technoabsurdist
technoabsurdist on Hacker News
Re:Running Kimi K3 on MI355X at Better Performance per Dollar Than B300
on Sunday August 02, 2026 @12:29PM (#49145961)
hi I work at wafer. yes we ran benchmarks. for example our Kimi K3 is live on open router and in order to host there you have to run accuracy checks. like tau/gpqa.
and then u must pass test regarding thinking, coherency, and tool calling
Re:GLM5.2 on AMD MI355X at 2626 tok/s/node at over 2x lower cost than ...
on Saturday July 04, 2026 @01:52AM (#48782959)
sounds good feedback taken, thanks beffjezos
Re:GLM5.2 on AMD MI355X at 2626 tok/s/node at over 2x lower cost than ...
on Friday July 03, 2026 @11:47PM (#48782484)
hi yes it’s not optimized for single stream it’s optimized for total node throughput
Re:GLM5.2 on AMD MI355X at 2626 tok/s/node at over 2x lower cost than ...
on Friday July 03, 2026 @08:22PM (#48781541)
this is exactly our thesis at wafer :) thank you for the support
Re:GLM5.2 on AMD MI355X at 2626 tok/s/node at over 2x lower cost than ...
on Friday July 03, 2026 @08:21PM (#48781534)
AMD MI355X uses 1,400W per GPU and NVIDIA B200 uses 1,200W. So AMD uses about 16% more power.
Re:GLM5.2 on AMD MI355X at 2626 tok/s/node at over 2x lower cost than ...
on Friday July 03, 2026 @07:53PM (#48781393)
yes it is 213 tok/s single stream (so per user)
Re:GLM5.2 on AMD MI355X at 2626 tok/s/node at over 2x lower cost than ...
on Friday July 03, 2026 @07:52PM (#48781390)
hi i work at wafer. no the margins are lower averaging at about ~40%. utilization is one of the highest order bits in determining margins here, yes.
Show HN: Wafer – Profile, inspect assembly, and iterate on CUDA within your IDE
Hi HN, I’m Emilio. We’re launching the Wafer extension for the popular IDEs (VS Code, Cursor and Antigravity).Wafer exists to make performance engineers more efficient. Most of the work perf engs do is extracting signal and turning it into the next experiment. You spend hours per kernel doing interpretation and bookkeeping: which counters matter, what changed, what hypothesis you’re testing,...
Show HN: GPU Profiling That's Useful in 60 Seconds
https://www.keysandcaches.com/
Hey HN! We're building a profiler for ML inference that actually shows what's happening at the hardware level without having to manually parse through flame graphs, or set up nsys and ncu.The problem: Current ML profilers either dump too much data (torch.profiler) or abstract away the details you need. You can't see why your model is actually slow - is it memory bandwidth? Kernel launch...
Show HN: We made PyTorch profiling usable for ML engineers
If you've ever tried to profile PyTorch or other Python code, you know that it can be painful to setup right. Nsight profiler is great, but it feels like using a sledgehammer to hang a picture frame when you're coming from a Python background.We're building Chisel to solve this. Instead of wrestling with nvtx events and complex CUDA interfaces, you just add decorators to the functions you want...
Chip Benchmark: Hardware-Centric Performance Insights for AI Workloads
Re:ChipBenchmark: Open-Source Benchmarking for LLM Performance Across ...
on Monday July 14, 2025 @10:56PM (#44567461)
^ We currently just have llama3.1-8b, so we'll be working on adding more models across more hardware options!
ChipBenchmark: Open-Source Benchmarking for LLM Performance Across Hardware
Re:ChipBenchmark: Open-Source Benchmarking for LLM Performance Across ...
on Monday July 14, 2025 @10:55PM (#44567455)
We just launched Chip Benchmark, an open-source tool for hardware-centric benchmarking of open-weight LLMs across accelerators like NVIDIA A100/H100/L40S and AMD MI300X. It measures throughput, latency, and time-to-first-token with transparent scripts and an interactive web dashboard—making apples-to-apples comparisons easier.
We're actively welcoming contributions, new hardware support, and benchmark requests.
Repo here: https://github.com/Herdora/chip-benchmark Dashboard: https://herdora.com/benchmark
Feedback and contributions welcome! We made it super easy to add other architectures by including the script we used for benchmarking.
Using AMD MI300X for High-Throughput, Low-Cost LLM Inference
Re:Profile CUDA kernels with one command, zero GPU setup
on Friday July 04, 2025 @05:52AM (#44463018)
We've been doing lots of GPU kernel profiling and optimization on cloud infrastructure, but without local GPU hardware, that meant constant SSH juggling: upload code, compile remotely, profile kernels, download results, repeat. Or, work entirely on cloud which is expensive, slow, and annoying. We were spending more time managing infrastructure than writing the kernels we wanted to optimize.
So we built Chisel: one command to run profiling commands on any kernel. Zero local GPU hardware required.
Next up we're planning to build a web dashboard for visualizing results, simultaneous profiling across multiple GPU types, and automatic resource cleanup. But please let us know what you would like to see in this project.
Available via PyPI: pip install chisel-cli
Github: https://github.com/Herdora/chisel
We're actively developing and would love community feedback. Feature requests and contributions always welcome!
Profile CUDA kernels with one command, zero GPU setup
Show HN: Profile GPU Kernels with One Command, Zero GPU Setup
https://github.com/Herdora/chisel
We've been doing lots of GPU kernel profiling and optimization on cloud infrastructure, but without local GPU hardware, that meant constant SSH juggling: upload code, compile remotely, profile kernels, download results, repeat. We were spending more time managing infrastructure than writing optimized kernels.So we built Chisel: one command spins up a cloud instance, syncs your GPU code, runs...
Show HN: Chisel – Profile GPU Kernels Without a GPU (Nvidia and AMD)
https://github.com/Herdora/chisel
We built Chisel to make GPU kernel profiling hardware-free. It lets you run chisel profile kernel.cu and get full Nsight/Ncompute or rocprofv3 reports without a GPU needed.It spins up remote H100, L40S, or MI300X machines (via DigitalOcean for now, but gonna expand backends soon), runs your code, and gives you back detailed traces (kernel timings, memory transfers, API calls, etc). Everything...
Re:Ask HN: Is anyone using AMD GPUs for their AI workloads?
on Wednesday June 25, 2025 @01:34AM (#44373918)
oh yeah, in my experience anything below ROCm6.x really sucks.
I tried to run qwen2.5-32B on ROCm5.x and it was running at <15tok/s lol.
Have you tried running any sort of LLM inference on your MI25, or what NN workloads are you running?
Ask HN: Is anyone using AMD GPUs for their AI workloads?
^ title. I've been renting MI300Xs coz they are cheaper than H100s and my experience has been generally OK (smoother than i expected based on people shitting on AMD so much online). ROCm 6.x seems decent out of the box now, and I'll happily spend 30 more minutes setting up my GPU if it means 20% cheaper. that being said, it's still annoying to run inference for LLMs on AMD's hardware (e.g. You...
Show HN: Chisel – GPU development through MCP
https://github.com/Herdora/chisel
We've been running lots of experiments on AMD MI300Xs. The price/performance is compelling compared to NVIDIA, and ROCm is finally usable.But, no local hardware means constant SSH juggling. Upload code, compile remotely, run rocprof, download results, repeat. We were spending more time managing infrastructure than optimizing kernels.We're currently competing in the GPU mode kernel optimization...
Show HN: Chisel – Profile AMD MI300X kernels locally
https://github.com/Herdora/chisel
We've been running lots of AMD MI300X experiments lately. The price/performance is compelling compared to NVIDIA, and ROCm keeps getting better.But, no local AMD hardware means constant SSH juggling. Upload code, compile remotely, run rocprof, download results, repeat. We were spending more time managing infrastructure than optimizing kernels.DigitalOcean's AMD GPU droplets ship with ROCm...
Re:Show HN: Chisel – AMD GPU development that feels local but runs in ...
on Sunday June 22, 2025 @01:01PM (#44348413)
correct github link: https://github.com/Herdora/chisel
Show HN: Chisel – AMD GPU development that feels local but runs in the cloud
https://pypi.org/project/chisel-cli/0.1.2/
We've been running lots of AMD MI300X experiments lately. The price/performance is compelling compared to NVIDIA, and ROCm keeps getting better.But, no local AMD hardware means constant SSH juggling. Upload code, compile remotely, run rocprof, download results, repeat. We were spending more time managing infrastructure than optimizing kernels.Context: DigitalOcean's AMD GPU droplets ship with...
Crackd – unbiased technical skill comparison among students
Re:Crackd – unbiased technical skill comparison among students
on Sunday May 11, 2025 @07:15PM (#43958098)
We built Crackd to solve a fundamental problem: helping students understand where they stand technically, and helping them learn how to become the absolute best.
Technical students have no reliable way to know how good they are. Grades don't work because they vary by professor and university. Getting internships can boil down to arbitrary indicators, and online advice is filled with contradictions.
Athletes have it better. A college football player knows exactly where they stand nationally (e.g. https://www.espn.com/college-sports/football/recruiting/play...).
More importantly, they know what the best players do differently: their training methods, their practice routines, the specific skills they're mastering.
CS students are flying blind by comparison. They face a collapsing job market with nothing but twitter threads giving them contradictory advice.
We built Crackd to fix this. Students judge pairs of technical students, and choose the person they believe is more “cracked.” After enough comparisons, the Elo algorithm generates a leaderboard. Humans are terrible at absolute judgments but good at comparisons. You can't reliably say if someone is a 7/10 developer, but you can usually tell who has a better profile based on technical achievement.
What matters about Crackd is letting other students see precisely what the top students do differently. Their projects, internships, skills, school clubs, etc. This is the information that turns rankings from a leaderboard into a roadmap for ambitious students.
For the time being, Crackd is restricted to college students as we refine our platform and build our initial community.
We're attempting to cover as many dimensions of technical skill as possible, covering projects, experiences, and including a miscellaneous section so students can add whatever info they feel showcases their best work. But we'd love your feedback on what other dimensions we should include to create the most comprehensive assessment of technical capability. We would love to hear from you at [email protected]
Show HN: Molecule-Rs – A Fast Protein Visualization Engine Written in Rust
https://molecule-rs.vercel.app/
rust-based molecular structure toolkit with webassembly integration for high-performance 3d visualization and analysis in the browser. parses PDB files, processes molecular structures, and renders interactive 3d models with near-native performance through wasm. built for researchers and developers who need efficient, technical visualization tools without the overhead of shitty javascript-only...