Introduction

The next chapters describe how to create programs running on DPUs, interacting with host applications.

It includes:

• Using the functionality provided by the Runtime Library in the DPU program

• Using the Host API that allows host applications to start and control the execution of DPUs

Those two components are tightly correlated, offering different categories of services:

• Tasklet management and synchronization: how to define threads, how they operate and interact

• Memory management: how to access the different memories within a DPU

• Communication with host applications: how to share data with the host

• Standard library functions: how to use the C standard library

• Exceptions: understanding the SDK exception model

• Controlling the execution of DPUs from host applications: how to control the DPU from the host side

• Logging: how to log messages from the DPU

• Performance Counters: how to access hardware performance counter

DPU chip characteristics

UPMEM curently provides two different models of DPU, named v1A and v1B.

The major differences between these two are as follows:

	v1A	v1B
Nominal frequency	350 MHz	400 MHz
Instruction Memory (IRAM) [1]	<= 4096 instructions	<= 3968 instructions
Working memory (WRAM) [1]	<= 65536 bytes	<= 63488 bytes
Tasklets [2]	<= 24	<= 16
Mutexes	<= 56	<= 64

[1] (1,2)

Part of the IRAM and WRAM in the v1B has been reserved for production and quality control purposes.

[2]

The DPU’s arithmetic performance is not affected by the maximum number of tasklets. Both types of chips execute the same number of instructions concurrently (11). In real-world scenarios, including applications developed internally and by our users, we have observed that exceeding 16 tasklets generally does not enhance performance.