News Intel Xe HP and Xe HPC: Packing a Whopping 40+ TFLOPS

[Admin]

Intel is going after big data with its Xe HP and Xe HPC GPUs. Die stacking and EMIB allow the current design to scale to a 4-tile solution that delivers over 40 TFLOPS of single precision compute.

Intel Xe HP and Xe HPC: Packing a Whopping 40+ TFLOPS : Read more

 

[setx]

Xe HP also supports the bfloat16 format, which offers similar accuracy to FP32 but at twice the performance

Please read something before you write such nonsense.
BFloat16 has the same range as FP32 but precision(accuracy) is much worse.

 

[JarredWaltonGPU]

Please read something before you write such nonsense.
BFloat16 has the same range as FP32 but precision(accuracy) is much worse.

I've clarified that this 'accuracy' is specifically for machine learning purposes.

 

[Deicidium369]

WCC had an exclusive stating that initially a 6nm TSMC process would be used, which is a slightly more dense 7nm - the updated exclusive is that will be on TSMC 5nm - as it is roughly the same density as Intel 7nm - this will NOT be a super high volume part - likely this arrangement just for the Aurora Exascale - which was 1st 7nm to have "shipped" - and the one put back 6-12 months.

Rocket Lake - Late 2020 - 14nm
Alder Lake - Late 2021 - 10nm
Meteor Lake - Late 2022 / Early 2023. - 7nm
~1 year cadence

Ice Lake SP - Late 2020 (very short platform life) 10nm
Sapphire Rapids - Mid to Late 2021 - 10nm
(will be a long platform, unlike Cooper Lake / Ice Lake SP)
Granite Rapids - 2022/ 2023 - 7nm

Sapphire Rapids reunified the Xeon for 1-8CPU - and will likely be a longer term platform - with Granite Rapids being just a shrink - later in the platforms life.

So looks like the PVC was the part pushed back 6-12 months - Not like we were going to get Rocket Lake and 6 months later Alder and 6 months after that Meteor

https://wccftech.com/exclusive-intel-ponte-vecchio-gpu-not-on-tsmc-6nm/