This item was Donated by Michael Thomson November 2024.
Table of Contents
Company and Product Blurb. 1
CHA Processor Highlights. 1
Centaur Technology History Highlights. 2
Donated Centaur Computer: Packing Slip & Notes. 5
System Setup and Power-on. 6
OS Installation Quirks. 8
Ubuntu Linux. 8
Windows 11. 9
Company and Product Blurb
Centaur Technology is the x86 company you’ve never heard of. Yet Centaur existed and shipped millions of units through top-tier OEMs from 1995 until 2021. That’s just shy of three decades!
Centaur Technology’s CHA processor showcased here was a significant achievement for the company in 2019, not just because of the processor’s capabilities, but also because of the company’s unusual history that led to this point. It would be an understatement to say that Centaur was an overachieving underdog against the heavyweight competitors like Intel and AMD.
Before I get into the company’s fun and interesting history, here are some brief details about the CHA processor. First silicon was received in the bringup / debug lab in mid-2019. At a high level, the CHA processor is made up of eight 64-bit, server-class x86 “CNS” cores, one 4,096-byte wide “Ncore” machine learning accelerator, and all the glue and typical “south bridge” interface logic integrated into a single silicon SoC. The CHA processor represents several notable achievements:
· From-scratch design, everything in it was new.
· First ever x86 CPU with an integrated ML accelerator (it beat Intel and AMD to the punch!)
· Played a role in the initial development and direction of the very first machine learning inference benchmark, MLPerf, that was fully embraced by the industry.
· Beat all other competitors in its MLPerf inference benchmark latency (including the heavyweight competitors you’re familiar with like NVIDIA and Intel).
· To put the “Ncore” machine learning accelerator’s 4,096-byte wide SIMD execution engine into perspective, it equates to a width of 32,768 bits (i.e., 4,096 bytes x 8 bits per byte). For those familiar with AVX-512 (which is 512 bits wide and currently the widest available SIMD engine in the x86 world), this accelerator is 64x wider than that. The company’s founder, Glenn Henry, often enjoyed saying that this accelerator is like having AVX-32,768.
· Centaur accomplished all of the above with only 110 employees based in Austin, Texas. Only 8 of those employees worked on the ML accelerator. For comparison, do you know how many employees AMD or Intel had at that same time?
o Intel had 110,000 employees, that is 1,000x the size of Centaur!
o AMD had 11,000 employees, that is 100x the size of Centaur!
o Centaur did the same end-to-end engineering work (architecture, RTL, chip design / place and route, verification, motherboard design, BIOS, drivers, software, etc.) as these other competitors, but with only 0.1% to 1.0% of the manpower. That’s incredible!
· Sadly, CHA also represents the last non-AMD and non-Intel x86 processor.
· Because the company was dissolved while lab debug was ongoing and before CHA began to ship, it is likely one of the rarest x86 processors in the world. Of the few specimens that exist, most of them accidentally became available to the public when Centaur’s parent company VIA Technologies inadvertently sold them during an auction where they were offloading all of Centaur’s equipment after shuttering the company. This unit is one of those limited specimens.
While the company existed, it could be characterized as both demanding and endearing. For a company of 110 people to do the same work as competitors 100 to 1,000 times its size, all of Centaur’s employees had diverse, wide-reaching responsibilities. It was common for employees to be engaged in all aspects of the engineering process, from architectural definition, to logic design and verification, to writing software and device drivers. Furthermore, since all employees wore many hats / roles, every team member continually became more valuable since they had to constantly learn and excel at new things. Rather than directly tasking employees to pick up these many hats of responsibility, the employees themselves were expected to discover what was needed and then just do it.
The company’s founder and President, Glenn Henry, would humorously describe this required self-driven mentality to everyone through an analogy about not leaving dog poo sitting in the hallway, but cleaning up the mess the moment it was encountered.
The same metaphor can also be used to explain why the company was originally founded. Back in the 1990’s, Glenn Henry (who had been an IBM Fellow and at that time was a Dell VP) saw what could be metaphorically described as a big dog poo steaming in the computer industry’s hallway. All processor manufacturers of the day were fanatically running to reach peak frequency and performance, but in doing so they were ignoring a large and unattended need that affected most consumers. Namely, no one was servicing the low-power, low-cost market. Glenn’s vision was to start a company that would produce cheaper CPUs for the masses that did not need ultra high performance.
Glenn pitched his idea to potential investors, asking for millions of dollars and to be left alone, with the promise of these low-cost and low-power x86 processors in return, within a short timeframe. The company IDT signed on, and with that, the company was founded by Glenn Henry, Terry Parks, Darius Gaskins, and Al Sato. Their first advertisement seeking new employees was taken out in the local newspaper in April 1995, and a copy of which is shown below.
The ragtag team grew, and Centaur Technology shipped several CPUs under the name WinChip in the 1990’s. There were four x86 companies at the time, Intel, AMD, Cyrix, and Centaur Technology (WinChip). Cyrix was taken out of the picture when VIA Technologies bought both Cyrix and Centaur Technology. At the time, Cyrix had nearly 10x more employees than Centaur but was producing processors at a slower pace, so VIA opted to keep the Centaur team intact and even use Centaur’s design in the final Cyrix CPU.
Another tenet of Centaur Technology’s culture was Glenn Henry’s assertion that if you treat your employees well, then they will work hard and want to stick around. In the company’s 26-year history, a large portion of the employees were the original company veterans from the 1990’s. A few of the employee-centric benefits included full-offices (no cubicles) for every employee, and lunch that was ordered daily from nearby restaurants at no cost to the employees (Centaur would pick up the bill). Additionally, freedom to explore and invent was encouraged. It was common for individuals’ pet projects to regularly and freely end up in silicon, as long as “chicken bits” were included to fuse off the logic in case it was defective. The machine learning accelerator began as one of these pet projects, which ultimately became fully-functional in the first silicon sample and recognized as the fastest entry in that year’s MLPerf inference benchmark latency results.
The development of that “Ncore” machine learning accelerator is worthy of a short summary, as it is a microcosm that reinforces the unique and efficient approach to engineering that company as a whole embraced. As previously mentioned, the entire Ncore team consisted of only 8 full-time employees that covered all aspects of the design: architectural definition, simulator creation, RTL, chip design / place & route, emulation, verification, device drivers, runtime API, custom compiler passes, performance optimization, and software integration into existing ML frameworks.
Ncore (within the CHA SoC) progressed through a relatively lightning-fast and successful timeline, from when it arrived in the lab as part of the first CHA silicon and all the way through official recognition in benchmarks and prestigious academic journals:
· May 2019: Received first taped-out silicon in our office / debug labs.
· A few days later: Fully booting system and our AI accelerator running image classification.
· October 2019: Submitted our best-in-class CPU + accelerator with full-stack software support in the competitive, industry-adopted machine learning inference benchmark MLPerf (MLPerf v0.5, see mlcommons.org).
· November 2019: Integration of our parts into third-party video analytics prototypes.
· January 2020: Wrote paper which was accepted in the industry’s premier academic journal, ISCA (International Symposium on Computer Architecture, 2020). Was also recognized in the same conference in MLPerf’s inference benchmarking introductory paper.
Centaur Technology came to an end in 2021 in what has been recognized as one of the industry’s stranger deals. Intel made a deal with Centaur’s parent company, VIA Technologies, to purchase all of Centaur’s engineers but not the company itself.
I tested that the system was functional and boots prior to shipping.
Package contents:
· Centaur motherboard
· Centaur CPU (already installed in motherboard)
· 2 memory DIMMs (already installed in motherboard in known working configuration)
· Heat sink + fan (please install, tightening screws incrementally in ‘X’ pattern)
· Centaur aluminum backplate (for open-air motherboard mounting)
· 500W power supply (same we used in Centaur debug lab)
o With power cable
· Small HDMI display (same we used in Centaur debug lab)
o With short HDMI cable
You need to provide your own:
· Graphics card
· Hard drive (either SATA or NVMe)
· Keyboard
· Mouse
· Micro-USB cord to power the small HDMI display
· Operating System (tested to work with 64-bit Ubuntu 22.04 and Windows 10 & 11)
Installation / Modules
· Centaur Motherboard: the motherboard, just like the CPU, was designed in-house by the folks at the Austin, TX facility. Since this is a lab bring-up version of the motherboard, it includes some interesting non-standard features:
o Physical power and reset buttons on the bottom-right corner (in addition to the typical header for PC case power and reset buttons).
o USB Type-C connector on the bottom edge of the motherboard.
o A header labeled “LPC Bus”, which when our custom 7-segment display is connected, it will flash the POST codes as the BIOS is booting (useful for debug, since if the system hangs, the last POST code is still shown).
o JTAG header.
o Various DIP switches and jumpers (do not change these).
o On-board serial port.
The motherboard includes 2 required power connectors: the 24-pin main power connector, and the 8-pin CPU power connector.
· Centaur CPU: I am shipping with the CPU pre-installed. The spring-action socket pins (like used with Intel and AMD) are easily damaged when they are not covered, so the CPU protects them during shipping. Some versions of the CPU have the Centaur Technology logo on them (there are two version of the etched logo), and some that are blank without a logo. The one shipped to you is one without the logo, since that is what I had left from the newer steppings.
· Aluminum backplate: I am shipping one of the custom aluminum slabs (pictured) that was used in Centaur’s debug lab for these systems. The threaded holes line up with motherboard installation holes. I have shipped the backplate with the threaded inserts already installed.
· CPU cooler: install the supplied heatsink (with thermal paste) and plug in the fan to the 4-pin connector labeled “CPU FAN” directly above the CPU and memory DIMM slots.
· BIOS: IMPORTANT! I do not recommend changing any of the BIOS settings. Since it is a lab bring-up version of the BIOS, some of the settings have the ability to permanently set fuses in the CPU. One BIOS setting that you can feel free to have fun with is the “concluder” option (Advanced à CPU Configuration à The Concluder), which allows the CPU to report itself to the system as if it is from another CPU vendor + model (e.g., Intel SkylakeX, which can be helpful to make compilers emit AVX512 code or to make software execute down faster codepaths).
· System memory: I am shipping with the DRAM already installed in DIMM slots A1 and B1, since some of these early motherboards can be a little finicky, varying from module to module and the slots they are installed in.
· USB 3.0 ports: IMPORTANT! There is a bug in the motherboard circuitry, where there is not sufficient power supplied to the USB ports. Low-power devices (e.g., mouse, keyboard, USB thumb drive) will work fine in the USB ports on the back panel of the motherboard. For anything higher power (e.g., USB wifi device, and some super high-speed USB drives), It is recommended to install a PCIe expansion card with USB 3.0 ports.
· Hard drive: connect either an SSD via the SATA ports or an NVMe SSD to the connector adjacent to the SATA ports (both NVMe slots function, but the one near the SATA ports is higher speed).
· GPU: The Centaur CPU does not have built-in graphics, so you must use a PCIe card (can be any range from a simple low-end to a good high-end card).
· Power supply: Connect the power supply’s 24-pin main power connector and 8-pin CPU power connector to the motherboard. If you are using a GPU that requires additional power connectors, then connect those as well.
· Various jumpers: Don’t change these.
· HDMI display: I have included one of the small HDMI displays (and a short HDMI cable) that was used in Centaur’s labs, but feel free to use any other display you want. To power the included display, you must connect a micro-USB cable to it (from one of the motherboard ports is fine). The plastic bracket on the display was custom 3D-printed so that it could hang on the open-air racks next to the open-air system setup that it was connected to. For a debug lab, easy access and using easily acquired parts was key. Here is a photo of the type of rack setup that was used in the debug labs:
No quirks, install per normal.
TPM 2.0
Since the company was dissolved prior to CHA becoming fully productized, it was never registered with Microsoft to be recognized as a Trusted Platform Module 2.0 (TPM 2.0) device, even though it technically supports it. As such, if you want to install Windows 11, then you must use a hidden feature of the Windows 11 installation media to allow bypassing the TPM 2.0 requirement. For installing Windows 11 in this way, please follow the procedure described here:
https://www.tomshardware.com/how-to/bypass-windows-11-tpm-requirement
(I am attaching a PDF of the TPM bypass webpage to this Word document as well.)
USB Port Detection
Additionally, the combination of the motherboard’s under-powered USB ports and Windows 11’s attempt to optimize boot speed leads to an unfortunate side effect. If too much power is drawn from the USB ports, they will turn themselves off to protect themselves. When this happens, normally rebooting the machine will re-enable them. However, Windows “remembers” that the USB devices disappeared, and so on the next boot does not even check if they have returned and will never detect anything on those USB ports. This is remedied by disabling Windows “Fast startup” (a web search will show you how to change this setting), which forces Windows to re-check for those USB devices each time it starts up.