27 comments

[ 3.5 ms ] story [ 50.3 ms ] thread
The US attempt to slow down China's technological development succeeds on the basis of preventing China from directly following the same path, but may backfire in the sense it forces innovation by China in a different direction. The overall outcome for us all may be increase efficiency as a result of this forced innovation, especially if Chinese companies continue to open source their advances, so we may in the end have reason to thank the US for their civilisational gate keeping
You mean, thank the US for their FAILED "civilizational" gate keeping.
Peaceful competition is a good thing. It's better than a unified one world government throttling everybody.
Does someone know if there's some equivalent of those engineering/research blogs for Chinese companies?

I used to follow the ones from Western companies, but honestly, after some point in time, I would like to see some cases from what I consider is a good benchmark for everyone that does not work in FAANG in terms of engineering.

Key paragraph:

> However, a small handful of models such as Alibaba’s Qwen and DeepSeek are most popular for inference, with most other models only sporadically called upon. This leads to resource inefficiency, with 17.7 per cent of GPUs allocated to serve only 1.35 per cent of requests in Alibaba Cloud’s marketplace, the researchers found.

Its easy enough for a a well resourced entity to take a pre trained model and deploy it on new hardware to save on the NVDA tax. It's far less likely for research and model training to happen outside the mature NVDA ecosystem.
Alibaba Cloud claims to reduce Nvidia GPU used for serving unpopular models by 82% (emphasis mine)

> 17.7 per cent of GPUs allocated to serve only 1.35 per cent of requests in Alibaba Cloud’s marketplace, the researchers found

Instead of 1192 GPUs they now use 213 for serving those requests.

So 82% of 17.7%?

14.5% is worth a raise at least. But it’s still misleading.

I don't think that's what this is saying, isn't it that 100 - ~82 = 17.7% ?
That is a confusing coincidence, but no.

> Reserving full GPU instances for these models leads to allocating 17.7% of our GPUs to serve only 1.35% of requests

> Deployment results show that Aegaeon reduces the number of GPUs required for serving these models from 1,192 to 213, highlighting an 82% GPU resource saving.

82% of their CPUs were serving 98.6% of all traffic. If they reduced the cluster size, they got it to 96.2% of their CPUs serving 98.6% of their traffic. If they reallocated those, which is more likely, then 96.8% of their CPUs are serving 98.6% of all requests, or around 17% more capacity for popular requests on the same hardware.

They are working with tiny models. Not sure how well it'd scale to bigger models (if at all).
To what extent is this practice applicable to other loads?
This is for platforms that serve many different models, most of which have very low usage. e.g. huggingface, civitai
Sounds like this virtual GPU is a separate scheduler. I wonder what kind of latency is introduced by marshaling all that data around.
How feasible is that in an horizon of 5 years new optimized "equations" will cut the need for more GPUs?
Lots of shareholders here, move along, there is nothing to read
Sounds like they stopped doing something stupid.
Would this make cloud providers running low volume fine-tuned models more economically viable?
Is this another nail in the gpu/ai stock market bubble coffin?
I still think nVidia has the most to loose in the AI race, optimisations like this will continue coupled with better ASIC's.
> Distributed executor: Inference engines support model parallelism via distributed executors (e.g., Ray [32] and NCCL [9]), whose initialization takes tens of seconds.

I mean, it really shouldn't take tens of seconds for those initialization(s) to occur. There's no good fundamental reason that it should take that long. It's just bloat.

Does anyone know how their KV cache sync mechanism compares to newer P2P communication layers like nixl, uccl p2p, etc.?

The authors mention that NCCL and Ray initialization were too slow (see quote below), but from the description it sounds like they’ve reimplemented a layer that’s increasingly being standardized by frameworks like nixl and uccl.

> Distributed executor: Inference engines support model parallelism via distributed executors (e.g., Ray [32] and NCCL [9]), whose initialization takes tens of seconds.