Max Hawkins
@hawkinsmax.bsky.social
CSE PhD student at Georgia Tech working to beneficially integrate computing at scale into society (or at least that's the goal).
Wow! You went to a lutefisk feed! I hope you had a blast. Did you like the lefsa?
September 26, 2025 at 1:42 AM
Wow! You went to a lutefisk feed! I hope you had a blast. Did you like the lefsa?
...(like a traditional roofline analysis), or under/overutilized data type. An example of the latter being not actually using FP8_E5M2's value range for a certain layer weight, or seeing gaps in the value histogram for types with more precision.
August 8, 2025 at 6:38 PM
...(like a traditional roofline analysis), or under/overutilized data type. An example of the latter being not actually using FP8_E5M2's value range for a certain layer weight, or seeing gaps in the value histogram for types with more precision.
For application-performance evaluation, now you can define channels across scale - from hardware pipelines to end-to-end applications. However, I think the kernel or instruction levels might be the most useful and easiest to analyze. This could highlight regions of underutilized hardware...
August 8, 2025 at 6:34 PM
For application-performance evaluation, now you can define channels across scale - from hardware pipelines to end-to-end applications. However, I think the kernel or instruction levels might be the most useful and easiest to analyze. This could highlight regions of underutilized hardware...
I see where you're coming from, and I think the discrepancy comes from the granularity of a 'channel.' For measuring hardware without considering an application, that means looking at each hardware pipeline's (FP64 ALU or Int8 TC) channel capacity (approximated as max input bit width).
...
...
August 8, 2025 at 6:28 PM
I see where you're coming from, and I think the discrepancy comes from the granularity of a 'channel.' For measuring hardware without considering an application, that means looking at each hardware pipeline's (FP64 ALU or Int8 TC) channel capacity (approximated as max input bit width).
...
...
August 8, 2025 at 2:53 PM
These are the first (and only) spoons I made (from firewood) in order from left to right. Functional and rewarding when I use them, but nothing like your work.
July 31, 2025 at 3:30 PM
These are the first (and only) spoons I made (from firewood) in order from left to right. Functional and rewarding when I use them, but nothing like your work.
I'd love to see what your ideas are for something analogous to the 'juicy salif' but for spoons...
July 31, 2025 at 3:27 PM
I'd love to see what your ideas are for something analogous to the 'juicy salif' but for spoons...
I really appreciate the whimsy you bring to an object we usually consider a tool (even naming it a 'utensil'). Your exploration of the art of spoons is refreshing and delightful. The designs and concepts you create are unique yet cohesive to the broader theme.
July 31, 2025 at 3:26 PM
I really appreciate the whimsy you bring to an object we usually consider a tool (even naming it a 'utensil'). Your exploration of the art of spoons is refreshing and delightful. The designs and concepts you create are unique yet cohesive to the broader theme.
Can I buy or trade for one of your spoons?
July 31, 2025 at 1:56 AM
Can I buy or trade for one of your spoons?
This is amazing. You’re really an inspiration.
July 31, 2025 at 1:55 AM
This is amazing. You’re really an inspiration.
NVIDIA math library talk involving emulation: youtu.be/Kx9Z-NCF8J4?...
How Math Libraries Can Help Accelerate Your Applications on Blackwell GPUs | NVIDIA GTC 2025
YouTube video by NVIDIA Developer
youtu.be
June 18, 2025 at 2:09 PM
NVIDIA math library talk involving emulation: youtu.be/Kx9Z-NCF8J4?...
This puts a B300's double precision performance between an M40 (released in 2015) and a P100 (2016 release). Single precision performance is pretty stellar still though.
Emulation of higher precision data formats is possible, but involves more perf/accuracy tradeoffs than native hardware support.
Emulation of higher precision data formats is possible, but involves more perf/accuracy tradeoffs than native hardware support.
June 18, 2025 at 2:05 PM
This puts a B300's double precision performance between an M40 (released in 2015) and a P100 (2016 release). Single precision performance is pretty stellar still though.
Emulation of higher precision data formats is possible, but involves more perf/accuracy tradeoffs than native hardware support.
Emulation of higher precision data formats is possible, but involves more perf/accuracy tradeoffs than native hardware support.
How do you think we can prevent a segmented benchmarking/performance metric landscape? I've been thinking about this for a bit, but what are your thoughts? I don't want to see yet another area HPC and AI diverge in.
June 18, 2025 at 2:46 AM
How do you think we can prevent a segmented benchmarking/performance metric landscape? I've been thinking about this for a bit, but what are your thoughts? I don't want to see yet another area HPC and AI diverge in.
Unless we have some sense of composable data formats that don't incur hardware implementation performance degradation...
June 16, 2025 at 2:55 AM
Unless we have some sense of composable data formats that don't incur hardware implementation performance degradation...
I've been thinking a lot about data formats in HPC lately, and our current arrangement does seem more arbitrary than I was led to believe. Many assumptions that may be worth revisiting...
The history of the IEEE 754 standard is also more interesting than I previously thought!
The history of the IEEE 754 standard is also more interesting than I previously thought!
June 16, 2025 at 2:51 AM
I've been thinking a lot about data formats in HPC lately, and our current arrangement does seem more arbitrary than I was led to believe. Many assumptions that may be worth revisiting...
The history of the IEEE 754 standard is also more interesting than I previously thought!
The history of the IEEE 754 standard is also more interesting than I previously thought!