Gigabyte, here is its new server with AMD EPYC and NVIDIA A100

Gigabyte, here is its new server with AMD EPYC and NVIDIA A100


Yesterday we told you about Gigabyte's rather unusual move related to the withdrawal from the market of the GeForce RTX 3090 Turbo graphics card, perhaps due to, but it is pure speculation, pressure from NVIDIA, which had observed a certain use of similar products for servers, an area for which the professional line of the Californian company is more suitable, whose most recent exponent is the A100. Today, Gigabyte announced its new server for AI, high-performance computing (HPC) and data analytics, called the G262-ZR0. It is one of the first in the industry with four NVIDIA A100 processing GPUs, accompanied by two AMD EPYC 7002 series 'Rome' processors with up to 64 cores per CPU. Recall that the four NVIDA A100 cards offer 13,824 FP64 CUDA cores, 27,648 FP32 CUDA cores and aggregate performance of 38.8 FP64 TFLOPS and 78 FP32 TFLOPS, as well as a lot of, and lightning fast, HBM2 memory.

The machine can be equipped with up to 16 RDIMM or LRDIMM DDR4-3200 memory modules, three M.2 SSDs with a PCIe 4.0 x4 interface and four 2.5 ”HDDs or SSDs with a SATA or Gen4 interface. U.2. The server also features two GbE ports, six low-profile PCIe Gen4 x16 expansion slots, one OCP 3.0 Gen4 x16 slot, one BMC ASpeed ​​AST2500, and two redundant 3000W 80+ Platinum certified power supplies. Gigabyte claims the G262-ZR0 will provide the highest possible GPU processing performance in a 2U chassis, which is its main competitive advantage.

Sadly, Gigabyte has not disclosed the pricing of its G262-ZR0 server at this time. but it will of course be much cheaper than an NVIDIA DGX A100 system (which sees the presence of eight A100s) or similar. Currently we just have to wait for more information directly from the manufacturer.

On Amazon there is a large assortment of Gigabyte graphics cards, even at a discounted price!

Gigabyte Axes GeForce RTX 3090 TURBO, Disrupting Partner Server Plans

Gigabyte has halted sales of its new GeForce RTX 3090 TURBO graphics card, according to sources familiar with the situation, disrupting plans by partners who were planning to sell the product in servers.

The Taipei City, Taiwan-based hardware manufacturer recently informed channel partners that the “blower edition” graphics card, which was only announced last September, has reached its end of life, prompting backorders to be canceled as a result, multiple sources told CRN. Gigabyte’s web page for the product now redirects to an error page.

[Related: Gigabyte: COVID-19 Surge At L.A. Ports Delaying Shipments ]

Gigabyte did not respond to a request for comment.

One partner, who had multiple backorders for Gigabyte’s GeForce RTX 3090 TURBO, told CRN that he was surprised and puzzled by the news, given that it had only been available for a few months.

“Why did you have to end-of-life a brand-new blower edition? It is very strange,” said the partner, who asked to not be identified because of his company’s relationship with Gigabyte.

While Nvidia has a dedicated lineup of GPU products for servers, which includes last year’s A100, some partners turn to the chipmaker’s GeForce graphics cards for desktop PCs to use in GPU servers. Multiple partners told CRN that GeForce graphics cards are preferred by some budget-conscious customers because they are significantly cheaper than those that are designed for servers.

“They’re budget-constrained, and this is how they can meet some of their needs for compute power,” the partner said of his public sector and university customers.

The challenge is that not every GeForce graphics card has the right form factor to fit inside servers. For instance, Nvidia’s GeForce RTX 3090 Founder’s Edition graphics card is a triple-slot card, which means it can’t fit into most rackmount servers, according to the partner with canceled backorders. But since Nvidia works with add-in board partners, such as Gigabyte, those add-in board partners can provide graphics cards containing Nvidia GPUs with different form factors, specifications and features.

The partner said Gigabyte’s GeForce RTX 3090 TURBO graphics card was a big deal in large part because it was a dual-slot card and could easily fit into standard server as a result. He added that other cards with a “blower” design — which indicates that the card blows hot air through the back grill to cool the GPU — have been popular for servers because of their dual-slot form factor.

With Gigabyte no longer selling the GeForce RTX 3090 TURBO, partners who were previously planning servers around the graphics card are now trying to figure out alternatives for customers, which is causing delivery timelines to be pushed back.

“It’s unfortunate that we can’t deliver these systems in a timely manner because the buyers are public sector, government labs and universities,” the partner said. “That’s who mostly buys this stuff.”

One AI researcher who relies on GPU servers told CRN that Gigabyte’s abrupt end-of-life for the GeForce RTX 3090 TURBO is slowing down research work at his lab because they were in the process of buying servers equipped with Gigabyte’s cards when the news was disclosed.

“It is slowing down our research since we have lots of jobs in queue with our current server infrastructure,” said the researcher, who asked to not be identified because of an affiliation with Nvidia. “Conference deadlines are missed, ablation studies are shorter and certain kinds of research can’t be even tackled. [Deep learning] research is very empirical, so fast experimentation is crucial.”

The researcher’s lab had turned to Gigabyte’s GeForce RTX 3090 TURBO because it offered high performance at a very affordable price in comparison to Nvidia’s V100 and A100 server GPUs, plus the 24 GB memory capacity made it “perfect” for deep learning. But because the Gigabyte partners they were trying to buy from had no alternative server configurations that could meet the lab’s budget constraints, they are now trying to look elsewhere.

Another partner, who had a large server order impacted by the card’s end-of-life, said the price-performance ratios of the GeForce RTX 3090 TURBO were likely to be very favorable in comparison to the more expensive A100, which shares the same underlying Ampere architecture. The partner said they compared a last-generation GeForce 2080 Ti card with the A100 and found that the A100 was only about 20 percent faster while costing several times more.

“There’s just a massive price difference,” the second partner said.

The only other rub with using GeForce graphics cards for servers is that Nvidia doesn’t endorse such use because the company wants customers to buy its cards that are purpose-built for servers, like the A100, according to this partner. Nvidia may market the GeForce RTX 3090 cards for researchers as a key demographic, but they are likely angling those cards for use in workstations, not servers, he added.

But while Nvidia stipulates that GeForce software may not be used for data center purposes in the software’s license agreement, the partner said the line for acceptable use in the field is blurry.

“They don’t want to lose the business, but they want to push everyone towards the A100,” he said.

An Nvidia spokesperson told CRN that GeForce and Titan graphics cards “were never designed for data center deployments with the complex hardware, software, and thermal requirements for 24x7 operation, where there are often multi-stack racks.”

However, while the GeForce software license is meant to discourage “potential misuse” of GeForce and Titan graphics cards in “demanding, large-scale enterprise environments,” Nvidia does recognize the value of such products in servers for “non-commercial uses or other research uses that do not operate at data center scale,” according to the spokesperson.

“Nvidia does not intend to prohibit such uses,” the spokesperson added.

Karl Freund, principal analyst at AI hardware research firm Cambrian-AI Research, told CRN that the sudden end-of-life for Gigabyte’s GeForce RTX 3090 TURBO underlines one of the risks customers and partners face when they decide to use Nvidia desktop PC and workstation products in servers.

GeForce graphics cards may offer high performance at a lower price in servers, according to Freund. But they lack the many benefits that Nvidia provides with its data center GPUs to address what it said are the “unique mechanical, physical, management, functional, reliability and availability needs of servers.”

Those benefits include a “three-year warranty covering data center workloads, Nvidia enterprise support, guaranteed continuity of supply and extended SKU life expectancy for data center components,” Nvidia said, which it has communicated since those products were first released.

That means if a customer running GeForce servers needs to replace a graphics card a year or two after deployment, the specific model used by the customer may not be available anymore because, unlike data center GPUs like the A100 or Quadro, Nvidia does not guarantee a long lifecycle for the product.

“For GeForce, it just costs a lot less money because Nvidia is not having to incur the expenses of inventory exposure, long lead times, accessibility and specifications have been validated through all the partners — that all costs money,” Freund said. “They have to recoup that. So if you need those things, then you got to find Quadro. If you don’t need those things, you can save a ton of money and get the work done essentially for a much lower total cost of ownership.”

The other potential downside with using GeForce graphics cards for servers is that they are not certified for data center applications while products like the A100 or Quadro are, according to Freund.

“If you’re designing an aircraft engine [using GeForce], you want to know that the finite element analysis application has been checked out and [that] they’ve been certified on that particular card,” he said. “And so that certification is something that vendors will often point to as a reason to spend the extra money for a Quadro card.”