News: 0001588055

  ARM Give a man a fire and he's warm for a day, but set fire to him and he's warm for the rest of his life (Terry Pratchett, Jingo)

AMD ROCm 7.1 Released: Many Instinct MI350 Series Improvements, Better Performance

([AMD] 42 Minutes Ago ROCm 7.1)


As expected after noting this morning that [1]ROCm 7.1 release preparations were underway , ROCm 7.1 is now officially released as the newest step-forward for this open-source GPU compute stack for Radeon and Instinct hardware.

ROCm 7.1 is coming just one and a half months after the major ROCm 7.0 release. ROCm 7.1 continues making a lot of enhancements around the AMD Instinct MI350X and MI355X support, including numerous performance optimizations and new features. Plus other performance improvements not specific to the Instinct MI350 series. There is also expanded Linux distribution coverage, HIP improvements for better compatibility against NVIDIA CUDA interfaces, and various other changes.

Some of the ROCm 7.1 highlights include:

- AMD Instinct MI350X and MI355X accelerator support is now available on Debian 13 with the AMDGPU DKMS packaged kernel driver.

- The AMD Instinct MI325X is now officially supported on RHEL 10, SLES 15 SP7, Debian 12, Debian 13, Oracle Linux 9, and Oracle Linux 10.

- The old AMD Instinct MI100 is officially supported now on SUSE Linux Enterprise Server 15 SP7.

- ROCm 7.1 guest OS support for RHEL 10.0 using KVM SR-IOV with the AMD Instinct MI350X/MI355X hardware.

- AMD SMI now allows setting a power cap in 1VF for the Instinct MI300X.

- Various virtualization improvements for the AMD Instinct MI350 series.

- New HIP runtime APIs to enhance compatibility with NVIDIA CUDA. AMD's HIP can now also support nested tile partitioning within cooperative groups as one of the improvements for better matching CUDA functionality. The HIP module loading latency has also been lowered.

- Performance improvements to hipBLASLt, including a number of kernel optimizations for the Instinct MI350 series.

- Significant performance enhancements to hipSPARSELt.

- RCCL has enhanced single-node performance for the Instinct MI350 series.

- Various ROCm profiler updates.

- Support for TensorFlow 2.20 and ONNX RT 1.23.1.

More details and downloads for today's ROCm 7.1 release via [2]rocm.docs.amd.com .



[1] https://www.phoronix.com/news/AMD-ROCm-7.1-Imminent

[2] https://rocm.docs.amd.com/en/latest/



"For a couple o' pins," says Troll, and grins,
"I'll eat thee too, and gnaw thy shins.
A bit o' fresh meat will go down sweet!
I'll try my teeth on thee now.
Hee now! See now!
I'm tired o' gnawing old bones and skins;
I've a mind to dine on thee now."

But just as he thought his dinner was caught,
He found his hands had hold of naught.
Before he could mind, Tom slipped behing
And gave him the boot to larn him.
Warn him! Darn him!
A bump o' the boot on the seat, Tom thoguht,
Would be the way to larn him.

But harder than stone is the flesh and bone
Of a troll that sits in the hills alone.
As well set your boot to the mountain's root,
For the seat of a troll don't feel it.
Peel it! Heal it!
Old Troll laughed, when he heard Tom groan,
And he knew his toes could feel it.

Tom's leg is game, since home he came,
And his bootless foot is lasting lame;
But Troll don't care, and he's still there
With the bone he boned from its owner.
Doner! Boner!
Troll's old seat is still the same,
And the bone he boned from its owner!
-- J. R. R. Tolkien