I’ve talked to the authors of this, it’s a very interesting project. GPU memory space used to be the limitation but the latest generations of GPUs have enormous shared memory capacity and need something like SiriusDB to manipulate and prepare data in-place before the AI algorithms get to work.
Some of the price performance improvement that is quoted is due to using $ from different cloud providers - eg a GH200 in Lambda Labs costs $1.5/hr, but the closest equivalent in AWS (p5.4xlarge) costs $6.88/hr. Which means, ~4.5x of the price performance benefits is not real ...
From their Rethinking Analytical Processing in the GPU Era paper,
> Sirius builds on GPU libraries such as
libcudf [6], RMM [14], and NCCL [11], reusing optimized implemen-
tations of core relational operators like joins, filters, aggregations,
and data shuffle. Thanks to its modular design, Sirius also allows
developers to easily switch the operator implementation between
these GPU libraries and custom CUDA kernels.
I wonder if the various other CUDA translation layers (ZLUDA, SCALE, HIP) can host this?
It'd be so nice to see a little more foothold for Vulkan in this space. There's some good work in AI for Vulkan, it's becoming quite capable. But for databases & GPGPU, it doesn't seem like there are good rallying points.
I expect whatever does eventually emerge will perhaps likely be based on Substrait too! What an awesome common grounds thats emerged for data processing work.
One downside is that you're paying for the GPU whether you're fully using it or not. It takes big queries to saturate a GH200, and if you're only using 10% of the capacity of the GPU it doesn't really matter that it's 10x faster.
In a typical company you'll have jobs, some scheduled, some ad-hoc, at a range of sizes. Most of them won't be cost-effective to run on a GPU instance, so you need a scheduling layer that estimates the size of the job and routes it to the appropriate hardware. But now what if the job is too big to run on your GPU machine? Now we either have to scale up our GPU cluster or retry it on our more flexible CPU cluster.
And this all assumes that your jobs can be transparently run across different executors from a correctness and performance standpoint.
There are niches where this makes sense (we run the same 100TB job every day and we need to speed it up), as well and large and sophisticated internal infra teams that can manage a heterogenous cluster + scheduling systems, but it's not mass-market.
GPU databases can run a small subset of production workloads in a narrow combination of conditions.
There are plenty of GPU databases out there: mapD/OmniSci/HeavyDB, AresDB, BlazingSQL, Kinetika, BrytlytDB, SQReam, Alenka, ... Some of them are very niche, and the others are not even usable.
I’ve talked to the authors of this, it’s a very interesting project. GPU memory space used to be the limitation but the latest generations of GPUs have enormous shared memory capacity and need something like SiriusDB to manipulate and prepare data in-place before the AI algorithms get to work.
Its sitting at the top in clickbench .Pretty cool https://benchmark.clickhouse.com/#system=-&type=-&machine=-c...
Wow! Now I got interested on reading the paper, thanks
improvement over DuckDb is kinda marginal (44%)
44% is not marginal. "Marginal" is what perceived by seller and buyer as negligible and it tops at 5%.
its marginal compared to promised 10x improvement.
It really is a SeriousDB
There is also a recent blog post about this: https://developer.nvidia.com/blog/nvidia-gpu-accelerated-sir...
Some of the price performance improvement that is quoted is due to using $ from different cloud providers - eg a GH200 in Lambda Labs costs $1.5/hr, but the closest equivalent in AWS (p5.4xlarge) costs $6.88/hr. Which means, ~4.5x of the price performance benefits is not real ...
Reminds me of Uber's AresDB: https://www.uber.com/blog/aresdb/
I wonder if the benefit is primarily for transactional vs analytical queries
it'll be purely for analytical queries
From their Rethinking Analytical Processing in the GPU Era paper,
> Sirius builds on GPU libraries such as libcudf [6], RMM [14], and NCCL [11], reusing optimized implemen- tations of core relational operators like joins, filters, aggregations, and data shuffle. Thanks to its modular design, Sirius also allows developers to easily switch the operator implementation between these GPU libraries and custom CUDA kernels.
https://arxiv.org/abs/2508.04701
I wonder if the various other CUDA translation layers (ZLUDA, SCALE, HIP) can host this?
It'd be so nice to see a little more foothold for Vulkan in this space. There's some good work in AI for Vulkan, it's becoming quite capable. But for databases & GPGPU, it doesn't seem like there are good rallying points.
I expect whatever does eventually emerge will perhaps likely be based on Substrait too! What an awesome common grounds thats emerged for data processing work.
Sounds amazing; what are the downsides that a company needs to consider? Memory bottlenecks or storage bus access?
One downside is that you're paying for the GPU whether you're fully using it or not. It takes big queries to saturate a GH200, and if you're only using 10% of the capacity of the GPU it doesn't really matter that it's 10x faster.
In a typical company you'll have jobs, some scheduled, some ad-hoc, at a range of sizes. Most of them won't be cost-effective to run on a GPU instance, so you need a scheduling layer that estimates the size of the job and routes it to the appropriate hardware. But now what if the job is too big to run on your GPU machine? Now we either have to scale up our GPU cluster or retry it on our more flexible CPU cluster.
And this all assumes that your jobs can be transparently run across different executors from a correctness and performance standpoint.
There are niches where this makes sense (we run the same 100TB job every day and we need to speed it up), as well and large and sophisticated internal infra teams that can manage a heterogenous cluster + scheduling systems, but it's not mass-market.
The website claims it’s 10x cheaper (“10x faster on same hardware costs”) and implements SQL execution.
I don’t understand why GPU saturation is relevant. If it’s 10x cheaper, it doesn’t matter if you only use 0.1% of the GPU, right?
Correctness shouldn’t be a concern if it implements SQL.
Curious for some more details, maybe there’s something I’m missing.
GPU databases can run a small subset of production workloads in a narrow combination of conditions.
There are plenty of GPU databases out there: mapD/OmniSci/HeavyDB, AresDB, BlazingSQL, Kinetika, BrytlytDB, SQReam, Alenka, ... Some of them are very niche, and the others are not even usable.