winwang

19 ms·

301.
by winwang 2y ago
Going through the application helps provide clarity to the founders themselves as well.
302.
by winwang 2y ago
Yep, although I do like not having clashing names. But I mean... somehow every other language solved this.
303.
by winwang 2y ago
something something lies are eventual consistency?
304.
by winwang 2y ago
Great context for us GPGPU fans! Do you think the L4 should have similar perf? Considering it has similar features and specs to L40, and I doubt the lower 300 GB/s bandwidth woudl be an issue.
305.
by winwang 2y ago
I think the author wanted to compare with a similar base algorithm. They should have stated that, however, instead of implying that a basic C++ algo would typically take that long. Something else to consider is cost of cloud hardware, thoug
306.
by winwang 2y ago
Curious, do you guys use Triton? I was surprised to find out PyTorch slowdown is palpable even compared to basic CUDA kernels.
307.
by winwang 3y ago
Well said! Have not seen the tension expressed in this way before.
308.
by winwang 3y ago
Do you think SQLite performance on user devices changes anything for UX/DX? i.e. if SQLite were 10x faster than today on user devices in the future.
309.
by winwang 3y ago
If I had to guess, it's because its a hardware change... like you're doing better at implementing software on physics :P
310.
by winwang 3y ago
Was mostly just checking to see if someone had already tested GPGPU, though I know some HPC labs like the MI300A. While I am starting a business, I'm not at the point of shipping software just yet (I wish!). Will definitely keep you in
311.
by winwang 3y ago
Nice! I'm pretty interested in GPGPU applications and MI300A, but I'm also just glad for more competition. Love that you hit up the LocalLLaMa sub. Do you know if anyone's tested CuPy stuff on MI300X?
312.
by winwang 3y ago
How's the overall software support for MI300 series? The hardware itself looks great. (also, +100 to valuing honesty and transparency)
313.
by winwang 3y ago
Fun fact, FFS (and its generalization, FNS) is in CUDA: https://docs.nvidia.com/cuda/cuda-math-api/index.html#group_... Another nice CUDA hardware intrinsic I like is log2.
314.
by winwang 3y ago
OP just needs a couple minutes of VPN
315.
by winwang 3y ago
Thank you for evangelizing -- this is amazing and kinda hilarious. Does it count as biohacking? lol.
316.
by winwang 3y ago
LLM the Ultimate Heuristic?
317.
by winwang 3y ago
+1. There are useful instructions besides increasing width, which is in fact less interesting. That being said, I'd have to look into how this was measured.
318.
by winwang 3y ago
> Boss and subordinate typically settle into a routine that is not really satisfactory but, aside from periodic clashes, is otherwise bearable for them. This sounds like a game theoretic setup. I'm annoyed that, in this scenario, th
319.
by winwang 3y ago
Are you seriously going to name "obvious" functions like `[x + 1 for x in xs]`?
320.
by winwang 3y ago
If you use the made up averaging scheme of ((5^p+50^p)/2^p)^(1/p), 25 is the limit as p -> inf (i think). ...Because it just simplifies to 50/2 lol.
321.
by winwang 3y ago
I'd say it depends on how confident/aggressive someone was. Perhaps OP didn't word it the best way, but it's also good to pause and reflect?
322.
by winwang 3y ago
How's the single core-to-core bandwidth?
323.
by winwang 3y ago
The author also just gave a talk on this at Nvidia GTC! https://www.nvidia.com/gtc/session-catalog/?search=S62506&ta...
324.
Leveraging Spark 3 and Nvidia's GPUs to Reduce Cloud Cost by Up to 70%(medium.com)
3 points by winwang 3y ago|1 comments
325.
by winwang 3y ago
(2019 paper) I find the idea of smarter register spilling interesting, especially since the register file is one of the most power-hungry elements of modern GPUs.
326.
RegDem: Increasing GPU Performance via Shared Memory Register Spilling(arxiv.org)
2 points by winwang 3y ago|1 comments
327.
by winwang 3y ago
I wonder if we should have a kind of "hidden type system", where we still take advantage of having a single type system to reason about, but the extra-specific "weird-ish" types can be hidden, almost like private variabl
328.
by winwang 3y ago
While network overhead always exists, the goal would be take advantage of the seemingly 10x performance headroom on a GPU as compared to a CPU. Not to mention, GPUs are getting more and more HBM capacity. Of course, you're right to won
329.
Show HN: Beating Microsoft's Garnet KV Store in less than 300 lines of code(github.com)
3 points by winwang 3y ago|2 comments
330.
by winwang 3y ago
Are you suggesting that it's essentially too easy for a dev to just set and forget? That's a pretty interesting viewpoint. Not sure how any BaaS could solve that human factor.

More ›