Items tagged with hbm2
Earlier this morning, Micron let it slip that NVIDIA's upcoming GeForce RTX 3090 Ampere graphics cards will ship with 12GB of GDDR6X memory. According to Micron's documentation, the GeForce RTX 3090 will be capable of achieving 1TB/sec of memory bandwidth. However, that's not all that Micron had in store via its Tech Brief, which is inexplicably still hosted on its server. Also mentioned is HBMnext, and Micron says that it is "fully involved in the ongoing JEDEC standardization" for the memory. Unfortunately, the company is rather light on details, only adding that it will "enter the market toward the end of 2022". It's not clear if HBMnext is a complete rethink of Micron's...
Read more...
It looks as though AMD might be readying a new Radeon Instinct accelerator, one that would finally feature an Arcturus GPU. The latest leak is essentially a continuation of a rumor we have been hearing about since 2018, that being the impending launch of the Radeon Instinct MI100 accelerator with an Arcturus GPU inside and HBM2 memory. The Radeon Instinct family of accelerators takes aim at machine intelligence and deep learning workloads in the high performance computing (HPC) segment—it's not a graphics card for gaming, nor is it a professional GPU for content creation, in other words. AMD's current Radeon Instinct product consists of the following: Radeon Instinct MI60: 7nm Vega, 4,096...
Read more...
Samsung today said it has developed the industry's first 12-layer 3D-TSV (Through Silicon Via) chip packaging technology, calling it "one of the most challenging packaging technologies for mass production of high performance chips." And that is precisely what it will be used for—it will enable Samsung to mass produce 24GB high bandwidth memory (HBM) sooner than later. What makes this chip packaging technology so challenging is the pinpoint accuracy it requires in order to vertically attach a dozen DRAM chips in a tiny three-dimensional configuration. These configurations consist of 60,000 TSV holes, each of which is just one-twentieth the width of a single strand of human hair. "The thickness...
Read more...
Not only has Samsung ramped up production of its sought-after 8GB HBM2 chip, today the company has announced Aquabolt, its significantly faster 2nd generation HBM2. Aquabolt boasts the fastest data transmissions speeds, delivering 2.4Gbps data transfer per pin at 1.2V. Today’s announcement marks a leap forward in HBM2, a field in which Samsung was already dominating. The company’s 1st generation 8GB HBM2 package reached speeds of 1.6Gbps pin speed at 1.2V, with a 256GB/s data bandwidth. The new 2nd generation Aquabolt delivers a 2.4Gbps pin speed at 1.2V with a 307GB/s bandwidth -- that 50% performance boost is the big takeaway here. By comparison, Aquabolt can achieve data...
Read more...
Samsung is currently the only memory chip maker that is producing 8-gigabyte (GB) HBM2 (second generation high bandwidth memory) chips. In order to keep up with demand from a growing field of customers and applications, Samsung is ramping up production of its 8GB HBM2 chips, which will find their way into systems tasked with artificial intelligence chores, HPC (high performance computing), advanced graphics, network systems, and enterprise servers. Both AMD and NVIDIA use HBM2 chips in some of their high-end graphics card solutions as well. With NVIDIA, HBM2 can be found on its Quadro GP100 card for professionals, which wields 16GB of HBM2. And for AMD, its recently announced Radeon RX Vega Frontier...
Read more...
AMD’s Vega graphics architecture is fast approaching a retail release, but unfortunately for gamers, the first available cards will be the professional workstation-targeted Radeon Vega Frontier Edition. If you’re in the market for such a card, preorders are now open from at least two retailers — one based in the U.S., with the other hailing from the UK. SaberPC has an air-cooled Radeon Vega Frontier Edition (100-506061) and the liquid-cooled version (100-506062) available to preorder. Both graphics cards include 16GB of HBM2 memory, 483GB/sec of memory bandwidth, 4096 stream processors 13.1 TFLOPS single precision compute performance and 26.2 TFLOPS double precision compute performance. TDP for...
Read more...
Earlier this morning, we learned that AMD Radeon RX Vega graphics cards will be announced May 31st at Computex, with general availability occurring on June 5th. It was also revealed that the will be three SKUs at launch: Core, Eclipse and Nova. According to the latest leaked information, at least one member of the Vega family will come with a 1600MHz core clock and a hefty 16GB HBM2 memory. According to the source, it is not known if this particular card is a member of the Radeon RX Vega, Radeon Pro or Radeon Pro Duo series, but we do know that is based on GFX900 (Vega) architecture. Given that most of the previous leaked specs on Vega that we have seen point to 8GB of HBM2, this could very well...
Read more...
AMD is on the cusp of releasing its hotly anticipated Radeon RX Vega graphics card lineup, which will be debut sometime this quarter. If all goes well for AMD, its Vega release will once again make it competitive with NVIDIA in the high-end GPU segment just as Ryzen made the company competitive with Intel in the performance CPU market. Until then, there are several questions that remain unanswered. One of them is whether Vega will largely be a paper release, and word on the web is that it could be due to supply constraints of second generation high bandwidth memory (HBM2).The folks at TweakTown seem to think they have it on good authority that AMD's Radeon RX Vega launch will consist of "only...
Read more...
It's taken a few years, but Intel has finally started sampling to customers its Stratix 10, the industry's first 14-nanometer field programmable array (FPGA). This is something chip designer Altera tapped Intel to do back in 2013, though it hit delays along the way. Nevertheless, when Intel acquired Altera last year for $16.7 billion, it assured clients and investors that Altera's ARM-based chips would still be developed going forward. Well, here we are. Stratix 10 meshes Intel's 14nm tri-gate process technology with a new architecture called HyperFlex. It features an embedded quad-core 64-bit ARM processor (Cortex-A53), 5.5-million logic elements, and second generation High Bandwidth Memory...
Read more...
We saw first generation High Bandwidth Memory (HBM) enter the enthusiast arena with the arrival of AMD’s Radeon R9 Fury lineup of graphics cards. HBM promised incredible performance with equally impressive low power consumption, and largely delivered on those claims. NVIDIA introduced us to HBM2 with its Tesla P100 accelerator, which includes up to 16GB of the good stuff onboard. While HBM2 isn’t even shipping in volume for mainstream (or even enthusiast-class) applications, that isn’t stopping Samsung from laying out its roadmap for HBM3. HBM3 is expected to hit production in the 2019 to 2020 time frame and take the “bigger, faster, stronger” approach compared to HBM2. Each die will now be capable...
Read more...
NVIDIA just pulled the wraps off its fastest GPU and compute engine yet: the Tesla P100. This new GPU powerhouse is based on NVIDIA’s next generation Pascal architecture, which means that it’s built on a 16nm FinFET process. But even with thoroughly modern FinFET process tech, the Tesla P100’s die size measures 600mm2 due in part to the 15.3 billion transistors that comprise the new compute engine — nearly twice that of Maxwell, for reference. The specifications of the Tesla P100 are second to none, as double-precision performance is rated at 5.3 teraflops (TFLOPs), while single-precision performance comes at 10.6 TFLOPs. And as you can probably surmise from those figures, half-precision performance...
Read more...
The big news yesterday from the AMD camp was with the official announcement of the Radeon Pro Duo, which was previously known as the Radeon R9 Fury X2. The water-cooled graphics card features two R9 Fury GPU cores clocked at 1GHz all operating within a 350W power envelope. But of course, we’re always looking to see what’s right around the corner, and in AMD’s case, that means Polaris. Polaris is destined to arrive about mid-way through 2016 and will move from a 28nm to a 14nm FinFET (Global Foundries) manufacturing processes. The move to FinFET allows for a radical decrease in power consumption with a new roadmap that was shared with us yesterday. So just how efficient is Polaris? If we use AMD’s...
Read more...
Most of us were introduced to High Bandwidth Memory (HBM) DRAM courtesy of AMD’s Fury family of graphics cards, each of which sports 4GB of HBM. HBM allows Fury GPUs to tout an impressive 512GB/sec of memory bandwidth, but perhaps that just isn’t enough to fuel the next generation of graphics cards in the pipeline.AMD's Fiji GPU With First Generation Stacked HBM Samsung is looking to make first generation HBM look decidedly old hat, and has announced that it has begun mass production of HBM2. Samsung’s 4GB HBM2 package is built on a 20 nanometer process (each package contains four 8-gigabit core dies built on top of a buffer die). Each 4GB package is capable of delivering 256GB/sec of bandwidth,...
Read more...
Although NVIDIA revealed some high level information on its next generation Pascal GPU architecture earlier this year, the company revealed some more interesting details this week at NVIDIA’s GPU Technology Conference in Japan. NVIDIA announced that Pascal will have close to a 2x performance-per-watt improvement over the current Maxwell GPU architecture, which is built on a 28nm process. It’s also been revealed that Pascal will be produced using TSMC’s 16nm FinFET process, which will put it right up against AMD which is moving to use 14nm/16nm with two new Graphics Core Next (GCN) products in 2016. Other improvements including replacing PLX PCIe Gen3 bridging with NVLink, which enables bi-directional...
Read more...
