IBM announces next generation Z processor: 7nm Telum chip, 22.5 billion transistors, 8 cores running on 5GHz clock

In fact, Telum is IBM's first microprocessor to integrate Artificial Intelligence (AI) on the processor chip.

Specifically, according to IBM, the next-generation Z cores are optimized along with an all-new cache and multi-chip hierarchy that enables up to 40% performance improvement per socket. Telum includes a total of 8 cores with dedicated L2 cache. This CPU model has a built-in SMT2 feature that provides 16 on-chip threads, while a maximum configuration of 32 cores and 64 threads is available with a 4-chamber system.

The clock speed is said to be over the 5GHz threshold, while each individual L2 cache is 32MB in size and the latency uses a 19-cycle load (~3.8 ns including CTR access). The dual-chip modular design contains 22 billion transistors and more than 30,000 meters of wiring across 17 metal layers.

Moving on to shared L3 and L4 caching across 8 cores, the IBM Z Telum chip comes with 256MB virtual 'on-chip' L3 cache and 2GB virtual L4 cache on up to 8 cores. L2 cache uses a 320 GB/s dual ring topology, while L3 cache is distributed through association with L2, and has an average latency of 12ns.

This chip's performance in AI Acceleration is rated at more than 6 TFLOPs per chip and over 200 TFLOPs in an IBM Z 4-chamber 4-chamber system. The internal Matrix array has 128 cells with 8-way FP-16 SIMD, high density multiplier and FPU accumulation.

While the Activation array consists of 32 cells with 8 dimensions FP16/FP-32 SIMD. The dual-chip configuration yields 116,000 inferences (1.1ms) while the 32-chip configuration yields 3,600,000 inferences (1.2ms).

The IBM Z Telum on-chip AI accelerator provides:

1. Very low and consistent inference latency
2. Calculating usability on a specific scale
3. Supports various AI models, from traditional ML to RNN and CNN
4. Security - provides enterprise-class virtualization and memory protection
5. Flexible scalability with future hardware and firmware updates.

The IBM Z Telum chip will be fabricated on Samsung's 7nm process and will have a die size of 530mm2. This chip will be geared towards enterprise & embedded workloads. A server system using Telum chip will be launched in the first half of 2022.