AMD Preparing More Linux Code For The Frontier Supercomputer

[phoronix]

Phoronix: AMD Preparing More Linux Code For The Frontier Supercomputer

Frontier as the first US exascale supercomputer being commissioned by Oak Ridge National Laboratory and the Department of Energy. while being powered by AMD CPUs/GPUs is in the process of seeing more Linux kernel changes for bringing up the new platform...

https://www.phoronix.com/scan.php?page=news_item&px=AMD-VMA-Changes-Frontier

[pipe13]

...is why https://www.phoronix.com/scan.php?pa...-Driver-Coming is likely Nvidia trying to stay in the game. I realize graphics driver is not quite the same as memory management. But still...

[torsionbar28]

...is why https://www.phoronix.com/scan.php?pa...-Driver-Coming is likely Nvidia trying to stay in the game. I realize graphics driver is not quite the same as memory management. But still...

Highly doubtful. For the big government supercomputers, NVDA likely does provide code... under strict NDA of course. Chances of NVDA releasing said code as FOSS? Pretty close to zero I'd wager.

[piorunz]

They gonna build supercomputer out of AMD GPUs? I know where all these Navi cards were going then, market is totally dry. They will be mining Ethereum with their thousands of cards 😂

[Azrael5]

It reminds me a kind of AGP aperture size.

[Jabberwocky]

They gonna build supercomputer out of AMD GPUs? I know where all these Navi cards were going then, market is totally dry. They will be mining Ethereum with their thousands of cards 😂

In the past supercomputers have taken tens of thousands of cards and there wasn't any shortage in the market. With the silicon shortage it probably has some noticeable effect. The most important will always come first. It's the same with nvidia's mining cards. Us consumers aren't willing to pay that much or place that big orders, at least not from a single entity that represents us all.

[ms178]

I hope this work also will be beneficial as cache coherent interconnects between GPU/CPU work their way down to consumer level hardware. After all CXL is the future (and I would be disapointed if rumors are true that we won't see this introduced on AM5 due to a lack of PCIe 5.0 support; it is also about time the industry comes together to introduce something better than the current PCIe connector as that technology reached its limits. The Gen-Z consortium came up with SFF-TA1002 which would allow for cheaper motherboards, better signal integrity and more flexibility and as I see it there is broad support behind that already. A new platform with new GPUs to make use of it would be perfect to introduce it (probably in servers first). But it seems that we have to wait yet another 2+X years to see that happen.

[skeevy420]

You know who else does DOE work with supercomputers?

OpenZFS and ZFSonLinux.

I wonder if AMD doing contractual work from the same place means that ZFS will get that NVME-to-GPU feature? OpenZFS already has all the bits and pieces in place for caching and sending data to all sorts of different high-speed disk controllers so sending data to the GPU shouldn't be that big of a deal (at least to a Layman).

AMD is building a system architecture for the Frontier supercomputer with a coherent interconnect between CPUs and GPUs.

Does this mean they're building an Epyc APU?

[wizard69]

If I remember correctly the "video" card support is likely for CDNA chips of which there should be 4 per CPU. I don't think there is a real need to put in RDNA 2 or 3 cards in these machines.

[zxy_thf]

They gonna build supercomputer out of AMD GPUs? I know where all these Navi cards were going then, market is totally dry. They will be mining Ethereum with their thousands of cards 😂

AMD secured more than one HPC order for their professional GPUs, and it seems they want Infinity Fabric to be the bus connecting Epyc and CDNA.
Otherwise there would be unnecessary overhead by using PCIe to achieve memory+cache coherency.

This also explains their weird strategy in GPU computing.
The benefits of market penetration from RDNA 1/2/3/... would appeal future developers and make the ecosystem healthier, but they must meed the demands from the biggest customers first and finish their ROCm on CDNA before the deadline in their signed contracts.

But still I'd say, an understandable strategy doesn't mean I'm happy about it.