Intel has taped out its third generation Gaudi3 design as a key driver for AI in the data centre.
The company is already sampling chiplet devices on 3nm process technology to data centre customers, and Intel sees AI silicon as a key driver with 1.8nm devices also currently taping out. Intel is not currently releasing details of the chip, process technology for product combination with other processor chiplets, but last year this was described as a 5nm shrink from the 7nm process used for the previous Gaudi2 chip.
The company yesterday doubled its estimate of the total addressable market (TAM) for data centre silicon from last year to $110bn, or more than a fifth of the total market. AI silicon is the key driver for this, and Intel sees an opportunity to add AI accelerators alongside its Xeon processors.
“When we talk about compute demand, we often look at the TAM through the lens of CPU units,” said Sandra Rivera, executive vice president and general manager of the Data Center and AI Group at Intel. “However, counting sockets does not fully reflect how silicon innovations deliver value to the market. Today, innovations are delivered in several ways, including increased CPU core density, the use of accelerators built into the silicon and the use of discrete accelerators.”
Rivera cited a couple of specific reasons for the increased TAM. With the integration of accelerator compute and advanced GPU offerings into its data center business after a restructuring last year, Intel says it is better positioned to serve a wider swath of customers.
“Customers want portability in their AI workloads. They want to build once and deploy anywhere,” Rivera said as she showed samples of the fifth generation Xeon processor (above) code named Emerald Rapids. “As we continue to deliver heterogenous architectures for AI workloads, deploying them at scale will require software that makes it easy for developers to program and a vibrant open and secure ecosystem to flourish.”
The acquisition of SYCL expert CodePlay Software in Scotland was key for this. Intel has made contributions to SYCL, an open C++-based programming model and SYCL is now included in oneAPI so customers can program and compile across CPUs, GPUs and accelerators from multiple vendors. Additional work on software optimizations upstream includes optimizations for PyTorch 2.0 and TensorFlow 2.9, as well as a collaboration with Hugging Face to train, tune and predict with the help of Intel Xeon and Gaudi 2.
This is expected to extend to Gaudi3.
Earlier Hugging Face demonstrated an open source AI framework with 176 billion parameters on the Gaudi 2. The BLOOM model is an open source large language AI model analogous to the 175B parameter GPT-3 model employed by ChatGPT. Hugging Face is also company also running deep-learning text-to-image model Stable Diffusion on 4th Gen Xeon with built-in Intel AMX for AI workload acceleration.