Wednesday, September 16th 2026

Axelera Launches Europa AI Processing Unit (AIPU) With Up to 629 TOPS

Axelera AI has formally launched Europa, its new AI Processing Unit architecture first announced back in October last year. The chip is now available in three form factors: as a chip for custom board designs, a half-height, half-length PCIe card (Edge 232p), and a full-height full-length PCIe card (Server 250p). The Europa AIPU has eight second-generation AI processing cores, 16 RISC-V vector processing cores, 128 MB of L2 SRAM, 200 GB/s DRAM bandwidth, PCIe 4.0 x4 interface and an onboard video decoder. The chip can reach 629 TOPS at a 45 W TDP, with INT4, INT8, and INT16 precision support. The Europa architecture targets inference workloads including agentic systems, vision-language models, generative AI, and computer vision and is the direct successor of Axelera Metis.

The Edge 232p is a single-chip HHHL card with up to 629 TOPS, reaching 83.4 tokens per second on Qwen 30B. It's available in validated configurations inside a Dell XE5 and a Supermicro 111AD. The Server 250p steps up to four Europa AIPUs for over 2,500 TOPS combined, delivering up to 4,205 tokens per second on Qwen3 8B at up to 32.1 tokens per second per watt. In a press release, Axelera says it has a sales pipeline that exceeds $1.5 billion and more than 600 customers across the globe. The company is now working on its first chiplet, the Titania, targeting rack-mount servers and supercomputers.
Source: Axelera
Add your own comment

7 Comments on Axelera Launches Europa AI Processing Unit (AIPU) With Up to 629 TOPS

#1
kondamin
600tops isn't exactly a lot of tops, I also doubt they will be cheap
Posted on Reply
#2
mashie
The 3dfx Voodoo 6000 is back!
Posted on Reply
#3
Happyguynpc
If you ask an LLM to tell a dirty joke about that thing, your name will be put on the Interpol list.
Posted on Reply
#4
igormp
kondamin600tops isn't exactly a lot of tops, I also doubt they will be cheap
It actually is a lot of tops, given that it's INT8 tops at 45W, as per their specs:

axelera.ai/hubfs/Product%20Brochure/Axelera%20Edge%20232p.pdf?hsLang=en

For comparison, here are some Nvidia desktop GPUs:
- 5090 does 838 INT8 TOPs @ 575W
- 4090 does 660 INT8 TOPs @ 450W
- 5080 does 450 INT8 TOPs @ 360W

Those 32/64GB of memory are also nice, with the downside being the bandwidth being capped at 200GB/s, which makes it closer to a RTX 4050 mobile or a Jetson AGX Orin in that regard. Still, it's an interesting product depending on the price, and would do good for models that can be quantized to INT8 without much issues and have high arithmetic intensity (instead of being memory-bound).

And I doubt it'll be cheap as well, given that the previous single-chip model with 16GB of memory and 214 INT8 TOPs went for ~$700, and the 4x chip with 16GB and 856 INT8 TOPs is over 2 grand:

store.axelera.ai/collections/
Posted on Reply
#5
Warigator
nice, good, but doesn't seem very cutting-edge
Posted on Reply
#6
Grimling
Warigatornice, good, but doesn't seem very cutting-edge
What is even the point of this? 128 MB of RAM?
Is this a joke?
What LLM fits into that?
And then 8 GB/s over that PCIe 4.0 x4 interface?
Lets say you have 512 GB of RAM and load a 300B LLM into it, this is gonna be extremely slow and unusable, because it is limited by the 8 GB/s of RAM-Speed!
80 tokens/s, like more 0,8 tokens/s!
My notebook looks like an amazing AI-Machine compared to this.
Posted on Reply
#7
igormp
GrimlingWhat is even the point of this? 128 MB of RAM?
Is this a joke?
128MB of SRAM. It has 32/64GB of RAM.
Thsi is not meant to do hybrid offloading from RAM. Either fit the model entirely in there, or don't use it.
And its main use case is not LLMs, given how constrained in memory bandwidth it is.
Posted on Reply
Nov 16th, 2026 03:56 CST change timezone

New Forum Posts

Popular Reviews

TPU on YouTube

Controversial News Posts