Torsten Hoefler 🇨🇭
@thoefler.bsky.social
Professor ETHZ, head of SPCL, Chief Architect ML at CSCS researching large-scale #HPC and #AI systems and #Climate computing - youtube: http://bit.ly/3h1VgIU
A special memory of Piz Daint - once Europe's fastest and first large #GPU #supercomputer
After 13 years of service at #ETH #CSCS, it retired from service. Those two blades shall live on as a milestone in Swiss and European history!
Better than spatulas #nvidia? Enjoy the festive season 🎄 #AI #HPC
After 13 years of service at #ETH #CSCS, it retired from service. Those two blades shall live on as a milestone in Swiss and European history!
Better than spatulas #nvidia? Enjoy the festive season 🎄 #AI #HPC
December 26, 2025 at 6:00 AM
Switzerland, where you go with your hybrid car by electric train to the mountains 🏔️. Always cool.
Happy holidays! ❄️
Happy holidays! ❄️
December 24, 2025 at 7:42 AM
Switzerland, where you go with your hybrid car by electric train to the mountains 🏔️. Always cool.
Happy holidays! ❄️
Happy holidays! ❄️
Reposted by Torsten Hoefler 🇨🇭
With the year slowly winding down, we took a bit of time to get together, celebrate and reflect on the amazing 2025 journey and all the learnings and achievements during our Xmas dinner last Friday.🎄
Happy holidays & 2026, here we come!🚀🥳
#HPC #AI @thoefler.bsky.social @csateth.bsky.social
Happy holidays & 2026, here we come!🚀🥳
#HPC #AI @thoefler.bsky.social @csateth.bsky.social
December 22, 2025 at 8:04 AM
With the year slowly winding down, we took a bit of time to get together, celebrate and reflect on the amazing 2025 journey and all the learnings and achievements during our Xmas dinner last Friday.🎄
Happy holidays & 2026, here we come!🚀🥳
#HPC #AI @thoefler.bsky.social @csateth.bsky.social
Happy holidays & 2026, here we come!🚀🥳
#HPC #AI @thoefler.bsky.social @csateth.bsky.social
NVIDIA explains why they used FP8 as a scaling factor in NVFP and not MXFP-style E8M0.
The idea is to preserve the largest value in each block.
buff.ly/SjwVr6w (16:00 ff)
We found this to be very true in a study and managed to recover some MXFP accuracy with micro-rotations: buff.ly/gXaCAhF
The idea is to preserve the largest value in each block.
buff.ly/SjwVr6w (16:00 ff)
We found this to be very true in a study and managed to recover some MXFP accuracy with micro-rotations: buff.ly/gXaCAhF
December 22, 2025 at 6:00 AM
NVIDIA explains why they used FP8 as a scaling factor in NVFP and not MXFP-style E8M0.
The idea is to preserve the largest value in each block.
buff.ly/SjwVr6w (16:00 ff)
We found this to be very true in a study and managed to recover some MXFP accuracy with micro-rotations: buff.ly/gXaCAhF
The idea is to preserve the largest value in each block.
buff.ly/SjwVr6w (16:00 ff)
We found this to be very true in a study and managed to recover some MXFP accuracy with micro-rotations: buff.ly/gXaCAhF
Programming modern accelerators uses more and more complex forms of parallelism for example pipelining loading and computing at all levels.
buff.ly/K85kokX
Check out Tiziano's FBLAS and SMI for some fundamentals behind pipelining in applications and networks!
buff.ly/wPnHfaJ
buff.ly/65mQo7q
buff.ly/K85kokX
Check out Tiziano's FBLAS and SMI for some fundamentals behind pipelining in applications and networks!
buff.ly/wPnHfaJ
buff.ly/65mQo7q
December 17, 2025 at 6:00 AM
Programming modern accelerators uses more and more complex forms of parallelism for example pipelining loading and computing at all levels.
buff.ly/K85kokX
Check out Tiziano's FBLAS and SMI for some fundamentals behind pipelining in applications and networks!
buff.ly/wPnHfaJ
buff.ly/65mQo7q
buff.ly/K85kokX
Check out Tiziano's FBLAS and SMI for some fundamentals behind pipelining in applications and networks!
buff.ly/wPnHfaJ
buff.ly/65mQo7q
Jeff Hammond from NVIDIA gives some historical context on NCCL's lineage going back to MPI, similarly for NVSHMEM. #AI is rooted in a strong #HPC networking history!
buff.ly/kMPnbBf (04:00)
Learn more about NCCL details in our paper: buff.ly/doin5bt
buff.ly/kMPnbBf (04:00)
Learn more about NCCL details in our paper: buff.ly/doin5bt
December 15, 2025 at 6:00 AM
Jeff Hammond from NVIDIA gives some historical context on NCCL's lineage going back to MPI, similarly for NVSHMEM. #AI is rooted in a strong #HPC networking history!
buff.ly/kMPnbBf (04:00)
Learn more about NCCL details in our paper: buff.ly/doin5bt
buff.ly/kMPnbBf (04:00)
Learn more about NCCL details in our paper: buff.ly/doin5bt
Nice overview of PCIe 6.0 key metrics and development - a key goal was to not increase the latency by more than 10ns. Total FEC latency goal <2ns!
Watch: buff.ly/c9hBJcV
Watch: buff.ly/c9hBJcV
December 8, 2025 at 6:00 AM
Nice overview of PCIe 6.0 key metrics and development - a key goal was to not increase the latency by more than 10ns. Total FEC latency goal <2ns!
Watch: buff.ly/c9hBJcV
Watch: buff.ly/c9hBJcV
Ilya Sutskever: "Compression is prediction and vice versa." (15:10).
A good compressor will exploit the shared structure - now this forms a strong mathematical reason for exploiting all information in the unlabeled data - independently of the later tasks.
Watch: buff.ly/Bnvazym
A good compressor will exploit the shared structure - now this forms a strong mathematical reason for exploiting all information in the unlabeled data - independently of the later tasks.
Watch: buff.ly/Bnvazym
December 3, 2025 at 6:00 AM
Ilya Sutskever: "Compression is prediction and vice versa." (15:10).
A good compressor will exploit the shared structure - now this forms a strong mathematical reason for exploiting all information in the unlabeled data - independently of the later tasks.
Watch: buff.ly/Bnvazym
A good compressor will exploit the shared structure - now this forms a strong mathematical reason for exploiting all information in the unlabeled data - independently of the later tasks.
Watch: buff.ly/Bnvazym
I talked about a path towards an "AI Climate Scientist" at the ADIA Lab Symposium in Abu Dhabi and the Spanish Royal Academy in Madrid.
Some ways to combine #AI and #HPC to address the data and computational challenge in #Climate sciences and accelerate insights.
Watch now: buff.ly/7qeGuqf
Some ways to combine #AI and #HPC to address the data and computational challenge in #Climate sciences and accelerate insights.
Watch now: buff.ly/7qeGuqf
December 1, 2025 at 6:00 AM
I talked about a path towards an "AI Climate Scientist" at the ADIA Lab Symposium in Abu Dhabi and the Spanish Royal Academy in Madrid.
Some ways to combine #AI and #HPC to address the data and computational challenge in #Climate sciences and accelerate insights.
Watch now: buff.ly/7qeGuqf
Some ways to combine #AI and #HPC to address the data and computational challenge in #Climate sciences and accelerate insights.
Watch now: buff.ly/7qeGuqf
Bill Dally on NVIDIA networking: "Jensen said: 'Absolutely not, we don't do networking, we're a GPU company'".
buff.ly/iUHv62C (4:30)
Then he talked the DOE into paying 100%, launching NVLINK
Example how a research lab changed the course of its organization. Another example is "all of AI" :-)
buff.ly/iUHv62C (4:30)
Then he talked the DOE into paying 100%, launching NVLINK
Example how a research lab changed the course of its organization. Another example is "all of AI" :-)
Insights From NVIDIA Research S73202 | GTC 2025 | NVIDIA On-Demand
The talk will give some highlights from NVIDIA Research for the past year. Detailed topics will be disclosed closer to the event.
buff.ly
November 26, 2025 at 6:00 AM
Bill Dally on NVIDIA networking: "Jensen said: 'Absolutely not, we don't do networking, we're a GPU company'".
buff.ly/iUHv62C (4:30)
Then he talked the DOE into paying 100%, launching NVLINK
Example how a research lab changed the course of its organization. Another example is "all of AI" :-)
buff.ly/iUHv62C (4:30)
Then he talked the DOE into paying 100%, launching NVLINK
Example how a research lab changed the course of its organization. Another example is "all of AI" :-)
Congrats to Saleh Ashkboos, SPCL's 13th PhD graduate (another prime number). A new expert in quantization and LLM optimization is born.
Thanks also to the great collaboration with Dan Alistarh from ISTA and James Hensman at Microsoft!
Thanks also to the great collaboration with Dan Alistarh from ISTA and James Hensman at Microsoft!
November 25, 2025 at 5:36 PM
Congrats to Saleh Ashkboos, SPCL's 13th PhD graduate (another prime number). A new expert in quantization and LLM optimization is born.
Thanks also to the great collaboration with Dan Alistarh from ISTA and James Hensman at Microsoft!
Thanks also to the great collaboration with Dan Alistarh from ISTA and James Hensman at Microsoft!
I am very honored to be part of to the two brilliant teams winning the 2025 Gordon Bell Prize for Climate Modeling and the 2025 Gordon Bell Prize (Honorable Mention). Both among the highest honors in #HPC. 🍾
buff.ly/gCkjDRN and buff.ly/4ewhdZ9
Congrats teams - Switzerland 🇨🇭 going strong - onward!
buff.ly/gCkjDRN and buff.ly/4ewhdZ9
Congrats teams - Switzerland 🇨🇭 going strong - onward!
November 24, 2025 at 6:00 AM
I am very honored to be part of to the two brilliant teams winning the 2025 Gordon Bell Prize for Climate Modeling and the 2025 Gordon Bell Prize (Honorable Mention). Both among the highest honors in #HPC. 🍾
buff.ly/gCkjDRN and buff.ly/4ewhdZ9
Congrats teams - Switzerland 🇨🇭 going strong - onward!
buff.ly/gCkjDRN and buff.ly/4ewhdZ9
Congrats teams - Switzerland 🇨🇭 going strong - onward!
Former member of SPCL, collaborator, and friend Daniele De Sensi speaks at both Broadcom's booth and at the main track of the Supercomputing conference.
Onward! We're all proud of you 🎉
Onward! We're all proud of you 🎉
November 21, 2025 at 12:18 PM
Former member of SPCL, collaborator, and friend Daniele De Sensi speaks at both Broadcom's booth and at the main track of the Supercomputing conference.
Onward! We're all proud of you 🎉
Onward! We're all proud of you 🎉
Reposted by Torsten Hoefler 🇨🇭
🌍 A 26-member team has been awarded the 2025 Gordon Bell Prize for Climate Modelling for their project, “Computing the Full Earth System at 1 km Resolution.”
Congratulations to this year’s outstanding team!
buff.ly/GKDYnH3
Congratulations to this year’s outstanding team!
buff.ly/GKDYnH3
November 20, 2025 at 8:30 PM
🌍 A 26-member team has been awarded the 2025 Gordon Bell Prize for Climate Modelling for their project, “Computing the Full Earth System at 1 km Resolution.”
Congratulations to this year’s outstanding team!
buff.ly/GKDYnH3
Congratulations to this year’s outstanding team!
buff.ly/GKDYnH3
Rajeev Thakur is kicking off our Advanced MPI tutorial at #SC25. It's always an honor to teach this long standing tutorial with esteemed colleagues including Bill Gropp and Pavan Balaji. Great attendance 👌, still some seats in 122.
We're looking forward to a productive session.
We're looking forward to a productive session.
November 17, 2025 at 2:36 PM
Rajeev Thakur is kicking off our Advanced MPI tutorial at #SC25. It's always an honor to teach this long standing tutorial with esteemed colleagues including Bill Gropp and Pavan Balaji. Great attendance 👌, still some seats in 122.
We're looking forward to a productive session.
We're looking forward to a productive session.
Kurt Ferreira opens the 14th addition of our ROSS workshop at #SC25! Supercomputer operating systems and middleware going strong! Packed room as always 😀.
Starting with an invited talk by NVIDIA's Jeff Hammond on communication systems.
Trivia: he probably traveled furthest ✈️
Starting with an invited talk by NVIDIA's Jeff Hammond on communication systems.
Trivia: he probably traveled furthest ✈️
November 16, 2025 at 8:15 PM
Kurt Ferreira opens the 14th addition of our ROSS workshop at #SC25! Supercomputer operating systems and middleware going strong! Packed room as always 😀.
Starting with an invited talk by NVIDIA's Jeff Hammond on communication systems.
Trivia: he probably traveled furthest ✈️
Starting with an invited talk by NVIDIA's Jeff Hammond on communication systems.
Trivia: he probably traveled furthest ✈️
Arrived at #SC25 and just moved into the SPCL den 🏠 - our homebase for the whole on-site team this week. Readying ourselves for a crazy time - my three talks for tomorrow should mainly be set 😀.
Looking forward to seeing all of you in person - for those who I have not already met 🤝.
Looking forward to seeing all of you in person - for those who I have not already met 🤝.
November 16, 2025 at 3:17 PM
Arrived at #SC25 and just moved into the SPCL den 🏠 - our homebase for the whole on-site team this week. Readying ourselves for a crazy time - my three talks for tomorrow should mainly be set 😀.
Looking forward to seeing all of you in person - for those who I have not already met 🤝.
Looking forward to seeing all of you in person - for those who I have not already met 🤝.
Very nice overview of the emerging UALink standard with nice features such as splitting packets in switches, in-network computing, high energy efficiency, and lowest silicon overhead: buff.ly/AgLvC1g
I'll be joining a panel at SC25 contrasting UALink and UEC next Wed: buff.ly/BeCMFcL Join us there
I'll be joining a panel at SC25 contrasting UALink and UEC next Wed: buff.ly/BeCMFcL Join us there
Introducing the UALink 200G 1.0 Specification Webinar
The Ultra Accelerator Link™ (UALink™) Consortium is an open industry standard group dedicated to advancing the UALink specification. The Consortium recently released the UALink 200G 1.0…
buff.ly
November 14, 2025 at 6:00 AM
Very nice overview of the emerging UALink standard with nice features such as splitting packets in switches, in-network computing, high energy efficiency, and lowest silicon overhead: buff.ly/AgLvC1g
I'll be joining a panel at SC25 contrasting UALink and UEC next Wed: buff.ly/BeCMFcL Join us there
I'll be joining a panel at SC25 contrasting UALink and UEC next Wed: buff.ly/BeCMFcL Join us there
Ram Velega (Broadcom) at OCP "The things that have confined us today that the scale-up domain is less than 100 are going to be changed very soon" (7:55 in buff.ly/9ioPH73)
I see this as a call to action for systems and software research.
UE 1.0 techniques such as LLR and CBFC play a crucial role!
I see this as a call to action for systems and software research.
UE 1.0 techniques such as LLR and CBFC play a crucial role!
2025 OCP APAC Summit Keynote - Broadcom
Ram Velaga (Broadcom)
Scale-up for AI: Balancing Compute, Memory and Networking
AI models are growing at an unprecedented pace, driving exponential increases in compute demands. As a single GPU…
buff.ly
November 12, 2025 at 6:00 AM
Ram Velega (Broadcom) at OCP "The things that have confined us today that the scale-up domain is less than 100 are going to be changed very soon" (7:55 in buff.ly/9ioPH73)
I see this as a call to action for systems and software research.
UE 1.0 techniques such as LLR and CBFC play a crucial role!
I see this as a call to action for systems and software research.
UE 1.0 techniques such as LLR and CBFC play a crucial role!
Ilya Sutskever "Three lines of math can prove all of supervised learning. That's nice" (4:33)
"I have not seen an exposition of unsupervised learning that I found satisfying" (7:50)
Its optimization objective has little relation to the actual objective you care about
Watch: buff.ly/Bnvazym
"I have not seen an exposition of unsupervised learning that I found satisfying" (7:50)
Its optimization objective has little relation to the actual objective you care about
Watch: buff.ly/Bnvazym
November 10, 2025 at 6:00 AM
Ilya Sutskever "Three lines of math can prove all of supervised learning. That's nice" (4:33)
"I have not seen an exposition of unsupervised learning that I found satisfying" (7:50)
Its optimization objective has little relation to the actual objective you care about
Watch: buff.ly/Bnvazym
"I have not seen an exposition of unsupervised learning that I found satisfying" (7:50)
Its optimization objective has little relation to the actual objective you care about
Watch: buff.ly/Bnvazym
Can we build an #AI #Climate Scientist? Asked at the ADIA Lab Symposium in Abu Dhabi last week - now online at buff.ly/6igSeyg :-).
Much work to be done - this is outlining some directions of indicative results with a lot of potential to accelerate AI for Science.
Much work to be done - this is outlining some directions of indicative results with a lot of potential to accelerate AI for Science.
November 9, 2025 at 9:24 AM
Can we build an #AI #Climate Scientist? Asked at the ADIA Lab Symposium in Abu Dhabi last week - now online at buff.ly/6igSeyg :-).
Much work to be done - this is outlining some directions of indicative results with a lot of potential to accelerate AI for Science.
Much work to be done - this is outlining some directions of indicative results with a lot of potential to accelerate AI for Science.
Collaborator and friend Dan Alistarh talks at ETH about using the new NvFP4 and MXFP4 block formats for inference.
Some going from "terrible" accuracy to acceptable using micro rotations to smoothen outliers in blocks.
arxiv.org/abs/2509.23202
Great collaboration and cool stuff
Some going from "terrible" accuracy to acceptable using micro rotations to smoothen outliers in blocks.
arxiv.org/abs/2509.23202
Great collaboration and cool stuff
November 5, 2025 at 8:32 AM
Collaborator and friend Dan Alistarh talks at ETH about using the new NvFP4 and MXFP4 block formats for inference.
Some going from "terrible" accuracy to acceptable using micro rotations to smoothen outliers in blocks.
arxiv.org/abs/2509.23202
Great collaboration and cool stuff
Some going from "terrible" accuracy to acceptable using micro rotations to smoothen outliers in blocks.
arxiv.org/abs/2509.23202
Great collaboration and cool stuff
Keren Bergman at the 2nd EFCL Workshop: "Huawei combines 3x less performant GPUs with a photonic scale-up network to build higher performance PODs than Nvidia."
Nvidia moving towards CPO 😀. Optics everywhere in scale-out.
Nice overview of optical networking at all distances.
Nvidia moving towards CPO 😀. Optics everywhere in scale-out.
Nice overview of optical networking at all distances.
November 4, 2025 at 12:49 PM
Keren Bergman at the 2nd EFCL Workshop: "Huawei combines 3x less performant GPUs with a photonic scale-up network to build higher performance PODs than Nvidia."
Nvidia moving towards CPO 😀. Optics everywhere in scale-out.
Nice overview of optical networking at all distances.
Nvidia moving towards CPO 😀. Optics everywhere in scale-out.
Nice overview of optical networking at all distances.