Asus has introduced a new flagship RTX 4090 graphics card that uses an all-in-one liquid cooling system combined with liquid metal thermal interface. Dubbed the ROG Matrix GeForce RTX 4090, Asus says that its advanced cooler combined with extremely efficient thermal interface will ensure the maximum boost clocks possible, with Asus taking clear aim of producing the fastest gaming graphics card on the market.

Proper power delivery and efficient cooling are main ways to enable consistently high CPU and GPU performance these days, so when designing its ROG Matrix GeForce RTX 4090, the company used its own proprietary printed circuit board (PCB) with an advanced voltage regulating module (VRM). Meanwhile cooling is being provided by an all-in-one liquid cooling system that removes heat not only from GPU, but also from memory and VRM, exhausting that heat via the attached “extra-thick” 360mm radiator.

But Asus says that its ROG Matrix GeForce RTX 4090 has a secret ingredient that its rivals lack: liquid metal thermal interface material (TIM) that ensures superior heat transfer from hot components to cooling systems. 

Asus does not disclose what type of liquid metal TIM it uses for graphics cards (it uses ThermalGrizzly’s Conductonaut Extreme for some laptops), bus usually such thermal interfaces are made from gallium or gallium alloys, which are liquid at or near room temperature and are great conductors of heat.

But there are also some risks and challenges associated with using liquid metal thermal interfaces. Firstly, they are electrically conductive, which means that if the material spills or is not properly contained, it could cause a short circuit. Secondly, these materials can be corrosive to certain metals like aluminum. Thirdly, applying liquid metal can be more complicated than using other types of thermal paste, requiring careful handling and precision.

Asus says that it has been using liquid metal TIMs in its laptops for years, so using them for graphics cards does not seem to be a big challenge for the company. 

Image Credit: Future/TechRadar

Asus is not disclosing the complete specifications of the ROG Matrix GeForce RTX 4090 for the moment, but it certainly hopes to make the graphics card the world’s fastest. It remains to be seen whether the product will indeed be the fastest out-of-box, but it will certainly offer a noteworthy overclocking potential when compared to regular GeForce RTX 4090 graphics boards with regular coolers.

The Asus ROG Matrix GeForce RTX 4090 will be a limited-edition card available for sale in Q3.

Source: AnandTech – Asus Details ROG Matrix GeForce RTX 4090: Liquid Cooling Meets Liquid Metal

Corsair has introduced its new Dominator Titanium series of DDR5 memory modules that will combine performance, capacity, and style. The new lineup of memory modules and kits will offer DRAM kits up to 192 GB in capacity at data transfer rates as high as DDR5-8000.

The Dominator Titanium DIMMs are based on cherry-picked memory chips and Corsair’s own printed circuit boards to ensure signal quality and integrity. Also, these PCBs are supplemented with internal cooling planes and external thermal pads that transfer heat to aluminum heat spreaders, with an aim on keeping the heavily overclocked DRAM sufficiently cooled.

With regards to performance, the retail versions of the Titanium kits will run at speeds ranging from DDR5-6000 to DDR5-8000. Which, at the moment, would make the top-end SKUs of the highest clocked DDR5 RAM on the market. Corsair is also promissing kits with CAS latencies as low as CL30, though absent a full product matrix, it’s likely those kits will be clocked lower. The DIMMs come equipped with AMD’s EXPO (AMD version) and Intel’s XMP 3.0 (Intel version) SPD profiles for easier overclocking.

As for capacity, the Titanium DIMMs will be available in 16GB, 24GB, 32GB, and 48GB configurations, allowing for kits ranging from 32GB (2 x 16GB) up to 192GB (4x 48GB). Following the usual rule curve for DDR5 memory kits, we’ll wager that DDR5-8000 kits won’t be avaialble in 192GB capacities – even Intel’s DDR5 memory controller has a very hard time with running 4 DIMMs anywhere near that fast – so we’re expecting that the fastest kits will be limited to smaller capacities; likely 48GB (2 x 24GB).

Corsair is not disclosing whose memory chips it uses for its Dominator Titanium memory modules, but there is a good chance that it uses Micron’s latest generation of DDR5 chips, which are available in both 16Gbit and 24Gbit capacities. Micron was the first DRAM vendor to publicly start shipping 24Gbit DRAM chips, so they are the most likely candidate for the first 24GB/48GB DIMMs such as Corsair’s. And if that’s the case, that would mark an interesting turn-around for Micron; the company’s first-generation DDR5 modules are not known for overclocking very well, which is why we haven’t been seeing them on current high-end DDR5 kits.

Image Credit: Future/TechRadar

Corsair has also taken into account aesthetic preferences by incorporating 11 addressable Capellix RGB LEDs into the modules. Users can customize and control these LEDs using Corsair’s iCue software. For those favoring minimalism, Corsair offers separate Fin Accessory Kits. These kits replace the RGB top bars with fins, bringing a classic look reminiscent of the original Dominator memory.

While Corsair’s new Dominator Titanium memory modules are already very fast, to commemorate their debut Corsair plans to release a limited run of First-Edition kits. These exclusive kits will feature even higher clocks and tighter timings – likely running at DDR5-8266 speeds, which Corsair is showing off at Computex. Corsair intends to offer only 500 individually numbered First-Edition kits.

Corsair plans to start selling its Dominator Titanium kits in July. Pricing will depend on market conditions, but expect these DIMMs to carry a premium price tags.

Gallery: Corsair Unveils Dominator Titanium: Up to 192GB DDR5-8000 Kits

Source: AnandTech – Corsair Unveils Dominator Titanium DDR5 Kits: Reaching For DDR5-8000

SK Hynix was one of the key developers of the original HBM memory back in 2014, and the company certainly hopes to stay ahead of the industry with this premium type of DRAM. On Tuesday, buried in a note about qualifying the company’s 1bnm fab process, the the manufacturer remarked for the first time that it is working on next-generation HBM3E memory, which will enable speeds of up to 8 Gbps/pin and will be available in 2024.

Contemporary HBM3 memory from SK Hynix and other vendors supports data transfer rates up to 6.4Gbps/pin, so HBM3E with an 8 Gbpis/pin transfer rate will provide a moderate, 25% bandwidth advantage over existing memory devices.

To put this in context, with a single HBM stack using a 1024-bit wide memory bus, this would give a known good stack die (KGSD) of HBM3E around 1 TB/sec of bandwidth, up from 819.2 GB/sec in case of HBM3 today. Which, with modern HPC-class processors employing half a dozen stacks (or more), would work out to several TB/sec of bandwidth for those high-end processors.

According to the company’s note, SK Hynix intends to start sampling its HBM3E memory in the coming month, and initiate volume production in 2024. The memory maker did not reveal much in the way of details about  HBM3E (in fact, this is the first public mention of its specifications at all), so we do not know whether these devices will be drop-in compatible with existing HBM3 controllers and physical interfaces.

Assuming SK hynix’s HBM3E development goes according to plan, the company should have little trouble lining up customers for even faster memory. Especially with demand for GPUs going through the roof for use in building AI training and inference systems, NVIDIA and other processor vendors are more than willing to pay premium for advanced memory they need to produce ever faster processors during this boom period in the industry.

SK Hynix will be producing HBM3E memory using its 1b nanometer fabrication technology (5th Generation 10nm-class node), which is currently being used to make DDR5-6400 memory chips that are set to be validated for Intel’s next generation Xeon Scalable platform. In addition, the manufacturing technology will be used to make LPDDR5T memory chips that will combine high performance with low power consumption.

Source: AnandTech – SK Hynix Publishes First Info on HBM3E Memory: Ultra-wide HPC Memory to Reach 8 GT/s

At Computex 2023, Phison is introducing a new, lower-cost SSD controller for building mainstream PCIe 5.0 SSDs. The Phison PS5031-E31T is a quad channel, DRAM-less controller for solid-state drives that is designed to offer sequential read/write speeds up to 10,8 GB/s at drive capacities of up to 8 TB, which is in line with some of the fastest PCIe 5.0 SSDs available today.

The Phison E31T controller is, at a high level, the lower-cost counterpart to Phison’s current high-end PCIe 5.0 SSD controller, the E26. The E31T is based around multiple Arm Cortex R5 cores for realtime operations, and in Phison designs these are traditionally accompanied by special-purpose accelerators that belong to the company’s CoXProcessor package. The chip supports Phison’s 7^th Generation LDPC engine with RAID ECC and 4K code word to handle the latest and upcoming 3D TLC and 3D QLC types of 3D NAND. The controller also supports AES256, TCG Opal, and Pyrite encryption.

The SSD controller is organized in four NAND channels with 16 chip enable lines (CEs) in total, allowing it to address 4 NAND dies per channel. For now Phison is refraining from disclosing NAND interface speeds the controller supports, though given the fact that the controller is set to support sequential read/write throughput of 10,800 MB/s over four channels, napkin math indicates they’ll need to support transfer rates of at least 2700 MT/s. This is on the upper-end of current ONFi/Toggle standards, but still readily attained. For example, Kioxia’s and Western Digital’s latest 218-layer BICS 3D NAND devices support a 3200 MT/s interface speed (which provides a peak sequential read/write speed of 400 MB/s).

Phison says that its E31T controller will enable M.2-2280 SSDs with a PCIe 5.0 x4 interface and a capacities of up to 8 TB. Phison’s DRAM-less controllers tend to remain in use in SSD designs for quite a while due to their mainstream posiitoning and relatively cheap price, so, unsurprisingly, Phison traditionally opts to plan for the long term with regards to capacity. 8 TB SSDs will eventually come down in price, even if they aren’t here quite yet.

Compared to the high-end E26 controller, the E31T supports fewer NAND channels and NAND dies overall, but enthusiasts will also want to take note of the manufacturing process Phison is using for the controller. Phison is scheduled to build the E31T on TSMC’s 7nm process, which although is no longer-cutting edge, is a full generation ahead of the 12nm process used for the E26. So combined with the reduced complexity of the controller, this should bode well for cooler running and less power-hungry PCIe 5.0 SSDs.

The smaller, mainstream-focused chip should also allow for those PCIe 5.0 SSDs to be cheaper. Though, as always, it should be noted that Phison doesn’t publicly talk about controller pricing, let alone control what their customers (SSD vendors) charge for their finished drives.

As for availability of drives based on Phison’s new controller, as Phison has not yet announced an expected sampling date, you shouldn’t expect to see E31T drives for a while. Phison typically announces new controllers fairly early in the SSD development process, so there’s typically at least a several month gap before finished SSDs hit the market. As Phison’s second PCIe 5.0 controller, the E31T should hopefully encounter fewer teething issues than the initial E26, but we’d still expect E31T drives to be 2024 products.

Source: AnandTech – Phison Unveils PS5031-E31T SSD Platform For Lower-Power Mainstream PCIe 5 SSDs

Teeing off an AI-heavy slate of announcements for NVIDIA, the company has confirmed that their Grace Hopper “superchip” has entered full production. The combination of a Grace CPU and Hopper H100 GPU, Grace Hopper is designed to be NVIDIA’s answer for customers who need a more tightly integrated CPU + GPU solution for their workloads – particularly for AI models.

In the works for a few years now, Grace Hopper is NVIDIA’s efforts to leverage both their existing strength in the GPU space and newfound efforts in the CPU space to deliver a semi-integrated CPU/GPU product unlike anything their top-line competitors offer. With NVIDIA’s traditional dominance in the GPU space, the company has essentially been working backwards, combining their GPU technology with other types of processors (CPUs, DPUs, etc) in order to access markets that benefit from GPU acceleration, but where fully discrete GPUs may not be the best solution.

In this first NVIDIA HPC CPU + GPU mash-up, the Hopper GPU is the known side of the equation. While it only started shipping in appreciable volumes this year, NVIDIA was detailing the Hopper architecture and performance expectations over a year ago. Based on the 80B transistor GH100 GPU, H100 brings just shy of 1 EFLOPS of FP16 matrix math throughput for AI workloads, as well as 80GB of HBM3 memory. H100 is itself already a huge success – thanks to the explosion of ChatGPT and other generative AI services, NVIDIA is already selling everything they can make – but NVIDIA is still pushing ahead with their efforts to break into markets where the workloads require closer CPU/GPU integration.

Being paired with H100, in turn, is NVIDIA’s Grace CPU, which itself just entered full production a couple of months ago. The Arm Neoverse V2-based chip packs 72 CPU cores, and comes with up to 480GB of LPDDR5X memory. And while the CPU cores are themselves plenty interesting, the bigger twist with Grace has been NVIDIA’s decision to co-package the CPU with LPDDR5X, rather than using slotted DIMMs. The on-package memory has allowed NVIDIA to use both higher clocked and lower power memory – at the cost of expandability – which makes Grace unlike any other HPC-class CPU on the market. And potentially a very big deal for Large Language Model (LLM) training, given the emphasis on both dataset sizes and the memory bandwidth needed to shuffle that data around.

It’s that data shuffling, in turn, that helps to define a single Grace Hopper board as something more than just a CPU and GPU glued together on the same board. Because NVIDIA equipped Grace with NVLink support – NVIDIA’s proprietary high-bandwidth chip interconnect – Grace and Hopper have a much faster interconnect than a traditional, PCIe-based CPU + GPU setup. The resulting NVLink Chip-to-Chip (C2C) link offers 900GB/second of bandwidth between the two chips (450GB/sec in each direction), giving Hopper the ability to talk back to Grace even faster than Grace can read or write to its own memory.

The resulting board, which NVIDIA calls their GH200 “superchip”, is meant to be NVIDIA’s answer to the AI and HPC markets for the next product cycle. For customers who need a more local CPU than a traditional CPU + GPU setup – or perhaps more pointedly, more quasi-local memory than a stand-alone GPU can be equipped with – Grace Hopper is NVIDIA’s most comprehensive compute product yet. Meanwhile, with there being some uncertainty over just how prevalent the Grace-only (CPU-only) superchip will be, given that NVIDIA is currently on an AI bender, Grace Hopper may very well end up being where we see the most of Grace, as well.

According to NVIDIA, systems incorporating GH200 chips are slated to be available later this year.

DGX GH200 AI Supercomputer: Grace Hopper Goes Straight To the Big Leagues

Meanwhile, even though Grace Hopper is not technically out the door yet, NVIDIA is already at work building its first DGX system around the chip. Though in this case, “DGX” may be a bit of a misnomer for the system, which unlike other DGX systems isn’t a single node, but rather a full-on multi-rack computational cluster – hence NVIDIA terming it a “supercomputer.”

At a high level, the DGX GH200 AI Supercomputer is a complete, turn-key, 256 node GH200 cluster. Spanning some 24 racks, a single DGX GH200 contains 256 GH200 chips – and thus, 256 Grace CPUs and 256 H100 GPUs – as well as all of the networking hardware needed to interlink the systems for operation. In cumulative total, a DGX GH200 cluster offers 120TB of CPU-attached memory, another 24TB of GPU-attached memory, and a total of 1 EFLOPS of FP8 throughput (with sparsity).

Look Closer: That’s Not a Server Node – That’s 24 Server Racks

Linking the nodes together is a two-layer networking system built around NVLink. 96 local, L1 switches provide immediate communications between the GH200 blades, while another 36 L2 switches provide a second layer of connectivity tying together the L1 switches. And if that’s not enough scalability for you, DGX GH200 clusters can be further scaled up in size by using InfiniBand, which is present in the cluster as part of NVIDIA’s use of ConnectX-7 network adapters.

The target market for the sizable silicon cluster is training large AI models. NVIDIA is leaning heavily on their existing hardware and toolsets in the field, combined with the sheer amount of memory and memory bandwidth a 256-node cluster affords to be able to accommodate some of the largest AI models around. The recent explosion in interest in large language models has exposed just how much memory capacity is a constraining factor, so this is NVIDIA’s attempt to offer a single-vendor, integrated solution for customers with especially large models.

And while not explicitly disclosed by NVIDIA, in a sign that they all pulling out all of the stops for the DGX GH200 cluster, the memory capacities they’ve listed indicate that NVIDIA isn’t just shipping regular H100 GPUs as part of the system, but rather they are using their limited availability 96GB models, which have the normally-disabled 6^th stack of HBM3 memory enabled. So far, NVIDIA only offers these H100 variants in a handful of products – the specialty H100 NVL PCIe card and now in some GH200 configurations – so DGX GH200 is slated to get some of NVIDIA’s best silicon.

Of course, don’t expect a supercomputer from NVIDIA to come cheaply. While NVIDIA is not announcing any pricing this far in advance, based on HGX H100 board pricing (8x H100s on a carrier board for $200K), a single DGX GH200 is easily going to cost somewhere in the low 8 digits. Suffice it to say, DGX GH200 is aimed at a rather specific subset of Enterprise clientele – those who need to do a lot of large model training and have the deep pocketbooks to pay for a complete, turn-key solution.

Ultimately, however, DGX GH200 isn’t just meant to be a high-end system for NVIDIA to sell to deep-pocketed customers, but it’s the blueprint for helping their hyperscaler customers build their own GH200-based clusters. Building such a system is, after all, the best way to demonstrate how it works and how well it works, so NVIDIA is forging their own path in this regard. And while NVIDIA would no doubt be happy to sell a whole lot of these DGX systems directly, so long as it gets hyperscalers, CSPs, and others adopting GH200 in large numbers (and not, say, rival products), then that’s still going to be a win in NVIDIA’s books.

In the meantime, for the handful of businesses that can afford a DGX GH200 AI Supercomputer, according to NVIDIA the systems will be available by the end of the year.

Source: AnandTech – NVIDIA: Grace Hopper Has Entered Full Production & Announcing DGX GH200 AI Supercomputer

As part of their efforts to push the boundaries on the largest manufacturable chip sizes, Taiwan Semiconductor Manufacturing Co. is working on its new Chip-On-Wafer-On-Substrate-L (CoWoS-L) packaging technology that will allow it to build larger Super Carrier interposers. Aimed at the 2025 time span, the next generation of TSMC’s CoWoS technology will allow for interposers reaching up to six times TSMC’s maximum reticle size, up from 3.3x for their current interposers. Such formidable system-in-packages (SiP) are intended for use by performance-hungry data center and HPC chips, a niche market that has proven willing to pay significant premiums to be able to place multiple high performance chiplets on a single package.

“We are currently developing a 6x reticle size CoWoS-L technology with Super Carrier interposer technology,” said said Yujun Li, TSMC’s director of business development who is in charge of the foundry’s High Performance Computing Business Division, at the company’s European Technology Symposium 2023.

Global megatrends like artificial intelligence (AI) and high-performance computing (HPC) have created demand for seemingly infinite amounts of compute horsepower, which is why companies like AMD, Intel, and NVIDIA are building extremely complex processors to address those AI and HPC applications. One of the ways to increase compute capabilities of processors is to increase their transistor count; and to do so efficiently these days, companies use multi-tile chiplet designs. Intel’s impressive, 47 tile Ponte Vecchio GPU is a good example of such designs; but TSMC’s CoWoS-L packaging technology will enable the foundry to build Super Carrier interposers for even more gargantuan processors.

The theoretical EUV reticle limit is 858mm² (26 mm by 33 mm), so six of these masks would enable SiPs of 5148 mm². Such a large interposer would not only afford room for multiple large compute chiplets, but it also leaves plenty of room for things like 12 stacks of HBM3 (or HBM4) memory, which means a 12288-bit memory interface with bandwidth reaching as high as 9.8 TB/s.

“The Super Carrier interposer features multiple RDL layers on the front as well as on the backside of the interposer for yield and manufacturability,” explained Li. “We can also integrate various passive components in the interpreter for performance. This six reticle-size CoWoS-L will be qualified in 2025”

Building 5148 mm² SiPs is an extremely tough tasks and we can only wonder how much they will cost and how much their developers will charge for them. At present NVIDIA’s H100 accelerator, whose packaging spans an interposer multiple reticles in size, costs around $30,000. So a considerable larger and more powerful chip would likely push prices higher still.

But paying for the cost of large processors will not be the only huge investments that data center operators will need to make. The amount of active silicon that 5148 mm² SiPs can house will almost certainly result in some of the most power-hungry HPC chips produced yet – chips that will also need equally powerful liquid cooling to match. To that end, TSMC has disclosed that it has been testing on-chip liquid cooling technology, stating that it has managed to cool down silicon packages with power levels as high as 2.6 kW. So TSMC does have some ideas in mind to handle the cooling need of these extreme chips, if only at the price of integrating even more cutting-edge technology.

Source: AnandTech – TSMC Preps 6x Reticle Size Super Carrier Interposer for Extreme SiP Processors

At its 2023 Technology Symposium TSMC revealed some additional details about its upcoming N4X technology that is designed specifically for high-performance computing (HPC) applications. This node promises to enable ultra-high performance and improve efficiency while maintaining IP compatibility with N4P (4 nm-class) process technology.

“N4X truly sets a new benchmark for how we can push extreme performance while minimizing the leakage power penalty,” said Yujun Li, TSMC’s director of business development who is in charge of the foundry’s High Performance Computing Business Division.

TSMC’s N4X technology belongs to the company’s N5 (5 nm-class) family, but it is enhanced in several ways and is optimized for voltages of 1.2V and higher in overdrive mode.

To achieve higher performance and efficiency, TSMC’s N4X improves transistor design in three three key areas. Firstly, they refined their transistors to boost both processing speed and drive currents. Secondly, the foundry incorporated its new high-density metal-insulator-metal (MiM) capacitors, to provide reliable power under high workloads. Lastly, they modified the the back-end-of-line metal stack to provide more power to the transistors.

In particular, N4X adds four new devices on top of the N4P device offerings, including ultra-low-voltage transistors (uLVT) for applications that need to be very efficient, and extremely-low threshold voltage transistors  (eLVT) for applications that need to work at high clocks. For example, N4X uLVT with overdrive offers 21% lower power at the same speed when compared to N4P eLVT, whereas N4X eLVT in OD offers 6% higher speed for critical paths when compared to N4P eLVT.

While N4X offers significant performance enhancements compared to N4 and N4P, it continues to use the same SRAM, standard I/O, and other IPs as N4P, which enables chip designers to migrate their designs to N4X easily and cost effectively. Meanwhile, keeping in mind N4X’s IP compatibility with N4P, it is logical to expect transistor density of N4X to be more or less in line with that of N4P. Though given the focus of this technology, expect chip designers to use this technology to get extreme performance rather than maximum transistor density and small chip dimensions.

TSMC claims that N4X has achieved its SPICE model performance targets, so customers can start using the technology today for their HPC designs that will enter production sometimes next year.

For TSMC, N4X is an important technology as HPC designs are expected to be the company’s main revenue growth driver in the coming years. The contract maker of chips anticipates HPC to account for 40% of its revenue in 2030 followed by smartphones (30%) and automotive (15%) applications.

Source: AnandTech – TSMC Details N4X Process for HPC: Extreme Performance at Minimum Leakage

Closing out the most recent earnings season for the PC industry is, as always, NVIDIA. The company’s unusual, nearly year-ahead fiscal calendar means that they get the benefit of being casually late in reporting their results. And in this case, they’ve ended up being the proverbial case of saving the best for last.

For the first quarter of their 2024 fiscal year, NVIDIA booked $7.2 billion in revenue, which is a 13% drop over the year-ago quarter. Like the rest of the chip industry, NVIDIA has been weathering a significant slump in demand for computing products over the past few quarters, which in turn has dented NVIDIA’s revenue and profitability. However, while NVIDIA’s consumer-focused gaming division has continued to take matters on the chin, the strong performance of NVIDIA’s data center group has kept the company as a whole fairly profitable, with the most recent quarter setting a segment record and helping NVIDIA to avoid the tough financial situations faced by rivals AMD and Intel.

To that end, while Q1’FY24 was not by any means a record quarter for NVIDIA, it was still a relatively strong one for the company. NVIDIA’s net income of $2 billion makes for one of their better quarters in that regard, and it’s actually up 26% year-over-year despite the revenue drop. That said, reading between the lines will find that NVIDIA paid their Arm acquisition breakup fee last year (Q1’FY23), so NVIDIA’s GAAP net income looks a bit better than it otherwise would; while non-GAAP net income would be down 21%. Meanwhile, NVIDIA’s gross margins have held strong in the most recent quarter, with NVIDIA posting a GAAP gross margin of 64.6%.

But even a solid quarter during an industry slump is arguably not the biggest news to come out of NVIDIA’s most recent earnings report. Rather, it’s the company’s projections for Q2’FY24. In short, NVIDIA is expecting revenue to explode in Q2, with the company forecasting $11 billion in sales. Should it come to fruition, such a quarter would blow well past NVIDIA’s previous revenue records – and shattering Wall Street expectations. As a result, NVIDIA’s stock has already taken off in overnight trading, and by the time the market opens a bit later this morning, NVIDIA is expected to be a $930B+ company, knocking on the door of crossing a market capitalization of a trillion dollars.

Source: AnandTech – NVIDIA Reports Q1 FY2024 Earnings: Bigger Things to Come as NV Approaches T Market Cap

Offering an update on its work with complementary field-effect transistors (CFETs) as part of the company’s European Technology Symposium 2023, TSMC has revealed that it has working CFETs within its labs. But even with the progress TSMC has made so far, the technology is still in its early days, generations away from mass production. In the meantime, ahead of CFETs will come gate-all-around (GAA) transistors, which TSMC will be introducing with its TSMC’s upcoming N2 (2nm-class) production nodes.

One of TSMC’s long-term bets as the eventual successor to GAAFETs, CFETs are expected to offer advantages over GAAFETs and FinFETs when it comes to power efficiency, performance, and transistor density. However, these potential benefits are theoretical and dependent on overcoming significant technical challenges in fabrication and design. In particular, CFETs are projected to require the usage of extremely precise lithography (think High NA EUV tools) to integrate both n-type and p-type FETs into a single device, as well as determining the most ideal materials to ensure appropriate electronic properties. 

Just like other chip fabs, TSMC is working on a variety of transistor design types, so having CFETs working in the lab is important. But it’s also not something that is completely unexpected; researchers elsewhere have previously assembled CFETs, so now it’s up to industry-focused TSMC to figure out how to bring about mass production. To that end, TSMC is stressing that CFETs are not in the near future.

“Let me make a clarification on that roadmap, everything beyond the nanosheet is something we will put on our [roadmap] to tell you there is still future out there,” said Kevin Zhang, senior vice president at responsible for technology roadmap, business strategy. “We will continue to work on different options. I also have the add on to the one-dimensional material-[based transistors] […], all of those are being researched on being investigated on the future potential candidates right now, we will not tell you exactly the transistor architecture will be beyond the nanosheet.”

Indeed, research projects take a long time and when you are running many of them in parallel, you never know which of them comes to fruition. Even at that point, it is hard to tell which of potential structure candidates TSMC (or any other fabs) will choose, Ultimately, fabs have to meet the needs of their larger customers (e.g., Apple, AMD, MediaTek, Nvidia, Qualcomm) at the time when this production node is ready for high volume manufacturing.

To that end, TSMC is going to use GAA structures for years to come, according to Zhang.

“Nanosheet is starting at 2nm, it is reasonable to project and that nanosheet will be used for at least a couple of generations, right,” asked Zhang rhetorically. “So, if you think about CFETs, we’ve leveraged [FinFETs] for five generations, which is more than 10 years. Maybe [device structure] is somebody else’s problem to worry, then you can continue to write a story.”

Source; TSMC European Technology Symposium 2023

Source: AnandTech – TSMC: We Have Working CFET Transistors in the Lab, But They Are Generations Away

Corsair has expanded the brand’s mini-ITX case lineup with the new 2000D Airflow series. The 2000D Airflow and 2000D RGB Airflow small-form-factor (SFF) cases cater specifically to compact but high-performance systems. With a volume of 24.4 liters, the Corsair 2000D series cases have enough landscape to house the most demanding hardware, including a 360 mm AIO CPU liquid cooler and full-size graphics cards up to a triple-slot design.

The 2000D Airflow is available with and without RGB-lit fans and in white or black colors. Therefore, the case comes in four different variants. Regardless, the 2000D Airflow is a mini-ITX case that prioritizes airflow for the components housed inside. For this same reason, Corsair designs the 2000D Airflow with removable steel mesh front, side, and rear panels for maximum ventilation from all directions. The case measures 18.03 x 10.67 x 7.87 inches and weighs just under 10 pounds. As a result, it doesn’t require much space whether users decide to put it on or under the desk. Being an SFF case, the 2000D Airflow only accepts mini-ITX motherboards.

The 2000D Airflow can accommodate up to eight 120 mm and two 140 mm cooling fans, doing the case’s name justice. If a user fits the 2000D Airflow with a single-slot graphics card, it opens the possibility of cooling the graphics card with two additional fan mounts. For CPU air cooling enthusiasts, the 2000D Airflow supports coolers with a maximum height of up to 6.69 inches. Given the generous amount of fan mounts, Corsair’s SFF case offers plentiful liquid cooling options. It supports 120 mm, 140 mm, 240 mm, 280 mm, and 360 mm radiators. Users can fit up to multiple radiators with an example combination of a 360 mm unit on the side and a 240 mm one at the rear in a scenario with a single-slot graphics card.

Gallery: Corsair 2000D Airflow

The 2000D Airflow has three case expansion slots, accommodating beefy graphics cards with up to three PCI slots in a vertical orientation. Consumers will have no problem fitting a GeForce RTX 4090 into the 2000D Airflow. However, they must ensure the graphics card is shorter than 14.37 inches since that’s the maximum length permitted inside the 2000D Airflow.

Storage options, however, are limited to three 2.5-inch drives, whether SSDs or hard drives, with the 2000D Airflow. In addition, one of the case’s caveats is that it only accepts SFX or SFL-L power supplies, reducing options to units with a length of up to 5.12 inches. Nevertheless, Corsair aficionados will have no issues finding an adequate unit within the brand’s ecosystem since the company offers the SF series and SF-L series with capacities varying from 600 watts to 750 watts on the former and 850 watts to 1,000 watts on the latter. Regarding the I/O design, the 2000D Airflow offers one USB 3.2 Gen 2 Type-C port, two USB 3.2 Gen 1 Type-A ports, and one 3.5 mm audio jack on the front panel.

The 2000D Airflow retails for $139.99. On the other hand, the 2000D RGB Airflow, which has three pre-installed Corsair AF120 RGB Slim fans in the front intake, will set consumers back $199.99. Corsair backs its 2000D Airflow cases with a two-year warranty. In the case of the RGB variant, the AF120 RGB Slim fans come with a three-year warranty.

Source: AnandTech – Corsair Launches 2000D Airflow SFF Cases For Triple-Slot GPUs

Last year, AMD unveiled their entry-level ‘Mendicino’ mobile parts to the market, which combine their 2019 Zen 2 cores and their RDNA 2.0 integrated graphics to create an affordable selection of configurations for mainstream mobile devices. Although much of the discussion over the last few months has been about their Ryzen 7040 mobile parts, AMD has launched four new SKUs explicitly designed for the Chromebook space, the Ryzen and Athlon 7020C series.

Some of the most notable features of AMD’s Ryzen/Athlon 7020C series processors for Chromebooks include three different configurations of cores and threads, ranging from entry-level 2C/2T up to 4C/8T, all with AMD’s RDNA 2-based Radeon 610M mobile integrated graphics. Designed for a wide variety of tasks and users, including and not limited to consumers, education, and businesses, AMD’s Ryzen 7020C series looks to offer similar specifications and features to their regular 7020 series mobile parts but expands things to the broader Chromebook and ChromeOS ecosystem too.

Source: AnandTech – AMD Launches Zen 2-based Ryzen and Athlon 7020C Series For Chromebooks

In the latest move in the tit-for-tat technology trade war between the United States and China, on Sunday the Cyberspace Administration of China announced that it was effectively banning Micron’s products from being purchased in the country going forward. Citing that Micron’s products have failed to pass its cybersecurity review requirements, the administration has ordered that operators of key infrastructure should stop buying products containing chips from the U.S.-based company.

“The review found that Meiguang’s products have serious hidden dangers of network security problems, which cause major security risks to China’s key information infrastructure supply chain and affect China’s national security,” a statement by CAC reads. “Therefore, the Cyber Security Review Office has made a conclusion that it will not pass the network security review in accordance with the law. According to the Cyber Security Law and other laws and regulations, operators of key information infrastructure in China should stop purchasing Micron’s products.”

The CAC statement does not elaborate on the nature of ‘hidden dangers’ and about the risks they pose. Furthermore, the agency did not detail which companies are considered as ‘operators of key information infrastructure,’ though we can speculate that these are telecommunication companies, government agencies, cloud datacenters serving socially important clients, and a variety of other entities that may deem crucial for the society or industries.

For U.S.-based Micron, while the Chinese market is a minor one overall, it’s not so small to be inconsequential. China and Hong Kong represent some 25% of Micron’s revenues, so the drop in sales is expected to have an impact on Micron’s financials.

“As we have disclosed in our filings, China and Hong Kong headquartered companies represent about 16% of our revenues,” said Mark Murphy, Chief Financial Officer at Micron, at the 51st Annual J.P. Morgan Global Technology, Media and Communications Conference. “In addition, we have distributors that sell to China headquartered companies. We estimate that the combined direct sales and indirect sales through distributors to China headquartered companies is about a quarter of our total revenue.”

The trade war implications aside, the ‘key information infrastructure’ wording of the government order leaves unclear for now on just how wide the Micron ban will be. Particularly, whether Micron’s products will still be allowed to be imported for rank-and-file consumer goods. Many of Micron’s Chinese clients assemble PCs, smartphones, and other consumer electronics sold all around the world, so the potential the impact on Micron’s sales could be significantly lower than 25% of its revenue so long as they are allowed to continue using Micron’s parts.

“We are evaluating what portion of our sales could be impacted by a critical information infrastructure ban,” Murphy added. “We are currently estimating a range of impact in the low single digits percent of our company total revenue at the low end and high single-digit percentage of total company revenue at the high end.”

The decision CAC decision comes after the U.S. government barred Chinese chipmakers from buying advanced wafer fab equipment, which is going to have a significant impact on China-based SMIC and YMTC, and years after the U.S. government implemented curbs that essentially drove one of China’s emerging DRAM makers out of business. Officially, whether or not the CAC decision has been influenced by the sanctions against Chinese companies by the U.S. government is an unanswered question, but as the latest barb between the two countries amidst their ongoing trade war, it’s certainly not unprecedented.

Sources: Micron, Reuters, SeekingAlpha, CAC.

Source: AnandTech – Micron Expects Impact as China Bans Its Products from ‘Critical’ Industries

Kioxia’s BG series of M.2 2230 client NVMe SSDs has proved popular among OEMs and commercial system builders due to their low cost and small physical footprint. Today, the company is introducing a new generation of products in this postage stamp-sized lineup. The BG6 series builds up on the Gen 4 support added in the BG5 by updating the NAND generation from BiCS5 (112L) to BiCS6 (162L) for select capacities. The increase in per-die capacity now allows Kioxia to bring 2TB M.2 2230 SSDs into the market. While the BG5 series came in capacities of up to 1TB, the BG6 series adds a 2TB SKU. However, the NAND generation update is only reserved for the 1TB and 2TB models.

The BG series of SSDs from Kioxia originally started out as a single-chip solution for OEMs either in a BGA package or a M.2 2230 module. The appearance of PCIe 4.0 and its demands for increased thermal headroom resulted in Kioxia getting rid of the single-chip BGA solution starting with the BG5 introduced in late 2021. The BG6 series continues the DRAMless strategy and dual-chip design (separate controller and flash packages) of the BG5.

While the performance numbers for the BG5 strictly placed it in the entry-level category for PCIe 4.0 SSDs, the update to the NAND has now amplified the performance to accepted mainstream levels for this segment. The DRAMless nature and use of the system DRAM (host memory buffer – HMB) for storing the flash translation layer (FTL) handicaps the performance slightly, preventing it from reaching high-end specifications. However, this translates to lower upfront cost and better thermal performance / lowered cooling costs – which are key constraints for OEMs and pre-built system integrators.

The company is focusing on the 1TB and 2TB SKUs with BG6 due to higher demand for those capacities in the end market. The 256GB and 512GB variants are under development. While the M.2 2230 form-factor is expected to be the mainstay, Kioxia is also planning to sell single-sided M.2 2280 versions for systems that do not support M.2 2230 SSDs.

In addition to client systems, Kioxia also expects the BG6 SSDs to be used as boot drives in servers and storage arrays. Towards this, a few features that are not considered essential for consumer SSDs (such as support for NVMe 1.4c specifications including interfacing over SMBus for tigher thermal management, encryption using TCG Pyrite / Opal, power loss notification for protection against forced shutdowns, and platform firmware recovery) are included.

The availability of performance numbers for the 1TB SKU allows us to note that the BG6 has more than 1.7x the sequential performance numbers of the BG5, and random reads are 1.3x better, while random write performance has doubled. These are obviously fresh out-of-the-box numbers (as typical of specifications for consumer / client SSDs). Power consumption numbers were not made available at the time of announcement.

Kioxia will be sampling the drives to OEMs and system integrators in the second half of the year. Systems equipped with these drives can be expected in the hands of consumers for the holiday season or early next year. Pricing information was not provided as part of the announcement, but Kioxia is demonstrating these at the Dell Technologies World 2023 being held in Las Vegas from May 22 – 25.

Source: AnandTech – Kioxia BG6 Series M.2 2230 PCIe 4.0 SSD Lineup Adds BiCS6 to the Mix

Micron this week officially said that it would equip its fab in Hiroshima, Japan, to produce DRAM chips on its 1γ (1-gamma) process technology, its first node to use extreme ultraviolet lithography, in 2025. The company will be the first chipmaker to use EUV for volume production in Japan and its fabs in Hiroshima and Taiwan will be its first sites to use the upcoming 1γ technology.

As the only major DRAM maker that has not adopted extreme ultraviolet lithography, Micron planned to start using it with its 1γ process (its 3rd Generation 10nm-class node) in 2024. But due to PC market slump and its spending cuts, the company had to delay the plan to 2025. Micron’s 1γ process technology is set to use EUV for several layers, though it does not disclose how many layers will use it. 

What the company does say is that its 1γ node will enable the world’s smallest memory cell, which is bold claim considering the fact that Micron cannot possibly know what its rivals are going to have in 2025.

Last year the 1-gamma technology was at the ‘yield enablement’ stage, which means that the company was testing samples of DRAMs extensive testing and quality control procedures. At this point, the company may implement innovative inspection to tools to identify defects and then introduce certain improvements to certain process steps (e.g., lithography, etching) to maximize yields.

“Micron’s Hiroshima operations have been central to the development and production of several industry-leading technologies for memory over the past decade,” Micron President and CEO Sanjay Mehrotra said. “We are proud to be the first to use EUV in Japan and to be developing and manufacturing 1-gamma at our Hiroshima fab.

To produce memory chips on its 1-gamma node at its Hiroshima fab, Micron needs to install ASML’s Twinscan NXE scanners, which cost about $200 million per unit. To equip its fab with advanced tools, Micron secured ¥46.5 billion ($320 million) grant from the Japanese government last September. Meanwhile, Micron says it will invest ¥500 billion ($3.618 billion) in the technology ‘over the next few years, with close support from the Japanese government.’

“Micron is the only company that manufactures DRAM in Japan and is critical to setting the pace for not only the global DRAM industry but our developing semiconductor ecosystem,” said Satoshi Nohara, METI Director-General of the Commerce and Information Policy Bureau. “We are pleased to see our collaboration with Micron take root in Hiroshima with state-of-the-art EUV to be introduced on Japanese soil. This will not only deepen and advance the talent and infrastructure of our semiconductor ecosystem, it will also unlock exponential growth and opportunity for our digital economy.”

Source: AnandTech – Micron to Bring EUV to Japan: 1γ Process DRAM to Be Made in Hiroshima in 2025

Samsung on Thursday said it had started high volume production DRAM chips on its latest 12nm fabrication process. The new manufacturing node has allowed Samsung to reduce the power consumption of its DRAM devices, as well as decrease their costs significantly compared to its previous-generation node.

According to Samsung’s announcement, the company’s 12nm fabrication process is being used to produce 16Gbit DDR5 memory chips. And while the company is already producing DDR5 chips with that capacity (e.g. K4RAH086VB-BCQK), the switch to the newer and smaller 12nm process has paid off both in terms of power consumption and die size. As compared to DDR5 dies made on the company’s previous-generation node (14nm), the new 12nm dies offer up to 23% lower power consumption, and Samsung is able to produce 20% more dies per wafer (i.e., the DDR5 dies are tangibly smaller). 

Samsung says that the key innovation of its 12nm DRAM fabrication process is usage of new high-k material for DRAM cell capacitors that enabled it to increase cell’s capacitance to boost performance, but without increasing their dimensions and die sizes. Higher DRAM cell capacitance means a DRAM cell can store more data and reduce power-draining refresh cycles, hence increasing performance. However, larger capacitors typically result in increased cell and die size, which makes the resulting dies more expensive.

DRAM makers have been addressing this by using high-k materials for years, but finding these materials is getting trickier with each new node as memory makers also have to take into account yields and production infrastructure they have. Apparently, Samsung has succeeded in doing so with its 12nm node, though it does not make any disclosures on the matter. That Samsung has succeeded in reducing their die size by a meaningful amount at all is quite remarkable, as analog components like capacitors were some of the first parts of chips to stop scaling down further with finer process nodes.

In addition to introducing a new high-k material, Samsung also reduced operating voltage and noise for its 12nm DDR5 ICs to offer a better balance of performance and power consumption compared to predecessors.

One of the aspects about Samsung’s 12nm DRAM technology is that it looks to be the company’s 3^rd Generation production node for memory that uses extreme ultraviolet lithography. The first D1x node was purely designed as a proof of concept and its successor D1a, which has been in use since 2021, used EUV for five layers. Meanwhile, it is unclear to what degree Samsung’s 12nm node is using EUV tools.

“Using differentiated process technology, Samsung’s industry-leading 12nm-class DDR5 DRAM delivers outstanding performance and power efficiency,” said Jooyoung Lee, Executive Vice President of DRAM Product & Technology at Samsung Electronics. 

Meanwhile, Samsung is also eyeing faster memory speeds with their new 12nm DDR5 dies. According to the company, these dies can run as fast as DDR5-7200 (i.e. 7.2Gbps/pin), which is well ahead of what the official JEDEC specification currently allows for. The voltage required isn’t being stated, but if nothing else, it offers some promise for future XMP/EXPO memory kits.

Source: AnandTech – Samsung Kicks Off DDR5 DRAM Production on 12nm Process Tech, DDR5-7200 in the Works

Solidigm has been extremely bullish on QLC SSDs in the data center. Compared to other flash vendors, their continued use of a floating gate cell architecture (while others moved on to charge trap configurations) has served them well in bringing QLC SSDs to the enterprise market. The company realized early on that the market was hungry for a low-cost high-capacity SSD to drive per-rack capacity. In order to address this using their 144L 3D NAND generation, Solidigm created the D5-P5316. While the lineup did include a 30TB SKU for less than $100/TB, the QLC characteristics in general, and the use of a 16KB indirection unit (IU) resulted in limiting the use-cases to read-heavy and large-sized sequential / random write workloads.

Solidigm markets their data center SSDs under two families – the D7 line is meant for demanding workloads with 3D TLC flash. The D5 series, on the other hand, uses QLC flash and targets mainstream workloads and specialized non-demanding use-cases where density and cost are more important. The company further segments this family into the ‘Essential Endurance’ and ‘Value Endurance’ line. The popular D5-P5316 falls under the ‘Value Endurance’ line.

The D5-P5430 being introduced today is a direct TLC replacement drive in the ‘Essential Endurance’ line. This means that, unlike the D5-P5316’s 16K IU, the D5-P5430 uses a 4KB IU. The company had provided an inkling of this drive in their Tech Field Day presentation last year.

Despite being a QLC SSD, Solidigm is promising very competitive read performance and higher endurance ratings compared to previous generation TLC drives from its competitors. In fact, Solidigm believes that the D5-P5430 can be quite competitive against TLC drives like the Micron 7450 Pro and Kioxia CD6-R.

Based on market positioning, the Micron 6500 ION launched earlier today is the main competition for the D5-P5430. The sequential writes and power consumption numbers are not particularly attractive for the Solidigm drive on a comparative basis, but the D5-P5430 does win out on the endurance aspect – 0.3 RDWPD for the 6500 ION against 0.58 RDWPD for the D5-P5430 (surprising for a QLC drive). Solidigm prefers total NAND writes limit as a better estimtate of endurance and quotes 32 PBW as the endurance rating for the D5-P5430’s maximum capacity SKU. Another key aspect here is that the D5-P5430 is only available in capacities up to 15.36 TB today. The 30 TB SKU is slated to appear later this year. In comparison, the 30 TB SKU for the 6500 ION is available now. On the other hand, the D5-P5430 is available in a range of capacities and form-factors, unlike the 6500 ION. The choice might just end up being dependent on how each SSD performs for the intended use-cases.

Source: AnandTech – Solidigm D5-P5430 Addresses QLC Endurance in Data Center SSDs