Comments on: Nvidia: There’s A New Kid In Datacenter Town

By: Michael Alan Bruzzone

Michael Alan Bruzzone — Mon, 28 Aug 2023 18:56:21 +0000

Correction, A800 (not A100) card cost $3302. mb

By: Michael Alan Bruzzone

Michael Alan Bruzzone — Mon, 28 Aug 2023 18:36:20 +0000

“If the rumors suggested in the Financial Times are true, Nvidia can only make 500,000 “Hopper” H100 GPU accelerators this year (calendar 2023 is almost the same as Nvidia fiscal 2024, offset by one month) and that Nvidia was looking to ship 1.5 million to 2 million units in 2024.”

The financial Times piece continues to perpetuate questionable data.

Questionable data #1 portrays Nvidia + 1000% margin on H100. I have x4.48 over card cost at $3824 for H100 and $3302 for A100. The other half goes to hyperscale and cloud direct customers, OEMs and design build engineering for applications and systems integration cost, training, education, field support and profit. This is how Nvidia prevents a reoccurrence of the 2022 GPU channel thefts on design/build and sales sticking with Nvidia for an equal revenue cut of product value.

Questionable data #2 is quarterly production volume. I have Nvidia DC all up at 489,219 units in q1 and 1,219,449 in q2. This is calculated net into total division revenue for product volume that recoups all costs. In the so said 20 million unit’s annual server market, another industry misnomer, delivering anything less than 5 M accelerators per year is a losing objective over the near term.

I have all DC gross at $17,127 per unit in q1 and $11,279 per unit in q2. Net after R&D, SG&A, tax accrual = $8756.81 in q1 and $8,394.17 in q2.

There are two other techniques. Driving gross take back into gross equals 84,487 units in q1 and 613,509 units in q2. Or driving net into net = 99,359 units in q1 and 170,394 units in q2. I disqualify both methods because nether takes into account actual production volume required to cover all costs; R&D, SG&A and tax accrual.

Nvidia data center production volume in q1 = 489,219 units and q2 = 1,229,424.

The Financial times on low volumes is a disservice to industry, AMD, Intel, Nvidia and especially investors.

Mike Bruzzone, Camp Marketing

By: ~ai

~ai — Sat, 26 Aug 2023 01:43:44 +0000

How much is per query cost for generative AI models? That will define the life of current AI hype. When the dust settles, how much carbon footprint these so called AI Datacenters have built? What’s the actual cost of “toying” around with ChatGPT on global warming? AI hyped cult will cancel you the moment you raise these issues.

By: EC

EC — Fri, 25 Aug 2023 04:22:22 +0000

In reply to Slim Albert. I think it's due to a very early realization single threaded applications were going to be overwhelmed by vast amounts of data and parallel processing was better suited for these types of workloads. But to your point, gaming was just the first application for what Nvidia refers to as accelerated computing. Totally concur on the enthusiasm, seems to be shifting to communities beyond gaming.

By: EC

EC — Fri, 25 Aug 2023 04:15:43 +0000

In reply to Timothy Prickett Morgan. touché

By: Slim Albert

Slim Albert — Thu, 24 Aug 2023 21:55:47 +0000

Nvidia’s success may have something to do with the intersection of its gaming roots and human psychology. It has (historically) been operating in the development, and selling, of tools to intricately render dreamscapes and produce alternate (or virtual) realities, a particularly enjoyable area of activity, where enthusiasm is broadly contagious, and fun-and-awe is the end goal. Rather distant from “serious” gray-hair computing, boring administrative databases, sedated stiffs in suits, and closer to leather-clad boundless youthful energy, “skateboarding”, X-gaming, “revolt”, street cred.

They’ve been outstandingly successful at jointly leveraging this enthusiasm, and the underlying advanced tech (triangle shading, ray-tracing), to meet the emerging opportunities presented by AI/ML — an area perceived also via a mixture of enthusiasm, fun-and-awe, alternative realism, and some degree of apprehension and worry. It’s a great match, and well-done! I’m not too sure though that IBM’2017 (or even ’23) could have absorbed that culture without damage to either, or both.

By: Timothy Prickett Morgan

Timothy Prickett Morgan — Thu, 24 Aug 2023 17:02:36 +0000

In reply to Carl Schumacher. Maybe an Extra Small Modular Nuclear Reactor? We don't need 300 megawatts yet. Maybe 30 megawatts. So a coupon for a 10XS SMR? HA! What the hell, 10 percent off?

By: Carl Schumacher

Carl Schumacher — Thu, 24 Aug 2023 15:44:30 +0000

“If you want to train a GPT5-level model, you apparently need somewhere on the order of 20,000 to 25,000 H100 GPU accelerators”…Perhaps if you buy in said quantity, they should ship with a 5% off coupon for an SMR.

By: Timothy Prickett Morgan

Timothy Prickett Morgan — Thu, 24 Aug 2023 10:51:56 +0000

In reply to EC. Come on, man. The Eagles? From Hotel California, the epic album of American excess and optimism? "Everybody loves you/ So don’t let them down."

By: EC

EC — Thu, 24 Aug 2023 07:52:42 +0000

Great article TPM, appreciate the amusements — mine and yours both, but don’t you think new *Sheriff* is more like it?