subsidiary’s lineup of game shows, courtroom shows and a daytime talk show makes business sense in 2026.
What’s Coming to the Bryce Jordan Center This Fall? Here’s Every Upcoming Concert and Show So Far
The Bryce Jordan Center is gearing up for a busy fall, with a series of concerts, comedy shows and other entertainment scheduled at the Penn State venue
Vera Rubin Observatory detects one of the smallest and faintest satellite galaxies ever seen — before the telescope's mission even began
Astronomers looking at Vera C. Rubin Observatory test data have discovered 'Aquarius IV', believed to be a brand new satellite galaxy of the Milky Way located 350,000 light-years away.
Social media and youth mental health
Dr. Jeffrey Rubin, psychotherapist and psychoanalyst, joined Fox 5 DC.
The night sky could get three times brighter as new satellites launch — all but ruining the Vera C. Rubin Observatory's survey of the universe
Extremely bright satellites and megaconstellations could make the night sky up to three times brighter than it is now, a new study warns. This would seriously hinder astronomical imaging devices, like the enormous camera at the Vera C. Rubin Observatory.
Melvin Robert Shines in KTLA Spotlight as Morning News Entertainment Anchor
Ever since Melvin Robert joined KTLA’s “Morning Show,” he has lit up the show with an incandescent smile and effervescent charm as he delivers entertainment hot topics. The Inglewood native replaced Sam Rubin last year when the iconic news personality suddenly died.
Inside the sports business of Fanatics CEO Michael Rubin
In an exclusive interview with The Post, Fanatics CEO Michael Rubin opened up Fanatics' rapid rise.
NVIDIA Blackwell-Next name spotted in Linux kernel patch
NVIDIA engineers are already preparing Linux support for hardware listed as “Blackwell-Next.” The name appears in a Linux kernel patch for the NVGrace-GPU VFIO…
Menopause and estrogen patches dominated DC conversations. Now what?
Menopause took center stage in Washington, DC, with meetings about testosterone, estrogen patches, and how to improve women's health care.
Fanatics CEO Michael Rubin: I fundamentally believe in the sports betting business long term
Fanatics founder and CEO Michael Rubin joins 'Squawk Box' to discuss the company's sports betting business, the increase in marketing spend on sports betting, his thoughts on prediction markets, staying private vs. going public, and more.
Titans of themed entertainment gathered in Orlando last week for panels and an awards show
Nearly a year after the grand opening of Epic Universe, the impact of Universal’s newest gate continues to echo through the attractions industry. This was
Bristol Myers Squibb Building Life Science Industry’s Most Advanced AI Factory on NVIDIA Vera Rubin
The pharmaceutical company has demonstrated how AI can transform how drugs discovery and is now scaling AI across R&D to accelerate the discovery of next-generation medicines.
Fanatics CEO Michael Rubin on two sports businesses his multi-billion-dollar conglomerate won't enter
Fanatics has become one of the most influential companies in sports, building out a massive merchandise, collectibles and sports gambling platform.
We're aiming for a 100% growth rate of the AI server business in 2026: Asus co-CEO
S.Y. Hsu, co-CEO of Asus, says it's "highly possible" that enterprise and commercial products will be the highest revenue grossing segment of the company, and that they are aiming to begin mass production of Nvidia Vera Rubin servers in 2026. He also expects the current shortage of memory chips to last until Q4 of 2027.
Sports Powerhouse Fanatics Pushes Into Entertainment With Launch of Fanatics Studios
The joint venture with OBB Media is launching with a Tom Brady docuseries, a deal to produce the ESPY Awards and deals with LA28, MLB and the WWE.
‘Like at McDonald’s’: Sports nurse practitioner sold opioids to athletes like ‘you gotta pay for your burger,” DEA investigation finds
Joan Rubinger had a client list that included hundreds of athletes—but the DEA found she doled out thousands of illegal prescriptions.
Nvidia Aims To Speed Deployment Of AI Factories
Nvidia introduced a design for an AI data center packed into a box about the size of a standard shipping container. Nvidia stock rose.
Bilimeters now in every Sanford Health Sioux Falls clinic
Bilimeters are handheld digital devices that give an instant reading of newborn babies' bilirubin levels, as opposed to a blood draw.
Fanatics CEO Michael Rubin on two sports businesses his multi-billion-dollar conglomerate won't enter
Fanatics has become one of the most influential companies in sports, building out a massive merchandise, collectibles and sports gambling platform.
South Korea to seek priority supply of Nvidia Vera Rubin GPUs, science minister says
South Korea will ask for priority supply of Nvidia's next-generation Vera Rubin graphics processing units, as deliveries are expected to be delayed, Science and ICT Minister Bae Kyung-hoon said on Monday.
Princeton collaborators celebrate start of observatory's ‘ultra-high-definition time-lapse record of our Universe’
Princeton researchers have played a major role in developing software pipelines to process the massive amounts of data the Legacy Survey of Space and Time generates as it scans the cosmos.
‘Extremely violating’: Tinder pauses AI photo enhancement feature
"It was extremely violating," Rubin said. "Those are not my teeth. That's not what my mouth looks like."
Pitt’s new Trivedi Institute will translate space science into human health breakthroughs
Former NASA astronaut and Professor Kate Rubins will lead the $25 million institute, funded by the Trivedi Family Foundation.
Cranbrook is getting a mastodon, and it's a mammoth project
With his campus closing, 10,000-year-old mastodon Elmer will be taken apart in Waterford and reassembled in Bloomfield Hills
Millions of exploding stars could soon reveal dark energy's secrets
A new AI-powered framework could transform how astronomers measure the expansion of the Universe. By analyzing images of Type Ia supernovae and modeling their environments in unprecedented detail, researchers can estimate cosmic distances with near-spectroscopic accuracy. The technique is designed for the flood of data expected from the upcoming Vera C. Rubin Observatory and may greatly improve our understanding of dark energy.
Plaque imaging, AI and structural heart disease in the spotlight at SCCT 2026 meeting
Coronary plaque imaging is an area many experts believe is driving a shift toward earlier detection and prevention of heart disease.
The Vera Rubin Telescope Begins Surveying Our Cosmos
The Vera C. Rubin Observatory is beginning the widest, deepest survey yet of the southern sky.
Rubin Observatory alerts scientists to 800,000 new asteroids, exploding stars and other cosmic phenomena in just one night
The Vera C. Rubin Observatory sent scientists nearly 1 million astronomy alerts in one night, showing off changes in the sky. Eventually, the telescope is expected to reach 7 million alerts per night.
Engrams Embedding Entendre: Codesign for Efficient DRAM/SSD Offloading
Engram extends standard token embeddings with learned multi-token lookups. Recurring local patterns retrieve vectors directly, reducing the need to reconstruct them through attention and feed-forward layers. With Engram model architecture optimization, it allows for lower HBM capacity to be needed for models at the same quality. This does not mean there won’t be an insane demand for HBM but it just means that model architecture will continue to innovate around constraints. This model architecture design is naturally codesigned for parameter offloading: each token accesses a few embedding rows whose addresses depend on token IDs, not hidden states. The runtime can prefetch those rows from host DRAM while earlier layers compute, keeping the table outside HBM without transferring entire weight matrices. Our Memory model contains our latest estimates of quarter by quarter HBM, DRAM, & NAND supply and demand. Offloading frees HBM for model weights and KV cache, potentially supporting larger batches or more concurrent sessions. When DRAM becomes the next constraint, NVMe offers another tier. Recommendation systems already cache frequently or recently accessed embedding rows in faster memory while backing colder rows with SSDs. After NVIDIA roadmap had to change due to massively despec’ing Rubin Ultra from 1024GB to now ~200GB of HBM per chip, model architecture optimizations like emgram maybe helpful. Our DeepSeek-V4.1-Flash configuration uses roughly 189 GiB of memory for Engram. We replace it with a memory-mapped (mmap) file and measure serving performance with offload to SSD. Later on in our report, we will show our Engram offloading experiments along with the official InferenceX agentic inference serving results on Engram models like DeepSeekv4.1 Flash across all 6 NVIDIA GPU SKUs along with MI355X. Unsurprisingly, the CUDA Moat is still mogging MI355X on the ultra popular DeepSeekV4.1 Flash model. We also show how even on high capacity HBM SKUs, offloading emgrams to DRAM could result in even better performance for most of the pareto than keeping the emgram in HBM. Our benchmark has been widely reproduced, validated and/or supported by almost every major buyer of compute from Google Cloud to Microsoft Azure to Oracle, to Meta and many more. Furthermore, it has the support of the ML community including from vLLM, LMCache, SGLang, PyTorch, Huggingface and the support of major labs like OpenAI, MiniMax, ZAI, Qwen, Moonshot Kimi, etc. Star the InferenceX GitHub repository if you find the open-source benchmark and data useful!. InferenceX is the only inference benchmark in the world to have TPUv7, Jalapeño, Nvidia Rubin NVL72, AMD, and soon, SambaNova and Trainium. Due to how realistic AgentX scenario is to real world agentic inference workloads, AMD has committed to collaborating on MI455X UALoE72 too. DeepSeek did not release the original paper’s two trained Engram models. We replicated its setup on fineweb-edu using the released code and training hyperparameters, at an estimated 6E18 FLOPs per run. We observed the same U-shape scaling: Engram improved performance over pure MoE baselines. We also reproduced DeepSeek’s results, where earlier-layer representations with Engram resembled those of later layers. Like the original Engram paper, it is possible to probe Engram’s gate scores to see what n-grams DeepSeek-V4.1-Flash makes the most use of. Our gate scan finds names, code fragments, relational phrasing, and boilerplate. These examples prioritize interesting-ness over gate strength. One unexpected result was Wright : Ace Attorney. These examples suggest learned memory optimizes the training objective, not a judgment of which facts deserve storage. Licenses, bibliography fragments, API scaffolding, and website furniture can provide prediction shortcuts, so the value of additional Engram capacity may depend on what survives data preparation. This does not show that table capacity is “wasted”: the evaluation-corpus scan establishes neither training exposure nor the capacity occupied by each category. For offloading, strong gates do not identify cache-hot rows. Low gates do not automatically save reads either: computing the gate requires the retrieved key and undoes the performance gain of a fused kernel. Skipping reads would require a separate usefulness predictor before retrieval. In the original paper’s inference-time ablation, factual-knowledge benchmarks retained just 29–44% of their original performance, while reading comprehension retained 81–93%. This is due to the training–inference mismatch. The resulting degradation therefore measures this trained model’s dependence on Engram, not the performance difference between models trained with and without it. In our ablations, suppressing Engram worsens token likelihood across all evaluated domains, especially encyclopedia text and several code corpora. Surprisingly, GSM8K accuracy stays within measured run-to-run variation and removing Engram has no effect. Engram is not a detachable dictionary beside an unchanged MoE. Removing it changes downstream features and expert selection. We tested whether rerouting hurts or compensates by holding tokens fixed in a teacher-forced experiment on CRUXEval, a code-reasoning benchmark of small Python functions where the model predicts a function’s output from its code and an input, and scoring the reference answer. Removing Engram raised answer loss from 0.2848 to 0.3093 bits/token. Forcing the ablated model to use the original Engram-on expert choices made it worse still, at 0.3375 bits/token. Rerouting partially compensates for the missing memory. Memory features and expert selection work together, rather than following a clean “memory stores facts; experts reason” division. On the same CRUXeval, removing Engram during either phase reduced accuracy and increased generated tokens; removing it throughout produced the largest changes. Keeping Engram for prefill leads to more correct answers than keeping for decode likely due to semantically richer KV cache transferred to decode workers, allowing it to mitigate some of the performance loss. Engram’s table is large, but each lookup is small. DeepSeek-V4.1-Flash requests 24 rows at each of two Engram layers, about 12.4 KiB per processed token position across the model, or 3.1 KiB per GPU when split across four GPUs. Currently as of Day 7 since Model Release, MI355X is still 2-4x worse performance per dollar compared to B200 even when normalized by Mi355X’s lower TCO. Our full total cost of ownership breakdown comes from our AI Cloud TCO Model along with monthly market surveys of over 100+ gpu clouds & gpu cloud customers. On the Day 0 release of DeepSeekv4.1 Flash, NVIDIA vLLM works out of the box with zero issues across all 6 SKUs: H100, H200, B200, B300, GB200, GB300! This was thanks to the amazing work by the NVIDIA & Interact teams! In comparison, AMD vLLM did not work on day 0 for DeepSeekv4.1 Flash.Source: SemiAnalysis InferenceX AMD’s vLLM documentation points to using vllm/vllm-openai-rocm:deepseekv41-flash-0909, but from hour 0 of the model release to hour 23, AMD has not publicly released the image. AMD claims, “SPEED IS THE MOAT,” yet it has still not released it by the 23rd hour. We wish that, going forward, the AMD team has a better process for hour 0 model releases. Eventually when they did publicly release for “day 0” image support, performance-wise, it is currently up to 14.8x worse perf per dollar than H200 and up to 42x worse perf per dollar than B200/B300. The power of the CUDA MOAT is NVIDIA’s collaboration with its massive 6 million-developer community ecosystem including most of the vLLM & SGLang & Tokenspeed maintainers which means that CUDA is optimized on day 0. Overall, AMD did make significant improvements but the performance per dollar is still currently 2-4x worse than B200. AgentX Engram DRAM Offloading Improving Performance The HBM and DRAM offload use the same GPU kernel to select and dequantize rows. With HBM, it reads GPU memory; with Unified Virtual Addressing (UVA), it reads pinned host memory directly. Both support full decode graphs. Moving the table into HBM accelerates only the sparse lookup, leaving decoder computation and communication unchanged, resulting in little overall benefit while consuming memory otherwise available to KV cache. Another benefit of Engram offloaded to DRAM which means you can reduce the communication overhead by using less HBM GPUs per replica. For example, when enabling Engram offloading on B300, we are able to switch from TP4 to now TP2 which improves the pareto curve by up to 1.6x. When iso-model quality, less HBM is required as the engrams could be offloaded to host DRAM. Thus HBM bandwidth matters way more than HBM capacity. For inference workloads where memory bandwidth matters the most, 4-hi HBM provides the best $/bandwidth and therefore lowest cost per token. If China continues to make more and more revolutionary model architecture innovations, soon it could potentially 0Hi HBM stacks. [ ](https://newsletter.semianalysis.com/p/long-live-the-short-king-why-4-hi) [ Long Live the Short King: Why 4-hi HBM Wins ](https://newsletter.semianalysis.com/p/long-live-the-short-king-why-4-hi) Myron Xie, Bryan Shan, and 3 others · Sep 13 [ Read full story ](https://newsletter.semianalysis.com/p/long-live-the-short-king-why-4-hi) Moreover, on B300 and week-0 stack, moving the Engram table back to HBM did not improve results and stay within run-to-run variance. This is a result of the work optimizing DRAM offload, such as async, and overlap. SSD Offloading On B200, we created a unoptimized vLLM fork and stored the Engram tables in memory-mapped files on local SSDs. File backing lets the OS reclaim table pages when other applications need RAM. Pages already cached in memory can be served without reading the SSD again. Furthermore, note that we were unable to turn on GDS The unoptimized SSD implementation changes how rows reach the GPU. It copies row IDs to the CPU, deduplicates them, gathers the requested rows into pinned buffers, copies those rows back to the GPU and dequantizes them. This work runs between segments of the GPU execution graph. Native UVA performs row selection and dequantization directly on the GPU, avoiding the CPU round trip. A warm filesystem cache removes physical SSD reads, but leaves the coordination, row gathering and transfers. This is why a file already cached in RAM can still perform worse than a pinned DRAM table. The comparison measures the whole serving path; it does not separate the time spent on each of these operations. B200 DRAM dominates both measured SSD serving curves in total tokens per dollar and P90 interactivity. Near 125 tokens/s/user, DRAM delivers 121 million total tokens per dollar versus 52 million for SSD. For production serving, SSD offloading is likely not worth the tradeoff. On the B200 configurations we measured, SSD offloading loses on both measures: every observed SSD point has a DRAM alternative that delivers higher P90 interactivity and more total tokens per dollar. The only points where Cheaper storage does not automatically produce a cheaper inference service. Moving Engram to SSD leaves the same four expensive GPUs and the rest of the server in place. Reclaiming RAM only creates an economic benefit if it enables a cheaper server configuration or additional useful capacity. The current unoptimized path provides neither benefit, and the filesystem cache still consumes RAM when table pages are resident. Next we will look at the mechanisms and specific implementation of ngrams from DeepSeekv4.1 Flash, LongCat, Qwen3.8 Flash Next.