Qualcomm Linux Driver Prepares For New "AIC080" Lower-Cost Cloud AI Accelerator
([Hardware] 110 Minutes Ago
Qualcomm Cloud AI 80)
- Reference: 0001496017
- News link: https://www.phoronix.com/news/Qualcomm-Cloud-AI-80-AIC080
- Source link:
The Qualcomm Cloud AI 100 accelerator caters to a variety of edge-to-cloud industries. While the Qualcomm Cloud AI hardware isn't talked about as much as the AI accelerators from other vendors, there is [1]the QAIC driver within the mainline Linux kernel for supporting the Cloud AI 100 along with associated [2]open-source compiler and user-space stack . It turns out the Qualcomm Cloud AI family is growing with a Cloud AI 80 "AIC080" accelerator coming to market at a lower-cost.
I haven't seen the Qualcomm AIC080 "Cloud AI 80" talked about publicly or mentioned prominently in any Google search results. But some interesting patches crossed the Linux kernel mailing list a few minutes ago and after some digging through the Qualcomm website was able to turn up some specs.
The [3]accel/qaic: Add AIC080 support patch series is the first time I'm seeing this new Qualcomm Cloud AI 80 product mentioned. The patch cover letter calls out this new AIC080 product quite clearly:
"Add basic support for the new AIC080 product. The PCIe Device ID is 0xa080. AIC080 is a lower cost, lower performance SKU variant of AIC100. From the qaic perspective, it is the same as AIC100."
So from the QAIC kernel driver side this Cloud AI 80 support is just a few lines of code for adding the new device ID bits and otherwise following the same driver paths as the AIC100.
The patches do not provide any other details around this cheaper, cut-down Qualcomm Cloud AI 80 accelerator hardware. While not prominently noted in Google search results or any press/news releases, on [4]this Qualcomm page the Qualcomm Cloud AI 80 Ultra and Cloud AI 80 Standard are mentioned.
The Qualcomm Cloud AI 80 Standard is a PCIe HHHL accelerator card rated for 75 Watts and able to deliver up to 190 TOPS of INT8 performance or up to 86 TFLOPS of FP16 performance. This is down from the 350 TOPS INT8 performance on the Qualcomm Cloud AI 100 Standard or up to 175 TFLOPS FP16. Both the AI 80 Standard and AI 100 Standard have 16GB of RAM. The Cloud AI 80 Ultra is a cut-down version of the AI 100 Ultra in the PCIe FH3/4L form factor. Both the 80 Ultra and 100 Ultra have a 150 Watt TDP and 128GB of LPDDR4x memory @ 548GB/s but the AI 80 Ultra delivers up to 618 TOPS INT8 and up to 222 TFLOPS FP16 while the AI 100 Ultra can achieve 870 TOPS INT8 and up to 288 TFLOPS FP16 performance.
I haven't seen any pricing information yet on the Qualcomm Cloud AI 80 Ultra/Standard PCIe accelerator cards, but at least the open-source Linux driver support is on the way.
[1] https://www.phoronix.com/news/Qualcomm-QAIC-User-Space
[2] https://www.phoronix.com/news/Qualcomm-QAIC-User-Space
[3] https://lore.kernel.org/dri-devel/20241004195209.3910996-1-quic_jhugo@quicinc.com/
[4] https://www.qualcomm.com/products/technology/processors/cloud-artificial-intelligence/cloud-ai-100
I haven't seen the Qualcomm AIC080 "Cloud AI 80" talked about publicly or mentioned prominently in any Google search results. But some interesting patches crossed the Linux kernel mailing list a few minutes ago and after some digging through the Qualcomm website was able to turn up some specs.
The [3]accel/qaic: Add AIC080 support patch series is the first time I'm seeing this new Qualcomm Cloud AI 80 product mentioned. The patch cover letter calls out this new AIC080 product quite clearly:
"Add basic support for the new AIC080 product. The PCIe Device ID is 0xa080. AIC080 is a lower cost, lower performance SKU variant of AIC100. From the qaic perspective, it is the same as AIC100."
So from the QAIC kernel driver side this Cloud AI 80 support is just a few lines of code for adding the new device ID bits and otherwise following the same driver paths as the AIC100.
The patches do not provide any other details around this cheaper, cut-down Qualcomm Cloud AI 80 accelerator hardware. While not prominently noted in Google search results or any press/news releases, on [4]this Qualcomm page the Qualcomm Cloud AI 80 Ultra and Cloud AI 80 Standard are mentioned.
The Qualcomm Cloud AI 80 Standard is a PCIe HHHL accelerator card rated for 75 Watts and able to deliver up to 190 TOPS of INT8 performance or up to 86 TFLOPS of FP16 performance. This is down from the 350 TOPS INT8 performance on the Qualcomm Cloud AI 100 Standard or up to 175 TFLOPS FP16. Both the AI 80 Standard and AI 100 Standard have 16GB of RAM. The Cloud AI 80 Ultra is a cut-down version of the AI 100 Ultra in the PCIe FH3/4L form factor. Both the 80 Ultra and 100 Ultra have a 150 Watt TDP and 128GB of LPDDR4x memory @ 548GB/s but the AI 80 Ultra delivers up to 618 TOPS INT8 and up to 222 TFLOPS FP16 while the AI 100 Ultra can achieve 870 TOPS INT8 and up to 288 TFLOPS FP16 performance.
I haven't seen any pricing information yet on the Qualcomm Cloud AI 80 Ultra/Standard PCIe accelerator cards, but at least the open-source Linux driver support is on the way.
[1] https://www.phoronix.com/news/Qualcomm-QAIC-User-Space
[2] https://www.phoronix.com/news/Qualcomm-QAIC-User-Space
[3] https://lore.kernel.org/dri-devel/20241004195209.3910996-1-quic_jhugo@quicinc.com/
[4] https://www.qualcomm.com/products/technology/processors/cloud-artificial-intelligence/cloud-ai-100
phoronix