November Update: Power, TOP500, the Kaufman Dinner, Fred Brooks, and an Award

Today is the last day in November, amazingly, and since I was on vacation last Friday, today is my regular monthly update covering topics that don't justify a whole post on their own or that are updates to a previously published post.

Kaufman Dinner

I wrote about this year's Kaufman Award in my post, 2022 Kaufman Award Honors Giovanni De Micheli. The 2022 Phil Kaufman Award ceremony and banquet honoring Dr. Giovanni De Micheli will be held Thursday, February 23, 2023, from 6:30 p.m. until 9:30 p.m. at The GlassHouse in San Jose, CA.

There Is Not Enough Power Redux

I wrote about datacenter power in my post There Is Not Enough Power. Some additional interesting data on the topic comes from Doug O'Laughlin of Fabricated Knowledge in his post Hyperscalers and Energy Costs. Unlike me, he listens to lots of companies' earnings calls.

Here's what Microsoft said about datacenter energy costs:

Microsoft cloud gross margin percentage decreased roughly 1 point driven...primarily due to higher energy cost. Therefore, with our first quarter results and lower expected OEM revenue for the remainder of the year as well as over $800 million of greater-than-expected energy cost, we now expect operating margins in U.S. dollars to be down roughly 1 point year-over-year.

So Microsoft has enough power, but it's costing them a lot to get it. Amazon has a similar story:

We're also seeing energy costs that are materially higher than they had in pre-pandemic, electricity and the impact of natural gas pricing. So those prices have up more than 2x over the last couple of years and contribute to about 200 basis point degradation versus 2 years ago.

Of course, chips are getting lower in power (thank you, Cadence Celsius Thermal Solver). Apparently, there is a way that this is measured, being Power Usage Effectiveness, or PUE. This is defined in the obvious way as: 

PUE = Total Building Power / IT Power (servers, GPUs, network equipment, etc.)

Doug had found the chart above from the Uptime Global Data Center Survey (yes, who knew that was a thing?). PUEs are decreasing (good) but not as fast as in the past. It fell really fast from 2007 to 2018 but has basically been flat since. Doug surmises that this is because Denard scaling is over, but Dennard scaling ended earlier than that, in 2003 or so. My suspicion is that it is neural network training, meaning that we have put a lot of very high-powered NVIDIA GPUs into datacenters. And that is not getting better. Each generation of NVIDIA training chips (such as the H100) has a lot more compute power (around 2X) than the previous generation but uses about 2X the electric power too.

Read Doug's whole piece (it's free) for a lot more on the topic.

Fast Company Next Big Things in Tech 2022

Cadence's Cerebrus won an award from Fast Company as one of The Next Big Things in Tech 2022. See the Fast Company page with the announcements.

I wrote about Cerebrus when we launched it in my post Cadence Cerebrus - Intelligent Chip Explorer.

Apple Emergency Services by Satellite

This is a mixture of seeing some really neat technology and a public service announcement. As of mid-November in US/Canada, and France/Germany/Ireland/UK in December, if you have an iPhone 14, you can make emergency calls even if there is no cellular or wifi coverage. It does this by using your phone to communicate with Globalstar satellites. You have to be outside. It doesn't work in buildings.

As Apple said in their announcement:

Some of the most popular places to travel are off the beaten path and simply lack cellular coverage. With Emergency SOS via satellite, the iPhone 14 lineup provides an indispensable tool that can get users the help they need while they are off the grid.
...
Additionally, if users want to reassure friends and family of their whereabouts while traveling off the grid, they can now open the Find My app and share their location via satellite.

Shame I only have an iPhone 13, although I hope never to need the service.

Fred Brooks

I wrote about Fred Brooks and his very influential book in my posts Fred Brooks: "It Is a Humbling Experience to Make a Multi-Million Dollar Mistake" and The Mythical Man-Month. Unfortunately, he passed away on November 17th a couple of weeks ago. I feel privileged to have met him.

TOP500 Supercomputers

It was recently SC22, the big supercomputing conference. This means that the latest results of the TOP500 list are announced. This list, now in its 60th edition (it comes out twice a year), lists the 500 most powerful computers in the world. For some details and previous posts, see Supercomputers and Japanese Arm-Powered Supercomputer Takes the TOP500 Crown. And supercomputers look really cool!

As it says in the press release:

With an HPL score of 1.102 EFlop/s, the Frontier machine at Oak Ridge National Laboratory (ORNL) did not improve upon the score it reached on the June 2022 list. That said, Frontier’s near-tripling of the HPL score received by second-place winner is still a major victory for computer science. On top of that, Frontier demonstrated a score of 7.94 EFlop/s on the HPL-MxP benchmark, which measures performance for mixed-precision calculation. Frontier is based on the HPE Cray EX235a architecture and it relies on AMD EPYC 64C 2GHz processor. The system has 8,730,112 cores and a power efficiency rating of 52.23 gigaflops/watt. It also relies on gigabit ethernet for data transfer. 

The Fugaku system at the Riken Center for Computational Science (R-CCS) in Kobe, Japan, previously held the top spot for two years in a row before being moved down by the Frontier machine. With an HPL score of 0.442 EFlop/s, Fugaku has retained its No. 2 spot from the previous list.

Look at those numbers. Frontier runs at 1 exaflops hand and has nearly 9M AMD cores. Number 2 Fugagku, which is described in more details in my post linked above, has less than half that performance on its 7.5M Arm cores.

Cray-Pi

Here's something fun. The original Cray-1, the first supercomputer, had a processor speed of 80Mhz, 8Mb of memory, and a performance of 160 megaflops. A Raspberry Pi has a 1Ghz processor, 512Mb of memory, and 24 gigaflops of performance, so a Raspberry Pi has more power than the original Cray-1. Actually, 150X more.

And this video gives some of the history of the Cray-1 and 3D prints a case that is a model of the Cray-1 for a single Raspberry Pi (12 minutes).

IEDM, Video Version

IEDM starts next Saturday (tutorials), Sunday (short courses), and Monday (start of the conference proper). Here's the official video.

 

Sign up for Sunday Brunch, the weekly Breakfast Bytes email.

.