Skip to content

Qualcomm Opens Dragonfly AI200 Inference Trials Through Cirrascale

News
Black data center accelerator card with a finned aluminum heatsink beside an open server slot on a test bench.
Qualcomm and Cirrascale have opened trials of the Dragonfly AI200. The system is built to run AI models in data centers. Some Cirrascale clients can now try it with their own models and tasks. Cirrascale will host the racks, so those clients can test the gear through the cloud.
The firms opened the trials on November 17 at the SC26 trade show in Chicago. They will help clients run tests and check that their code works with the chips. Teams can also try out a small version of a planned service. The firms say teams can move on to live use after the trial.

Why the Dragonfly AI200 has more room for data

An AI model needs a place to store what it learned in training. It also needs room to hold data as it works. Chat models, for instance, save parts of the work they have done on past words to save time on the next ones. Long chats can take up more space. So can a spike in the number of people who use the model at once.
Qualcomm’s AI200 specs list 768 GB of LPDDR memory per card. A full rack has about 43 TB. More room on each card can help teams run a large model or serve more users at once. The speed at which the chips can read all that data matters too. The trials give clients a way to test both with their own workloads.
The cards share data over PCIe links. The racks can also link up over Ethernet. Qualcomm uses liquid to cool the gear and rates the rack at 140 kW. That figure helps a host plan how much power it must be able to supply. To work out the cost of one reply, a team needs to measure how much power the setup draws as it serves real requests.

What clients can learn from the trials

Each test can match the kind of work a team plans to do. A chat or voice tool needs to reply fast when many people use it at the same time. A team that sorts long files may care more about how many pages it can get through in an hour. Both teams can track cost as well as speed. Qualcomm says the system can also run models that work with more than one type of input. A team might test how fast it responds to a mix of photos and text.
Cirrascale says it charges a flat rate for its cloud services. It also says it has no fees for data sent to or from its cloud. The firms have yet to name trial clients or say what it costs to use an AI200 rack. They gave no date for broader access.
The trial adds a new way to try a chip built for AI inference. Google, for its part, has added a TPU with more memory. Qualcomm’s chip roadmap also lists the AI250 and AI300. For now, the AI200 trial is where Cirrascale clients can see how much it costs to run their own models on Qualcomm’s rack.
loading

Loading