Folding Forum

With the release of Intel Core Ultra Series 3 (Panther Lake), I am requesting the addition of the new Xe3 integrated graphics to the official whitelist.

This architecture represents a significant jump in compute density on the 18A process, with the top-tier Ultra X9 388H featuring 12 Xe-cores. Early testing indicates these units can match entry-level discrete GPU performance, making them viable for sustained folding in small form factor systems.

Current technical details:

Architecture: Intel Xe3 (Arc B390)

Driver: Intel Production Compute Runtime 26.01 (OpenCL 3.0 support confirmed)

Peak TDP: Scalable up to 80W+ in certain Mini PC configurations

I am prepared to provide the specific Hardware PCI ID from Device Manager as soon as the unit is initialized to assist with the GPUs.txt update. Has anyone else on the science team started internal testing on 18A silicon for core compatibility?

Does it support FP64?

muziqaz wrote: ↑Thu Jan 15, 2026 7:53 pm Does it support FP64?

The Arc B390 is a different beast than the old Xe-LP/Gen 12 stuff where Intel basically disabled hardware FP64.

Since Panther Lake's Xe3 architecture is part of the Battlemage family, hardware-level FP64 is definitely back. It's not at "pro-grade" ratios like 1:2, but early info points to a 1:16 ratio for the Xe3 cores. That's a massive win over the slow software emulation we saw on Meteor Lake and should be plenty to clear project deadlines without the validation errors we used to get on Intel iGPUs.

I’ll have a unit soon and can post a clinfo log or a quick clpeak test to confirm the exact double-precision FLOPs for you guys.

Please provide PCI ID
and maybe FP32 number

Thanks

And FP64 number.

Info I have seen so far is that FP64 will be supported in hardware only on the Battlemage discrete GPU cards, not on the iGPUs. Whether that means a separate chip series without the hardware for FP64, or just the function disabled in binned chips I haven't seen.

The problem with FP64 number is that might be produced by software emulated stuff during clpeak run. While other software might not care if FP64 is simulated or not, but FAH crashes

Just to clear up the confusion regarding FP64 on the iGPU—Intel has officially confirmed that the Xe3 architecture in Panther Lake (Core Ultra Series 3) does have native hardware FP64 support. It is no longer emulated in software as it was in the older Xe-LP units.

Since Xe3 is a direct derivative of the Battlemage architecture, it includes the same hardware FP64 blocks found in the discrete B-series cards. The ratio is 1:16 relative to FP32. This hardware-native support should provide the stability needed for scientific cores that require high-precision math without the "numerical drift" or crashes seen on previous generations.

Once my unit arrives, I'll run a clpeak test to verify the double-precision GFLOPS and post the results here for the science team.

Do you havre a link to that official confirmation? I do know when originally announced the Xe2 Battlemage architecture it was going to have two series of chips - the Xe2-LPG for iGPUs and Xe2-HPG. Lunar Lake processors shipped with Xe2-SPG chips for their iGPUs and do not support FP64, the discrete cards are based on Xe2-HPG. So wondering if that is going to be the same for Xe3.

Joe_H wrote: ↑Thu Jan 15, 2026 10:57 pm Do you havre a link to that official confirmation? I do know when originally announced the Xe2 Battlemage architecture it was going to have two series of chips - the Xe2-LPG for iGPUs and Xe2-HPG. Lunar Lake processors shipped with Xe2-SPG chips for their iGPUs and do not support FP64, the discrete cards are based on Xe2-HPG. So wondering if that is going to be the same for Xe3.

The confusion around FP64 usually comes from the Xe-LP and Xe2-LPG (Lunar Lake) chips, which were strictly low-power designs that lacked these compute blocks. Panther Lake is different because the Xe3 architecture is a direct derivative of the Battlemage family.

Intel has confirmed that each Xe3 core in the Arc B390 includes hardware-native FP64 blocks with a 1:16 ratio relative to FP32. This is built into the silicon itself, not emulated in the driver, so it should provide the stability needed to avoid the client crashes seen on older iGPUs.

I’ll have a unit in hand by the 27th and will provide the exact PCI ID and a clpeak report to confirm the hardware double-precision throughput for the science team.

No no, confusion is coming purely from Intel
While their driver team is doing wonderful things, the rest of them are just talking for the sake of talking.
They rarely distinguish between hardware and software emulation. They just state supported and leave it at that

muziqaz wrote: ↑Fri Jan 16, 2026 9:22 am No no, confusion is coming purely from Intel
While their driver team is doing wonderful things, the rest of them are just talking for the sake of talking.
They rarely distinguish between hardware and software emulation. They just state supported and leave it at that

muziqaz, you’re 100% right to be skeptical given how the old Xe-LP and Xe-LPG (Meteor Lake) drivers handled things. Those were strictly low-power designs where hardware FP64 was literally removed from the silicon, leaving only unstable software emulation that FAH hated.

The reason Panther Lake is a different story is that the Xe3 architecture is a direct derivative of the Battlemage family. Unlike the previous "LPG" chips, the Xe3 Vector Engines in the Arc B390 have hardware-native "Extended Math and FP64" blocks built into the silicon. Intel has confirmed a hardware-native 1:16 ratio for FP64 on this architecture.

Because this is native silicon and not a driver-level workaround, it provides the fixed-function stability needed to avoid the numerical drift and crashes we saw in the past. I’ll post the clpeak --compute-dp results as soon as I have the unit on the 27th to prove the hardware GFLOPS are real.

You still haven't posted any link to any official Intel statements on this. Interpreting those statements depend a lot on who is making them and the exact wording.

Joe_H wrote: ↑Fri Jan 16, 2026 5:35 pm You still haven't posted any link to any official Intel statements on this. Interpreting those statements depend a lot on who is making them and the exact wording.

https://newsroom.intel.com/client-compu ... el-18a?hl=

https://wccftech.com/intel-xe3-graphics ... amily/?hl=

https://www.techpowerup.com/review/inte ... 8.html?hl=

Ncard00 wrote: ↑Sat Jan 17, 2026 10:55 am

Joe_H wrote: ↑Fri Jan 16, 2026 5:35 pm You still haven't posted any link to any official Intel statements on this. Interpreting those statements depend a lot on who is making them and the exact wording.

https://newsroom.intel.com/client-compu ... el-18a?hl=

https://wccftech.com/intel-xe3-graphics ... amily/?hl=

https://www.techpowerup.com/review/inte ... 8.html?hl=

Neither Intel newsroom, nor techpowerup mention anything about FP64. Since today's PR people don't know how to count past FP8
However, let me know when you get your device, give me pci id, and we will see if it folds or not. If it doesn't fold reliably, I can easily disable it.
Thanks

Okay, can find a single illustration indicating FP64 in the Xe Vector Engine graphic a third of the way through the wccftech writeup. Assuming that graphic is accurate as supplied by Intel, then yes it appears to support FP64 in hardware. The final say will be in the developer notes issued by the software engineers at Intel writing up support for the new iGPUs.