News Posts matching #Rubin

Return to Keyword Browsing

NVIDIA Announces Rubin HPC GPU to Accelerate Agents for Science, Engineering and Analytics

Announced today at SC26, NVIDIA Rubin HPC is a new GPU in the NVIDIA Vera Rubin platform designed to accelerate simulation and data processing for science, engineering and analytics workflows.

As agents help scientists and engineers explore more candidate materials, molecules and designs, they rely on simulations to test those ideas and guide their next steps. Rubin HPC will accelerate those simulations, shortening the path from exploration to scientific discovery and industrial design.

SK hynix Bets on HBM4 Demand Beyond NVIDIA With New Equipment Orders

SK hynix reportedly plans to add at least 20,000 wafers per month of HBM capacity. According to Korea's Electronic Times (ETNews), the company is in talks with equipment suppliers over a fresh round of HBM4 orders, mostly back-end tools for bonding, dicing, and inspection. Suppliers are aiming to have it all installed in the first half of 2027. The likely site is P&T6 in Cheongju, a former fab SK hynix picked for HBM in 2025. For scale, ETNews put SK hynix's total HBM capacity at around 150,000 wafers per month at the end of last year. The 20,000 figure isn't entirely new, though, since ETNews floated a similar number for P&T6 in February.

Importantly, this comes while NVIDIA is still dealing with the fallout of the Jensen Huang situation, which led to Huang resigning as CEO on November 12. Apparently, demand from AMD and Broadcom is seen as strong enough to carry HBM orders even if NVIDIA's AI business takes a hit. NVIDIA's move to lower the HBM4 speed for Vera Rubin from 11.7 Gbps to 9.6 Gbps per pin also eases the manufacturing burden for suppliers, and the report says it likely factored into the decision.

TSMC Could Charge Up to 20% More for CoWoS Packaging Amid AI Crunch

TSMC is reportedly considering raising prices for its CoWoS (Chip on Wafer on Substrate) advanced packaging by as much as 20%, according to supply chain sources cited by Taiwan's Commercial Times. Foundries and outsourced packaging firms are apparently weighing another round of 10% to 20% increases, as CoWoS demand has spread beyond AI GPUs to server CPUs and custom ASICs, leaving a growing backlog of unfilled orders. TSMC hasn't confirmed anything yet.

Testing is turning into a bottleneck of its own. Nomura estimates, via Seoul Economic Daily, that NVIDIA's Blackwell takes roughly four times longer than Hopper in final testing, while Rubin could need seven times as long. Packages keep growing too, with TSMC already mass producing CoWoS-L at 5.5 times the reticle limit. Tool lead times have reportedly stretched from three to six months to as long as eight months. TSMC, which usually negotiates for discounts or extra features, apparently placed this year's equipment orders at existing prices instead, which says a lot about how tight things are.

NVIDIA Cuts HBM4 Speed and NVLink Bandwidth Targets for "Vera Rubin"

NVIDIA has officially implemented changes across its "Vera Rubin" product line, reducing target specifications for memory and interconnect fabric. According to the updated product specification sheet, the first change involves HBM4 memory. Previously, NVIDIA used HBM4 with a speed of 11.7 Gbps per pin, which has now been reduced to 9.6 Gbps per pin. In a "Vera Rubin" NVL72 rack-scale solution, NVIDIA now offers a total memory bandwidth of 1,400 TB/s, down from the previously advertised 1,580 TB/s. While this issue does not stem directly from NVIDIA's suppliers, it indicates that the company's implementation of memory bus width and overall electrical signaling is insufficient to support higher speeds.

However, the changes are not limited to memory. NVIDIA has also reduced the previously advertised NVLink bandwidth per GPU, decreasing it from 1.8 TB/s to 1.5 TB/s. This results in a reduction from 3.6 TB/s to 3 TB/s in bidirectional NVLink GPU bandwidth. The underlying issue is related to signal integrity, as NVIDIA reportedly lowered the NVLink Nyquist frequency target from 56 GHz to 47 GHz. Lowering the Nyquist frequency reduces the electrical demands on the SerDes, cables, and materials, which can improve manufacturing yields and overall feasibility, even though it requires accepting a small reduction in peak interconnect bandwidth. Thus, NVIDIA chose to sacrifice some bandwidth to enhance system stability. Below is the new system specification in light mode, and old system specification in dark mode.

NVIDIA Eyes Software Revenue With a $2,000 per GPU Yearly Figure From Huang

NVIDIA CEO Jensen Huang has put a rough price on the software side of the company's AI business. During a Q&A session at GTC Berlin last week, Huang said software licensing will become a growing part of NVIDIA's business, and that it costs about $2,000 per GPU per year to run all the NVIDIA software that powers agents in AI factories. The Register was first to report it, and Huang seems to have offered it as an example rather than a formal price.

The number is a bit confusing, since NVIDIA already sells its AI Enterprise suite on a per-GPU basis. The company's own licensing guide lists it at $4,500 per GPU for a one-year subscription with support, or $18,000 per GPU for five years. NVIDIA hasn't said whether Huang's $2,000 figure covers the same software, a narrower agent-focused bundle, or a future licensing model, so it probably shouldn't be read as a price cut. There's also no word on timing.

Rumor: NVIDIA RTX 6060 May Be Years Away, and Possibly Slower Than the PS6

Anyone holding out for an affordable next-gen NVIDIA card might have a long wait. Hardware leaker Kepler_L2 claims on NeoGAF that the GeForce RTX 6060 will "launch much later," adding that, based on recent trends, he doubts it will even match the PlayStation 6. He was replying to a forum member who argued the PS6 would compete with the RTX 6060, in the same thread where he claimed the console's RAM and storage alone cost over $680. Kepler didn't give an actual date, and the 2029 figure going around seems to be an extrapolation. Both Kepler and kopite7kimi have placed NVIDIA's Rubin-based RTX 60 series in 2028, and the x60 cards usually trail the flagship by a few months. The RTX 5060 arrived under four months after the RTX 5090, while the RTX 4060 took over eight months to follow the RTX 4090.

With a flagship launch around mid-2028 or later, that could push the RTX 6060 into 2029. That said, the "much later" comment seems to be relative to the PS6, which is reportedly targeting late 2027, so the exact timing is still anyone's guess. Kepler didn't say why, but the ongoing memory shortage looks like the obvious suspect. NVIDIA reportedly postponed its RTX 50 SUPER refresh earlier this year to prioritize AI products that use the same GDDR7 memory, and the company recently admitted that RTX 5090 supply is tight. Of course, NVIDIA hasn't announced the RTX 60 series officially yet, so treat this as a rumor. Anyone looking to buy a new NVIDIA GPU should probably expect the current RTX 50 lineup to stick around for quite a while.

Frore Systems Announces New LiquidJet Diamond Coldplates for AI Factories

Frore Systems today announced LiquidJet Diamond, a new AI Factory liquid cooling coldplate featuring an integrated diamond spreader. The accelerating demand for AI Tokens, coupled with energy scarcity, is spurring hyperscalers to find new ways to increase the efficiency of AI Factories. LiquidJet Diamond Coldplate boosts AI Tokens/Watt and revenue by 35% compared to traditional skived coldplates, with the increased revenue flowing directly to profit.

LiquidJet Diamond builds on the previously announced Frore Systems LiquidJet coldplate. LiquidJet coldplate employs a unique 3D ultra short-loop multi-stage design which can be precisely customized to the power map of each GPU, including extreme hotspots. This precision is critical as GPUs grow in complexity and power draw to meet AI demand. LiquidJet delivers 12°C reduction in GPU die temperature which directly translates to 25% higher tokens/watt efficiency. LiquidJet Diamond adds an additional spreader with integrated diamond wafers to the coldplate bottom, positioned precisely at the hotspot locations. LiquidJet Diamond delivers an additional 10 °C reduction in hotspot die temperature, boosting increase in Tokens/Watt to 35%.

Samsung HBM4E Passes Testing for NVIDIA's "Rubin Ultra" GPU

Samsung's upcoming HBM4E memory has reportedly passed testing for NVIDIA's next-generation "Rubin Ultra" AI GPU accelerator family. This is great news for the company, as its previous HBM3E didn't meet NVIDIA's requirements, leading NVIDIA to choose Samsung for GDDR7 only, instead of HBM3E. The new HBM generation brings significant advancements, and Samsung is one of the leading manufacturers with designs currently running at 14 Gbps per pin and 2,048 pins per HBM4E stack. Samsung can achieve a maximum bandwidth of 3.6 TB/s, and at 16 Gbps, this maximum memory bandwidth will increase to 4 TB/s for a single HBM4E stack.

With about a dozen memory stacks per AI accelerator or GPU, the total memory bandwidth will be remarkable. The company is currently offering this memory in a 12-layer (or 12-high) configuration with a memory density of 48 GB per stack. However, customer requirements will also include 8-layer 32 GB and 16-layer 64 GB stacks in the future.

NVIDIA GeForce RTX 60-Series Reportedly Slips to 2028, Per New Leak

Kopite7Kimi, a leaker with a strong track record on NVIDIA hardware, says the company's next-generation GeForce RTX 60-series based on the Rubin architecture is a 2028 product. "Unfortunately, GR20x is a product from 2028. That means they have postponed again," the leaker posted on X. Their earlier estimate was the second half of 2027, so this is another "delay," if you can even call it that. NVIDIA hasn't said anything official about RTX 60-series timing, so treat it as a rumor for now.

Interestingly, this new leak seems to align with what prominent leaker Kepler_L2 claimed a few days ago: that Rubin gaming GPUs would only arrive in 2028. It also goes against leaker Moore's Law Is Dead, who said about a week ago that an NVIDIA source told them gaming Rubin was greenlit for the first half of 2027. Two leakers against one doesn't settle anything, and we don't know who is right yet, but the 2028 timing does seem more likely given the current industry conditions. Of course, these plans can change before 2028, as the memory supply situation is currently very volatile, and hardware roadmaps are changing almost on a weekly basis.

Rumor: No NVIDIA "Rubin" and AMD RDNA 5 Gaming GPUs Until 2028

According to hardware leaker Kepler_L2, known for accurate hardware predictions, the next generation of gaming GPUs might be skipped next year, with a full-stack refresh expected in 2028. Reportedly, NVIDIA's GeForce RTX 60-Series "Rubin" and AMD's RDNA 5-based Radeon RX 10000 series will only be available in 2028, with some select RDNA 5 GPUs from AMD being made available throughout 2027. The specific model is yet to be determined, but it reportedly carries the codename "AT2." Perhaps AMD will roll out this GPU as either a new flagship design and launch the remainder of the stack later on, or simply as a mid-range or low-end SKU to kickstart the RDNA 5 launch. Thus, we still need to wait and gather more information before making conclusions, so treat this only as a rumor for now.

For example, the NVIDIA GeForce RTX 50-Series SUPER refresh had been planned before, but the recent DRAM scarcity has turned the launch into speculation, with some believing it is entirely canceled, while others believe it will be a stopgap solution before RTX 60 "Rubin" arrives. For AMD, we hope to see RDNA 5 finally compete in the flagship SKU sector again and see more powerful designs from the start.

NVIDIA Secures TSMC A16 Node for Next-Generation "Feynman" GPUs

According to supply chain sources and those close to TSMC, NVIDIA is already working on prototyping the replacement for its "Rubin" and "Rubin Ultra" GPU designs with the next-generation "Feynman" GPU microarchitecture. As noted in the latest report from DigiTimes, NVIDIA is opting for TSMC's A16 node, bypassing the N2 family of nodes entirely in favor of the more advanced 1.6 nm A16 design. Mass production is scheduled for the second half of 2028. The "Feynman" designs will rely on the backside power delivery in the A16 node, as well as on packaging technologies like 3D chiplets using SoIC, CoWoS-L 2.5D integration, and next-generation panel-level packaging.

Initially, NVIDIA will use SoIC 3D stacking technology to stack multiple chiplets into a single package, which will shorten the travel time between chips and reduce latency. These 3D chips will then be packaged side-by-side in a CoWoS-L or CoPoS packaging, creating a multi-kilowatt design capable of processing dozens of PetaFLOPS. For memory, NVIDIA will work with its partners to develop customized HBM, likely HBM4E, by the time "Feynman" reaches mass production. This customized memory includes a specialized base die that can contain logic elements like a memory controller or packet processing unit. Think of it as a way to pre-process data before it enters memory, or as a simple speedup, depending on what NVIDIA aims to achieve.

NVIDIA's Six-Years-Old A100 "Ampere" GPU to Remain in Use Until 2029

With the expansion of AI, every bit of available computing power is being utilized, and even NVIDIA's six-year-old A100 "Ampere" accelerator is being fully utilized at CoreWeave. According to CoreWeave, and confirmed by NVIDIA's CEO Jensen Huang, the company has signed an agreement with a customer to rent chips released more than half a decade ago for several more years, specifically through 2029. This demonstrates that NVIDIA's platforms remain in high demand, with chips released in 2020 expected to have a usable lifespan extending nearly to the end of the decade. CoreWeave customers have rented these chips through 2029 for AI training and inference. In late 2020, NVIDIA introduced a GA100 GPU SKU for the A100, featuring 6,912 CUDA cores and up to 80 GB of HBM2E memory. By today's standards, the core count may not match modern solutions, but in an era where memory shortages affect everyone, even older HBM2E is acceptable when there is 80 GB of it per card.

Despite the later-introduced "Hopper," "Blackwell," and modern "Rubin" generations offering significant, multifold improvements in efficiency, performance, and memory capacity, every available computing resource is currently being utilized. CoreWeave confirmed that a customer is renting this six-year-old GPU through 2029, indicating that it still holds value for someone. This GPU likely costs much less to access per hour than modern solutions, but it probably performs well enough that customers are still willing to pay for it. On his social media feed, NVIDIA CEO Jensen Huang posted the following message.

NVIDIA Details the "Rubin" Architecture: Die Annotation, Vera CPU, HBM4, and Disaggregated Inference

NVIDIA has released a technical details on its next-generation "Rubin" GPU, including a fully annotated die shot. "Rubin" represents a significant generational leap in NVIDIA's mission to scale computing across numerous racks, systems, and data centers. The company is promoting this generation as the first "agentic AI" native design. This AI paradigm means that inference workloads are continuously active, engage in reasoning, planning, tool utilization, and execute many sequential steps. In this context, inference refers to a completed model running on the GPUs, rather than being trained, which is what Rubin is designed for. NVIDIA claims up to 10x more agentic throughput per unit of energy compared to "Blackwell." This improvement relies not only on the new "Vera" CPU and rack-level power management but also on the entire system, as customers now prioritize overall system performance rather than just the performance of a single GPU.

In its full configuration, "Rubin" combines two compute dies in a single package, linked with a high-speed connection that NVIDIA calls NV-HBI. This chiplet approach uses CoWoS-L packaging from TSMC, chosen because each die is already as large as the chip factory can print in one go, reaching the reticle limit. The GPU contains 336 billion transistors, up to 224 SMs, 896 Tensor Cores with a third-generation Transformer Engine, and 288 GB of HBM4 memory, capable of delivering up to 50 PetaFLOPS of inference and about 35 PetaFLOPS of training at the ultra-compact NVFP4 format. This is NVIDIA's proprietary 4-bit data format that enhances efficiency in inference and training with minimal to no accuracy loss compared to the 8-bit and 16-bit data formats used today for training and inference. However, this is quite different from the FP16/FP32 formats used in gaming GPUs. Below, we will break down not just the GPU, but the entire platform NVIDIA has built around it.

Legendary SEGA Game Titles Come to NVIDIA's RTX Spark PC Platform

Japan's legendary SEGA and NVIDIA have announced a new partnership, marking a significant addition to NVIDIA's latest RTX Spark PC platform. Celebrating three decades of collaboration, the two companies are bringing VIRTUA FIGHTER CROSSROADS and future SEGA-branded titles to NVIDIA's RTX Spark. When NVIDIA unveiled its RTX Spark at the Computex 2026 international show, many gamers were curious whether a new Arm-based PC platform would receive the same support as its competing x86-based platforms. It turns out that NVIDIA's industry partnerships with major game studios are proving fruitful, as SEGA's upcoming titles will be running on RTX Spark laptop and desktop PCs.

Interestingly, we asked NVIDIA whether the support means a full native port of VIRTUA FIGHTER CROSSROADS or if it will just be an emulation of the native game running without issues. We are still waiting for a response and will update the article once we learn more. Modern emulation of games is usually good enough to ensure performance and playability, but a native port is always preferable as it allows for more optimizations. Either way, the release brings RTX Spark one step closer to fulfilling NVIDIA's dream of powering slim laptops and compact mini PCs with its own hardware.

NVIDIA's "Blackwell-Next" GPU Appears in Linux Kernel Patches

NVIDIA's next generation of graphics cards, under the "Blackwell-Next" name, began their integration into the Linux kernel some time ago. The latest set of patches now includes updates for handling Compute Express Link (CXL). This particular patch is intriguing because it refers to the next generation of NVIDIA GPUs with the "Blackwell-Next" architecture, even though NVIDIA has publicly confirmed the codenames for its upcoming GPU architectures multiple times. From the "Blackwell" generation, we have seen another variant called "Blackwell Ultra" in the B300 series. However, the next-generation architecture has been publicly confirmed to be called "Rubin." Therefore, it is unclear why NVIDIA's engineers are referring to the next generation GPU design as "Blackwell-Next," given that the codenames have been public for some time.

Furthermore, this is not a case of separating consumer and server products, as both consumer GeForce RTX GPUs are available in "Blackwell" and will be available in "Rubin" variants. This is also true for server-grade AI accelerators. We have learned that NVIDIA plans to update its RTX Spark SoC with the "Vera" CPU and "Rubin" GPU, ensuring name consistency across both server and consumer products. This makes the "Blackwell-Next" patch an isolated case, but it serves as a good reminder that progress toward the general launch of the first "Rubin" GPUs is underway.

NVIDIA Already Planning the 2nd and 3rd Generations of RTX Spark, Secures Supply for Robust Growth

NVIDIA in its interactions with the press at Computex 2026 says that it has secured enough supply of the 3 nm N1X silicon from TSMC to ensure robust growth of the RTX Spark processor in the PC industry. The company sees RTX Spark not as a niche product for power users and AI developers—for which it has the purpose-built DGX Spark—but it wants to win across all mainstream and performance segments of the PC market. The RTX Spark will be differentiated in many product models ranging from the entry-level to the top-spec part shown at Computex.

At Computex, NVIDIA also said that it is already working on the 2nd- and 3rd generations of the silicon behind the RTX Spark—the N2X and N3X, implementing future process nodes, and tapping into future microarchitectures. These could be the reported "Vera" CPU and "Rubin" graphics/compute, and the future "Feynman" graphics/compute. It took over 3 years of sustained co-development with Microsoft for NVIDIA to enter the PC market, and the company is here to stay.

Palit Confirms Next-Gen GALAX HOF and KFA2 GPUs Already in Development

After the unification of Palit's sub-companies, GALAX and KFA2, under one roof, Palit has confirmed that the company is already developing next-generation GALAX Hall-of-Fame (HOF) graphics cards, as well as regular GALAX and KFA2-branded GPU models. The company claims that the current generation of GPUs is here to stay, but what's especially interesting is the note about next-generation GPUs being in development. This means that the upcoming NVIDIA GeForce RTX 60-Series "Rubin" gaming GPUs will also be available in GALAX HOF and KFA2 variants, allowing extreme overclockers to access more HOF generations in the future. As Palit is one of NVIDIA's biggest add-in card partners, we expect that future product launches will remain NVIDIA-exclusive.

Additionally, Palit has emphasized that the leader of Galax Brazil, Ronaldo from TecLab, will continue with GALAX to work on the GPUs and enhance GALAX branding and market presence in every possible way.

NVIDIA's "Rubin Ultra" Reportedly Faces Issues With CoWoS-L Packaging

NVIDIA is reportedly experiencing manufacturing issues with its next-generation "Rubin Ultra" GPU design, one of the company's most ambitious chip development projects, due to the limitations of modern packaging technology. The world's largest company is already shipping customer samples of the standard "Rubin" GPUs, with mass shipments set to begin this summer. However, the current roadmap for the upgraded "Rubin Ultra" design may be encountering technological limitations, as NVIDIA's design goals are too ambitious for TSMC's packaging capabilities. Reportedly, NVIDIA plans to double the regular "Rubin" two-die package with 8 HBM4 modules into a new "Rubin Ultra" package that will include four silicon dies and 16 HBM4E modules in a single package. This configuration is scheduled for 2027, but the sheer volume of silicon may be too much for TSMC's packaging, according to Global Semi Research.

In a typical CoWoS package, TSMC usually combines multiple smaller dies and multiple HBM memory modules into a unified package that supports the entire AI build-out. However, with the ambitious "Rubin Ultra" design, NVIDIA planned to use CoWoS-L, which was expected to handle the design and concept that "Rubin Ultra" was based on. It is rumored, however, that in a 2+2 die package—meaning four dies as in this architecture—TSMC is encountering warping issues. The package—which includes a substrate—is bending in multiple directions, causing the compute dies of "Rubin Ultra" to not make complete contact with the underlying substrate. This instability means that TSMC has to explore alternatives within its packaging portfolio. One of these alternatives is a panelized approach called CoPoS, which stands for Chip-on-Panel-on-Substrate.

NVIDIA Readies Rubin-based GeForce RTX 60-series with Massive RT Performance Gains

The rumor mill has started grinding about NVIDIA's next-generation gaming GPU, and it looks like NVIDIA does not want to allow the current market environment to get in the way of implementing its product roadmap, with a roughly 2-year GeForce generation product launch cadence. The next-generation GeForce RTX 60-series will be powered by the "Rubin" graphics architecture. "Rubin" already debuted on NVIDIA's bread-winning AI GPU series, and is making its way to GeForce, reports RedGaming Tech.

The first slice of rumors about GeForce RTX 60-series predicts that NVIDIA will stick to a more conservative approach with foundry nodes, and not go with a sub-2 nm nanosheet-based node. GeForce "Rubin" will be built on some variant of the current TSMC 3 nm FinFET node. They need not be the same N3 node that's in use by Apple, Intel, and others; and NVIDIA might collaborate with TSMC on an exclusive variant just the way it created the NVIDIA 4N node, derived from TSMC N5. Chips in the series will follow the numbering scheme "GR20x," with examples being "GR202" for the biggest part powering the flagship product. The 3 nm node will allow NVIDIA to maintain GPU clock speeds ranging between high 2 GHz and low 3 GHz, which is a minor increase over the current "Blackwell."

Supermicro Reveals DCBBS with New NVIDIA Vera Rubin NVL72, HGX Rubin NVL8, and Vera CPU Systems

Super Micro Computer, Inc. (NASDAQ: SMCI), a Total IT Solution Provider for Cloud Computing, AI/ML, Storage, and 5G/Edge, today unveiled its upcoming system portfolio powered by the NVIDIA Vera Rubin platform. As data centers transform into AI factories, producing intelligence at massive scale, agentic reasoning, long-context AI, and Mixture-of-Experts (MoE) workloads are driving demand for an entirely new class of compute and storage infrastructure. Supermicro's NVIDIA Vera Rubin NVL72, NVIDIA HGX Rubin NVL8, NVIDIA Vera CPU systems are being designed and built with Supermicro's Data Center Building Block Solutions (DCBBS) advanced liquid-cooling technology stack to accelerate time-to-market for customers.

"We are entering a new era where every organization requires an AI factory to win in the marketplace, as the demand for inference workloads is reshaping what data center infrastructure must deliver," said Charles Liang, president and CEO of Supermicro. "Supermicro's DCBBS technology stack is being engineered to empower upcoming NVIDIA Vera Rubin NVL72, HGX Rubin NVL8, and Vera CPU systems to give our customers a fast, clear path to deploying next-generation AI factories, at scale. We are excited to provide an early look at these solutions as a testament to being first to market with the infrastructure that will power the next frontier of AI."

NVIDIA Launches BlueField-4 STX Storage Architecture With Broad Industry Adoption

NVIDIA today announced NVIDIA BlueField-4 STX, a modular reference architecture that enables enterprises, cloud and AI providers to easily deploy accelerated storage infrastructure capable of the long-context reasoning required for agentic AI. Traditional data centers provide high-capacity, general-purpose storage but lack the responsiveness required for seamless interaction with AI agents that work across many steps, tools and sessions. Agentic AI demands real-time access to data and contextual working memory to keep conversations and tasks fast and coherent. As context grows, traditional storage and data paths can slow AI inference and reduce GPU utilization.

NVIDIA STX allows storage providers to build infrastructure that keeps data close and accessible at scale, so agentic AI factories can deliver higher throughput and responsiveness across inference, training and analytics. The first rack-scale implementation includes the new NVIDIA CMX context memory storage platform, which expands GPU memory with a high-performance context layer for scalable inference and agentic systems - providing up to 5x tokens per second compared with traditional storage.

NVIDIA Lowers HBM4 Specs for "Vera Rubin" VR200 as Memory Suppliers Miss 22 TB/s Target

NVIDIA has reportedly lowered its performance requirements for the HBM4 memory used in "Rubin" GPUs, as SK hynix and Samsung are reportedly struggling to meet the ambitious performance targets set by NVIDIA. According to a new note from SemiAnalysis, NVIDIA is reducing its specification requirements for the upcoming GPU generation. Originally, NVIDIA targeted a total bandwidth of 22 TB/s for the Rubin chip, but memory suppliers seem to be having difficulty meeting these requirements. Initial shipments are expected to achieve closer to 20 TB/s, which translates to approximately 10 Gbps per pin for HBM4. This indicates that NVIDIA's aggressive upgrade plan for "Vera Rubin" is facing a setback, and the final performance will differ slightly.

Interestingly, NVIDIA's initial target for the "Vera Rubin" VR200 NVL72 system was 13 TB/s in March 2025, which was later upgraded to 20.5 TB/s by September. At CES 2026, NVIDIA confirmed that the VR200 NVL72 system is now operating at 22 TB/s of bandwidth. Compared to AMD's Instinct MI455X accelerator, which has 19.6 TB/s, NVIDIA initially had lower system bandwidth. They addressed this by using faster DRAM and improving interconnects between CPUs, GPUs, and the entire system. However, as memory makers like SK hynix and Samsung struggle to meet NVIDIA's performance requirements, we will see HBM4 speeds of about 20 TB/s for the entire "Vera Rubin" system.

NVIDIA Ships First "Vera Rubin" VR200 Samples to Customers

NVIDIA reported its full-year 2025 results with massive revenue of $215.9 billion, with $68.1 billion coming in the fourth quarter. The company's earnings call after the results were published contained some interesting information and confirmed that the first "Vera Rubin" VR200 racks are shipping to customers as samples, with volume shipping to commence in the second half of 2026. NVIDIA confirmed that the upcoming platform, which includes the Vera CPU, Rubin GPU, NVLink 6 switch, ConnectX-9 SuperNIC, BlueField-4 DPU, and Spectrum-6 Ethernet switch, will be powering the next-generation trillion-parameter models with only one-fourth of the GPUs compared to the previous-generation "Blackwell" and will reduce inference costs by up to 10 times.
NVIDIA CFO Colette KressWe shipped our first Vera Rubin samples to customers earlier this week, and we remain on track to commence production shipments in the second half of the year. Based on its modular, cable-free tray design, Rubin will deliver improved resiliency and serviceability relative to Blackwell. We expect every cloud model builder to deploy Vera Rubin.

Meta Partners with NVIDIA to Deploy Millions of Blackwell and Rubin GPUs

NVIDIA today announced a multiyear, multigenerational strategic partnership with Meta spanning on-premises, cloud and AI infrastructure. Meta will build hyperscale data centers optimized for both training and inference in support of the company's long-term AI infrastructure roadmap. This partnership will enable the large-scale deployment of NVIDIA CPUs and millions of NVIDIA Blackwell and Rubin GPUs, as well as the integration of NVIDIA Spectrum-X Ethernet switches for Meta's Facebook Open Switching System platform.

"No one deploys AI at Meta's scale—integrating frontier research with industrial-scale infrastructure to power the world's largest personalization and recommendation systems for billions of users," said Jensen Huang, founder and CEO of NVIDIA. "Through deep codesign across CPUs, GPUs, networking and software, we are bringing the full NVIDIA platform to Meta's researchers and engineers as they build the foundation for the next AI frontier."

AMD Instinct MI455X Reportedly Facing Delays, Only Low-Volume Production This Year

According to SemiAnalysis, which recently released its InferenceX benchmark comparing the inference performance of the latest AI accelerators, AMD's next-generation Instinct MI455X accelerators, designed to compete with NVIDIA's "Vera Rubin," are facing serious manufacturing difficulties, causing delays in AMD's roadmap execution. Only low-volume production will occur this year, while customers will be generating AI tokens in Q2 2027. This means that NVIDIA's "Vera Rubin" VR200 rack-scale system will be the most powerful system shipping this summer, without competition. As AMD's first rack-scale system, the "Helios" AI rack with the MI455X UALoE72 configuration packs 72 GPUs and 6th Gen AMD EPYC "Venice" CPUs in a heterogeneous system designed to run as a single homogeneous GPU, delivering massive AI performance with multi-ExaFLOP compute output.
SemiAnalysisEngineering samples and low volume production of AMD's first rack scale MI455X UALoE72 system will be in H2 2026 while due to manufacturing delays, the mass production ramp and first production tokens will only be generated on an MI455X UALoE72 by Q2 2027.
Return to Keyword Browsing
Nov 18th, 2026 13:24 CST change timezone

New Forum Posts

Popular Reviews

TPU on YouTube

Controversial News Posts