Erica Virtual Assistant (6/18) Note: We assume a start at zero users from Erica’s launch in 6/18. Pilot users excluded. Source: Bank of America (2/21, 4/24, 2/25) Bank of America Erica Virtual Assistant 10/21 3/22 8/22 1/23 6/23 11/23 4/24 9/24 2/25 Cumulative Client Interactions with Erica Virtual Assistant (MM) Note: Erica is a conversational AI built into Bank of America’s mobile app that helps + High Valuations + High Investment Levels = Good News for Consumers… Others TBD…176 *Select media reports have xAI revenue being as high as $1B as of 4/25. Note: OpenAI annualized revenue estimated

buffer to hide memory load latency 。 storage align to reduce bank conflicts of shared memory 。 Virtual threads for data reuse (on going) Performance on V100 (FP16) 计算平台事业部 COMPUTING PLATFORM 512, 16 下和全于由 loss = loss_fn() opt = tf.Adamoptimizer(learning_rate=...) # Choose a 1oss Scale manager which decides how to pick the right loss scale # throughout the training process. 1oss_scale_manger original optimizer in a LossScale0ptimizer . loss_scale_optimizer = LossScaleOptimizer(opt，1oss_scale_manager) # Call minimize() on the loss scale optimizer. train_op = loss_scale_optimizer.minimize(1oss) Loss

model in TVM ● Support Any-dim in typing ● Use shape function to compute the type at runtime ● Virtual machine as a new runtime for Relay ● Dynamic codegen (WIP) ○ Kernel dispatch for a single op ○ Data dependent© 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. Relay virtual machine Relay Executable relay.vm.compile Relay Object (hardware independent) Code segment VM a data type using the entries from a register. AllocClosure Allocates a closure with a lowered virtual machine function. If Jumps to the true or false offset depending on the condition. Goto Unconditionally

areas: o Core Infrastructure Improvements to TVM o_uTVM: support for microcontrollers in TVM o_ Virtual Machine and dynamic NNs support (w/ AWS folks) o_ Improved NLP support, with focus on transformers o QQ octoML BERT has many reshape operations, which are currently implemented using copy, 10 Virtual Machine e Many improvements from contributors at UW, AWS, and OctoML. e Initial implementation

remote_mod = remote.load_module(“mylib.so") func = remote_mod[“npufunction0"] func(remote_a, remote_b)Virtual Machine: Supporting Dynamic Workload Dynamic shape workloads More runtime objects: Arrays, Tuples

Product Note: Operator Operator is an example of OpenAI’s agentic approach. Leveraging its own virtual browser, Operator can navigate the web, click on buttons, fill in forms, and gather data just like

ticket to a human. Orchestration Agents themselves can serve as tools for other agents—see the Manager Pattern in the Orchestration section. Refund agent, Research agent, Writing agent. 9 A practical requirements, our experience with customers highlights two broadly applicable categories: Manager (agents as tools) A central “manager” agent coordinates multiple specialized agents via tool calls, each handling specializations. Multi-agent systems can be modeled as graphs, with agents represented as nodes. In the manager pattern, edges represent tool calls whereas in the decentralized pattern, edges represent handoffs