Gamze İslamoğlu

Gamze İslamoğlu, Moritz Scherer, Gianna Paulin et al.

Transformer networks have emerged as the state-of-the-art approach for natural language processing tasks and are gaining popularity in other domains such as computer vision and audio processing. However, the efficient hardware acceleration of transformer models poses new challenges due to their high arithmetic intensities, large memory requirements, and complex dataflow dependencies. In this work, we propose ITA, a novel accelerator architecture for transformers and related models that targets efficient inference on embedded systems by exploiting 8-bit quantization and an innovative softmax implementation that operates exclusively on integer values. By computing on-the-fly in streaming mode, our softmax implementation minimizes data movement and energy consumption. ITA achieves competitive energy efficiency with respect to state-of-the-art transformer accelerators with 16.9 TOPS/W, while outperforming them in area efficiency with 5.93 TOPS/mm$^2$ in 22 nm fully-depleted silicon-on-insulator technology at 0.8 V.

Philip Wiese, Gamze İslamoğlu, Moritz Scherer et al.

One of the challenges for Tiny Machine Learning (tinyML) is keeping up with the evolution of Machine Learning models from Convolutional Neural Networks to Transformers. We address this by leveraging a heterogeneous architectural template coupling RISC-V processors with hardwired accelerators supported by an automated deployment flow. We demonstrate Attention-based models in a tinyML power envelope with an octa-core cluster coupled with an accelerator for quantized Attention. Our deployment flow enables end-to-end 8-bit Transformer inference, achieving leading-edge energy efficiency and throughput of 2960 GOp/J and 154 GOp/s (0.65 V, 22 nm FD-SOI technology).

Thorir Mar Ingolfsson, Wajeeha Tahir, Anna Tegon et al.

Recursive reasoning models solve hard puzzles by applying compact, weight-tied blocks over many refinement steps. Because these blocks are reused many times, quantizing them creates a unique dynamical problem: the quantization error is incurred at every step. While 8-bit quantization (integer or float) preserves accuracy, moving to a per-tensor 4-bit format causes a systematic bias to accumulate. The ensuing drift catastrophically degrades exact-solution accuracy on Sudoku from 84.1% to 0.0% (only ~25% of cells correct). In this work, we show that this collapse is caused by activation-scaling granularity rather than bit-width or number format. Crucially, moving to per-block scaling completely restores the transition. To implement this, we apply MXInt4, a blockwise integer activation format, to recursive reasoning models. It is competitive with blockwise float formats on our tasks, while keeping integer elements and power-of-two block scales. Finally, recursion depth and reuse modulate quantization sensitivity, with the deepest architecture we test (the EqR equilibrium model) the most sensitive. Yet blockwise scaling overcomes this vulnerability, staying robust across these architectures and transferring to the open-ended ARC-AGI benchmark.

Run Wang, Gamze Islamoglu, Andrea Belano et al.

While Transformers are dominated by Floating-Point (FP) Matrix-Multiplications, their aggressive acceleration through dedicated hardware or many-core programmable systems has shifted the performance bottleneck to non-linear functions like Softmax. Accelerating Softmax is challenging due to its non-pointwise, non-linear nature, with exponentiation as the most demanding step. To address this, we design a custom arithmetic block for Bfloat16 exponentiation leveraging a novel approximation algorithm based on Schraudolph's method, and we integrate it into the Floating-Point Unit (FPU) of the RISC-V cores of a compute cluster, through custom Instruction Set Architecture (ISA) extensions, with a negligible area overhead of 1\%. By optimizing the software kernels to leverage the extension, we execute Softmax with 162.7$\times$ less latency and 74.3$\times$ less energy compared to the baseline cluster, achieving an 8.2$\times$ performance improvement and 4.1$\times$ higher energy efficiency for the FlashAttention-2 kernel in GPT-2 configuration. Moreover, the proposed approach enables a multi-cluster system to efficiently execute end-to-end inference of pre-trained Transformer models, such as GPT-2, GPT-3 and ViT, achieving up to 5.8$\times$ and 3.6$\times$ reduction in latency and energy consumption, respectively, without requiring re-training and with negligible accuracy loss.