Under The Wire: Nearly HPC News (June 13, 2024)

By Doug Eadline

June 13, 2024

As managing editor of the major global HPC news source, the term “news fire hose” is often mentioned. The analogy is quite correct. In any given week, there are many interesting stories, and only a few ever become headlines on HPCwire. We present “Off The Wire,” which delivers real-time curated press releases in addition to more in-depth news and technology stories.

Even with this level of coverage, I run across many news items that are “not quite a full story” or “there is no official press release.” In an effort to bring some, but not all, of these HPC and AI news tidbits to your attention, I have begun this bi-weekly (or so) column entitled “Under The Wire: Nearly HPC News,” where I devote a few paragraphs and a link to some research paper, project news, or anything else that finds its way to my desk. I expect to cover three to five topics in each installment to keep it readable.

Fair warning. These news snippets are what I find interesting and may, at times, seem “out in left field” (apologies for the US baseball analogy; substitute “odd” or “strange” if it helps). As a little bit of a guide, I do hunt around open source and commodity hardware news and projects because one never knows how these little projects turn out. Many years ago, I learned of someone doing something weird by using commodity video cards to run BLAST sequence alignments. Using GPUs as accelerators, go figure.

This week’s menu includes GPU power facts, LLM fails, the RISC-V Panic, and ill-advised Sierra Forest benchmarks. Let’s get started.

Fun GPU Facts

We recently ran an article on the number of GPUs sold in 2023. Obviously, Nvidia leads the pack, with 3.76 million data-center GPUs. Throw in AMD and Intel, and the needle moves slightly higher to 3.85 million GPUs. When I see big GPU shipment numbers, I first think of cost, but then I start considering power.

Let’s do some power-a-home equivalent math and assume 700W capacity per GPU (this number will be getting bigger next year), which means you need a capacity of 2.7 GWatts (that “G” means 10E9) to power all of them or the equivalent capacity needed to power 2.2 million homes (1.2 KW/home average per day).

The actual number for the GPU is higher because I did not include the rest of the hardware support for the GPUs, but it is reasonable to assume that a modern GPU’s power appetite is the equivalent of powering a home in the US.

Rule of thumb: Every time a data center GPU fires up, it will use the equivalent power of a home in the US.

Expect another 4-5 million houses, sorry, GPUs, next year. Note the estimate did not include those gammers or HPC wonks (cough, cough) that run GPUs in their home, the CPUs behind the GPU, the fans blowing air on the CPUs and GPUs, or the requisite coffee pot.

This result leads us to the “how we are going to power these things department.” It seems AWS has acquired Talen’s Energy nuclear data center campus in Pennsylvania. The deal is $650 million with plans for a 960MW campus, which is good for about 800K new GPU homes (assuming it all goes to GPUs, it won’t).

Go Ask Alice

Part of modern Sci-Tech journalism’s fun is pointing out how spectacularly an LLM can fail. We need look no further than a Computex presentation by AMD CEO Lisa Su demonstrating OpenAI’s Wanderlust, a travel assistant built on GPT-4. As reported by PCGamer, good old Chatty GPT4 got the location of the Computex show completely wrong. Evidently, it believed it was at Changan Junior High School on the other side of town.

Who has not made an embarrassing mistake in front of a large crowd at a large international show? Let’s not be too hard on those intelligent programs that are going to take over the world at some point because they can pass the bar exam. We can, however, keep embarrassing them by asking some basic questions about family relationships. Take, for example, the “Alice in Wonderland problem” (AIW) that seems to break the best and the brightest LLMs.

The AIW problem is quite simple:

Alice has N brothers, and she also has M sisters. How many sisters does Alice’s brother have?

Most humans can figure this out (if you can’t, go sit in the corner with the LLMs).

Turns out not many LLMs can get past this question consistently. The best performance was GPT-4o, which achieved a 65% correct response rate. The results are described in a paper, “Alice in Wonderland: Simple Tasks Showing Complete Reasoning Breakdown in State-Of-the-Art Large Language Models.” Remember, LLMs provide “an answer,” not “the answer,” based on probabilistic token completion.  

The paper also presents the LLMs with a more challenging puzzle called AIW+:

Alice has 3 sisters. Her mother has 1 sister who does not have children – she has 7 nephews and nieces and also 2 brothers.

Alice’s father has a brother who has 5 nephews and nieces in total, and who has also 1 son. How many cousins does Alice’s sister have?

A more difficult puzzle that requires taking different paternal sides, that of mother and father, ignoring some data, carefully calculating the number of cousins, taking care of subtracting Alice and her sister, and summing up the total number of cousins from both sides, for instance:

  1. On the mother’s side: 7 (total nephews and nieces) – 4 (Alice and her sisters) = 3 cousins;
  2. On the father’s side: 5 (total nephews and nieces) + 1 (own son of the father’s brother) – 4 (Alice and her sisters) = 2 cousins; summing up 3 + 2 = 5 cousins which Alice and any of her sisters have.

How did the LLMs do with this more challenging puzzle? No LLM got the correct answer more than 4% of the time, which, like many humans, was probably a good guess. The results are summarized in the figure below.

As we welcome our new LLM’s overloads with swanky law degrees, let’s take stock of the fact that at least we know where we live and how many cousins we have.

Collapse of most state-of-the-art LLMs on the AIW and AIW+ (inlay) problem. (Source: https://arxiv.org/pdf/2406.02061)

RISC-V Panic

A recent headline in the South China Morning Post (SCMP), “Scientists find security risk in RISC-V open-source chip architecture that China hopes can help sidestep US sanctions.”

Gulp. Not RISC-V, the future of sovereign computing for the EU and China (and everywhere else that wants to make its own CPUs). These types of CPU flaws are reserved for the big boys and get really cool names like Spectre and Meltdown.

Admittedly, the SCMP was light on details and, probably due to poor translation, seemed to indicate that the flaw can somehow help China overcome US sanctions on an open-source processor API. A bit more searching found this paper, “The first domestic independently discovered medium-risk vulnerability research in RISC-V processor design” (use Google Translate).

Combing through the paper we get some good techno babel about the flaw.

“... multiplication, and division units use the same write port… port contention … the division instruction may be delayed by the younger multiplication or addition instructions, … timing side channel …

The next question is, “Is this a design flaw or an implementation flaw?”

It turns out the answer is neither, and the RISC-V designers did not introduce a flaw. Some further searching revealed the following from YCombiner:

Assuming they really found a problem, it’s in just one RISC-V core design, a student project, which is not used commercially — at least directly.”

Deep breath. RISC-V panic is over. Processor sovereignty is safe, and vetting sensational web headlines and articles before you write a full story is still a good idea, that is, of course, unless you are scraping content for your LLM.

Don’t Do This with Xeon Sierra Forest E-Cores

Recently, a story by one of the most useful sites on the internet, Phoronix, tested the new Intel Xeon 6 Sierra Forest processors. These new processors are composed entirely of the Intel E-cores (Efficient cores, the Granite Rapids Performance cores (P-cores) arriving next).

Phoronix tested the new 144-core server processors, the Xeon 6766E and the Xeon 6780E. Both are 144-core processors, but the 6766E is a 250-watt part, while the 6780E, the flagship of the 6700E series, is rated for a 330-watt TDP.

Phoronix mentions the Intel proviso stating that the Xeon 6 E-core processors are not designed for some workloads including HPC, so results will not be as expected. But Phoronix was curious and decided to run NAMD on the new Sierra Forest processors. As expected, the performance was much slower than the past generation Ice Lake/Sapphire Rapids/Emerald Rapids processors due to the lack of AVX-512 support.

As Phoronix points out, however, there’s a big boost in performance per Watt with Sierra Forest compared to prior Intel Xeon CPUs. As indicated in the power usage figure below, Sierra Forrest is now in the 140 W range along with the AMD Epyc gang and not in the 560 Watt range of their Max brothers. Good news for Intel: the performance and power numbers should continue to look better when they get their AVX-512/AVX10 mojo back in Granite Rapids later this year. Refer to the full story for more details.

Phoronix power profile for Intel Sierra Forest processors 6780E/6766E running NAMD. (Source: https://www.phoronix.com)

 


Header Graphic: North East Line Tunnels No machine-readable author provided. Calvin Teo assumed (based on copyright claims). CC BY-SA 2.5, via Wikimedia Commons

Subscribe to HPCwire's Weekly Update!

Be the most informed person in the room! Stay ahead of the tech trends with industry updates delivered to you every week!

ARM, Fujitsu Targeting Open-source Software for Power Efficiency in 2-nm Chip

July 19, 2024

Fujitsu and ARM are relying on open-source software to bring power efficiency to an air-cooled supercomputing chip that will ship in 2027. Monaka chip, which will be made using the 2-nanometer process, is based on the Read more…

SCALEing the CUDA Castle

July 18, 2024

In a previous article, HPCwire has reported on a way in which AMD can get across the CUDA moat that protects the Nvidia CUDA castle (at least for PyTorch AI projects.). Other tools have joined the CUDA castle siege. AMD Read more…

Quantum Watchers – Terrific Interview with Caltech’s John Preskill by CERN

July 17, 2024

In case you missed it, there's a fascinating interview with John Preskill, the prominent Caltech physicist and pioneering quantum computing researcher that was recently posted by CERN’s department of experimental physi Read more…

Aurora AI-Driven Atmosphere Model is 5,000x Faster Than Traditional Systems

July 16, 2024

While the onset of human-driven climate change brings with it many horrors, the increase in the frequency and strength of storms poses an enormous threat to communities across the globe. As climate change is warming ocea Read more…

Researchers Say Memory Bandwidth and NVLink Speeds in Hopper Not So Simple

July 15, 2024

Researchers measured the real-world bandwidth of Nvidia's Grace Hopper superchip, with the chip-to-chip interconnect results falling well short of theoretical claims. A paper published on July 10 by researchers in the U. Read more…

Belt-Tightening in Store for Most Federal FY25 Science Budets

July 15, 2024

If it’s summer, it’s federal budgeting time, not to mention an election year as well. There’s an excellent summary of the curent state of FY25 efforts reported in AIP’s policy FYI: Science Policy News. Belt-tight Read more…

SCALEing the CUDA Castle

July 18, 2024

In a previous article, HPCwire has reported on a way in which AMD can get across the CUDA moat that protects the Nvidia CUDA castle (at least for PyTorch AI pro Read more…

Aurora AI-Driven Atmosphere Model is 5,000x Faster Than Traditional Systems

July 16, 2024

While the onset of human-driven climate change brings with it many horrors, the increase in the frequency and strength of storms poses an enormous threat to com Read more…

Shutterstock 1886124835

Researchers Say Memory Bandwidth and NVLink Speeds in Hopper Not So Simple

July 15, 2024

Researchers measured the real-world bandwidth of Nvidia's Grace Hopper superchip, with the chip-to-chip interconnect results falling well short of theoretical c Read more…

Shutterstock 2203611339

NSF Issues Next Solicitation and More Detail on National Quantum Virtual Laboratory

July 10, 2024

After percolating for roughly a year, NSF has issued the next solicitation for the National Quantum Virtual Lab program — this one focused on design and imple Read more…

NCSA’s SEAS Team Keeps APACE of AlphaFold2

July 9, 2024

High-performance computing (HPC) can often be challenging for researchers to use because it requires expertise in working with large datasets, scaling the softw Read more…

Anders Jensen on Europe’s Plan for AI-optimized Supercomputers, Welcoming the UK, and More

July 8, 2024

The recent ISC24 conference in Hamburg showcased LUMI and other leadership-class supercomputers co-funded by the EuroHPC Joint Undertaking (JU), including three Read more…

Generative AI to Account for 1.5% of World’s Power Consumption by 2029

July 8, 2024

Generative AI will take on a larger chunk of the world's power consumption to keep up with the hefty hardware requirements to run applications. "AI chips repres Read more…

US Senators Propose $32 Billion in Annual AI Spending, but Critics Remain Unconvinced

July 5, 2024

Senate leader, Chuck Schumer, and three colleagues want the US government to spend at least $32 billion annually by 2026 for non-defense related AI systems.  T Read more…

Atos Outlines Plans to Get Acquired, and a Path Forward

May 21, 2024

Atos – via its subsidiary Eviden – is the second major supercomputer maker outside of HPE, while others have largely dropped out. The lack of integrators and Atos' financial turmoil have the HPC market worried. If Atos goes under, HPE will be the only major option for building large-scale systems. Read more…

Everyone Except Nvidia Forms Ultra Accelerator Link (UALink) Consortium

May 30, 2024

Consider the GPU. An island of SIMD greatness that makes light work of matrix math. Originally designed to rapidly paint dots on a computer monitor, it was then Read more…

Comparing NVIDIA A100 and NVIDIA L40S: Which GPU is Ideal for AI and Graphics-Intensive Workloads?

October 30, 2023

With long lead times for the NVIDIA H100 and A100 GPUs, many organizations are looking at the new NVIDIA L40S GPU, which it’s a new GPU optimized for AI and g Read more…

Shutterstock_1687123447

Nvidia Economics: Make $5-$7 for Every $1 Spent on GPUs

June 30, 2024

Nvidia is saying that companies could make $5 to $7 for every $1 invested in GPUs over a four-year period. Customers are investing billions in new Nvidia hardwa Read more…

Nvidia Shipped 3.76 Million Data-center GPUs in 2023, According to Study

June 10, 2024

Nvidia had an explosive 2023 in data-center GPU shipments, which totaled roughly 3.76 million units, according to a study conducted by semiconductor analyst fir Read more…

AMD Clears Up Messy GPU Roadmap, Upgrades Chips Annually

June 3, 2024

In the world of AI, there's a desperate search for an alternative to Nvidia's GPUs, and AMD is stepping up to the plate. AMD detailed its updated GPU roadmap, w Read more…

Some Reasons Why Aurora Didn’t Take First Place in the Top500 List

May 15, 2024

The makers of the Aurora supercomputer, which is housed at the Argonne National Laboratory, gave some reasons why the system didn't make the top spot on the Top Read more…

Intel’s Next-gen Falcon Shores Coming Out in Late 2025 

April 30, 2024

It's a long wait for customers hanging on for Intel's next-generation GPU, Falcon Shores, which will be released in late 2025.  "Then we have a rich, a very Read more…

Leading Solution Providers

Contributors

Google Announces Sixth-generation AI Chip, a TPU Called Trillium

May 17, 2024

On Tuesday May 14th, Google announced its sixth-generation TPU (tensor processing unit) called Trillium.  The chip, essentially a TPU v6, is the company's l Read more…

Nvidia H100: Are 550,000 GPUs Enough for This Year?

August 17, 2023

The GPU Squeeze continues to place a premium on Nvidia H100 GPUs. In a recent Financial Times article, Nvidia reports that it expects to ship 550,000 of its lat Read more…

IonQ Plots Path to Commercial (Quantum) Advantage

July 2, 2024

IonQ, the trapped ion quantum computing specialist, delivered a progress report last week firming up 2024/25 product goals and reviewing its technology roadmap. Read more…

Choosing the Right GPU for LLM Inference and Training

December 11, 2023

Accelerating the training and inference processes of deep learning models is crucial for unleashing their true potential and NVIDIA GPUs have emerged as a game- Read more…

The NASA Black Hole Plunge

May 7, 2024

We have all thought about it. No one has done it, but now, thanks to HPC, we see what it looks like. Hold on to your feet because NASA has released videos of wh Read more…

Nvidia’s New Blackwell GPU Can Train AI Models with Trillions of Parameters

March 18, 2024

Nvidia's latest and fastest GPU, codenamed Blackwell, is here and will underpin the company's AI plans this year. The chip offers performance improvements from Read more…

Q&A with Nvidia’s Chief of DGX Systems on the DGX-GB200 Rack-scale System

March 27, 2024

Pictures of Nvidia's new flagship mega-server, the DGX GB200, on the GTC show floor got favorable reactions on social media for the sheer amount of computing po Read more…

MLPerf Inference 4.0 Results Showcase GenAI; Nvidia Still Dominates

March 28, 2024

There were no startling surprises in the latest MLPerf Inference benchmark (4.0) results released yesterday. Two new workloads — Llama 2 and Stable Diffusion Read more…

  • arrow
  • Click Here for More Headlines
  • arrow
HPCwire