Pre-AVX2 x86-64 // compatibility builds
Terabytes for
mere mortals.
Old enterprise iron makes absurd machines attainable: four sockets, hundreds of gigabytes of RAM, and enough fans to levitate a small cat. The catch? You have to meet the hardware where it is. New software packages compiled for old war horses.
01 / Why this exists
Old machines deserve
current software.
The R820 is where this started. The compatibility problem is much bigger than one chassis.
My farm is built from 48-core Ivy Bridge R820s with 512 GB to 3 TB of RAM apiece. I could never justify that much memory in new machines. A retired enterprise server turns it into something a mere mortal can own, fill with data, and learn from.
There are tradeoffs: power, noise, and instruction sets frozen before AVX2. But old is not the same as useless. This project closes the software gap with binaries compiled for the instructions actually present. If a machine matches the target, the build runs; it does not need to be a Dell.
02 / Know your silicon
The instruction flags are the compatibility test.
Dell shipped the R820 with two CPU generations. The same build works beyond Dell wherever the required instruction set is present.
| R820 example | Processor family | F16C | AVX2 | TensorFlow build |
|---|---|---|---|---|
| Sandy Bridge first generation | E5-4600 |
No | No | Coming July 23 |
| Ivy Bridge second generation | E5-4600 v2 |
Yes | No | Available now |
The compatibility tax
No AVX2.
No surrender.
Stock TensorFlow 2.21 does not run on pre-AVX2 CPUs. Its standard Linux x86-64 wheel expects AVX2 instructions. This build does not.
Many current x86-64 binary packages assume AVX2. On an older machine, that can
turn a perfectly good server into an Illegal instruction before
your program gets started. TensorFlow is a common offender.
The answer is not to pretend the instruction exists. It is to compile for the real target—and keep separate builds where F16C support demands it.
03 / AVX1 + F16C build
TensorFlow 2.21
without AVX2.
A Linux x86-64 wheel for CPython 3.13, compiled for AVX1 with F16C and
CUDA support—without requiring AVX2. Use the companion non-F16C build
on AVX1 machines that do not advertise the f16c CPU flag.
- CPU targetAVX1 x86-64 with F16C
- ISAAVX1 + F16C; no AVX2
- RuntimeLinux x86-64 / CPython 3.13
- AccelerationCUDA support
$ python3.13 -m pip install \
https://r820.goblinreactor.com/ivybridge-f16c-cuda/tensorflow-2.21.0-cp313-cp313-linux_x86_64.whlVerify the checksum before installing. CUDA driver and runtime compatibility still depend on your system. Browse this build →