ml for systems at nvidia opinions are my own
- Tweets 208
- Following 385
- Followers 39
- Likes 1,696
what is @chatgpt suddenly making mermaid charts for all explanations (i support!) but not actually rendering them in the UI (either on the web or the ios app)?
cc @thsottiaux
1
17
this is not new. nvidia’s hiring has been very team-specific and non-leetcode focused for a while now
11
1,024
"We asked it, open-endedly, to find and solve an unsolved cipher, and it chose this one."
Did you guys do multiple trials to conclude that there was intuition behind picking that cipher (or something similar) and not just luck? Awesome work btw!
1
6
2,606
I don't think AI labs are using the macs for their gpus? Looks like mainly a way to get macos in the hands of computer use agents for RL? DGX spark is definitely not compatible for that market...
1
3
289
datacenter and graphics cards are all fighting for the same wafer capacity so i don’t see how this can be fixed anytime soon
9
200
I dont know what big tech you are looking at but in India, big tech salaries for fresh undergrads start at atleast $20-25/hr
1
37
3,018
asking open source to not be regulatory captured is not the same as saying everything should be open sourced
2
322
They literally used EDA tools. If your argument is that they were able to do this with open-source EDA tools, I would say the results would have been much better with SOTA (closed) EDA tools. Saying EDA is cooked based on this is like saying Python interpreter is cooked because agents are crazy good at using it!
1
1
92
7,493
what you call greed is just the difference in price elasticity between apple’s customers and the dram suppliers’ customers
5
696
I think this is slightly misleading. There are certain things that need so much context that you can only give to siri (apple) and not chatgpt. I really hope siri can do those things well because those are not possible (at least yet) with any other app
1
1
25
1,910
From my understanding so far, Apple‘s approach is more on the training side. NVIDIA is primarily a hardware platform. Even though they optimize the software stack, the main selling point is that you can run anything. I don’t see a reason why they won’t adopt the hardware to rely less on DRAM and more on other storage options like NAND, SSD and NVMe if the industry moves in that direction.
1
6
1,082
NYC is land-constrained. If you don’t live in your second home worth more than $5 million, pay extra or sell, which will reduce the said land-constrain. Either sounds like a good outcome for the city, no?
2
11
2,212
I used claude code to both create my vault as well as managing it. It automatically processes my articles, tracking projects, extracting concepts and creating concept maps: github.com/leakyfilter/obsid…
3
351
1. Triton has a reasonable (~80%) gap in performance as compared to the nvidia stack. This is expected since Triton's focus is more on developer productivity. Some sources:
- x.com/stuart_sul/status/1990…
- modular.com/blog/democratizi…
2. GPUs remain un-defeated when it comes to experimenting new architectures (think funky control flow, custom ops, non-standard data layouts). Unless you think the current transformer is enough for AGI, TPUs by design will only be competitive for scalable deployment after the architectural improvements have been agreed upon and factored into designing the chips.
(4/6) Existing libraries are slow to adapt. NCCL’s main API is bottlenecked by channeling and host-side comms, NVSHMEM relies on the slow st and ld instructions for intra-node RMA, Triton Distributed is slow on H100s and doesn’t work on Blackwell, and many existing hand-tuned kernels are optimized for specific shapes only.
2
847