09:21PM EDT - The final talk today at Hot Chips is from Habana, who is discussing its approach to how to scale AI compute.

09:21PM EDT - Goya and Gaudi

09:22PM EDT - Recapping Training vs Inference requirements

09:24PM EDT - Goya processor architecure

09:24PM EDT - 3 engines, RPC, GEMM, and DMA. Work Concurrently with shared SRAM

09:24PM EDT - TPC is VLIW SIMD core, C-programmable

09:24PM EDT - PCIe Gen 4.0 x16

09:24PM EDT - Two DDR4-2666 channels, built on TSMC 16

09:25PM EDT - Supports UINT8 to FP32

09:25PM EDT - Dedicated HW and TPC ISA for special function acceneration

09:25PM EDT - Have to adjust quantization to mix accuracy vs power

09:26PM EDT - PCIe card - Software stack is more important.

09:26PM EDT - Habana is a software company that just happens to do hardware

09:27PM EDT - Graph compiler with built-in quantization engine

09:27PM EDT - Multiple recipes can be loaded for the hardware

09:28PM EDT - Goya supports models trained on any processor: CPU, GPU, TPU, Gaudi etc

09:28PM EDT - Users can create custom layers and kernels

09:29PM EDT - Still market leader since benchmarks made 11 months ago vs common CPU/GPU

09:29PM EDT - New for today, natural language benchmark results

09:30PM EDT - Support BERT architecture on Goya

09:30PM EDT - GEMMs and TPCs are fully utilized

09:30PM EDT - Chip was designed long before BERT was invested

09:30PM EDT - invented

09:30PM EDT - High degree of accuracy when quantized

09:30PM EDT - Software managed SRAM

09:31PM EDT - Now Gaudi, the training processor

09:31PM EDT - Performance at Scale, high throughput at low batch size, high power efficiency

09:32PM EDT - Enable native ethernet scale out - on chip RDMA over Converged Ethernet

09:32PM EDT - Open Compute Project Accelerator Module: OAM = (OCP)AM

09:32PM EDT - Framework and ML compiler support, rich TPC Kernet Library

09:32PM EDT - Architecture looks similar to Goya

09:33PM EDT - Networking has changed, memory has changed

09:33PM EDT - PCIe 4.0 x16, 4x8GB HBM

09:33PM EDT - 10x 100 GbE, or 20x50 GbE

09:33PM EDT - Supports UINT8 to FP32 and BF16

09:34PM EDT - SW supports profiling tools

09:34PM EDT - Only AI Training chip with RoCE v2

09:35PM EDT - NVIDIA was first to showcase RoCE v2 for AI, but they haven't implemented it yet

09:36PM EDT - NVIDIA GPU is much more complex with RoCE v2 support via Mellanox

09:36PM EDT - Gaudi integrates both

09:36PM EDT - Supports Lossless and Lossy fabrics

09:36PM EDT - Advanced congestion controls

09:37PM EDT - Customers can buy OAM cards or an 8 card Server

09:38PM EDT - Server box has no CPU, up to customer to config to needed. Uses mini-SAS HD

09:38PM EDT - Ethernet connectivity for point-to-point links with non-blocking full mesh

09:38PM EDT - 3 ports per card for scale up

09:39PM EDT - Can choose ratio of CPUs to Gaudi cards

09:39PM EDT - Gaudi vs DGX

09:40PM EDT - Unlike DGX, do not force user to separate PCIe between management and scaleout. Gaudi offers separate PCIe ports

09:41PM EDT - PCIe card dual slot also available

09:41PM EDT - HL-200

09:41PM EDT - Data parallel possible, model parallel possible

09:44PM EDT - Can leapfrog performance over DGX-2 due to better connectivity. Can connect 64 gaudi chips with non-blocking throughput

09:45PM EDT - Q&A time

09:46PM EDT - Q: What type of quantization requires a processor? There is no quantization processor. There's a software engine that takes an FP32 model and can quantize to data types that are more efficient and gives the feedback on the accuracy

09:47PM EDT - Q: Can you comment on interconnectivity of GEMM? A: It's one functional unit.

09:48PM EDT - Q: What is the minimum viable for an IoT gateway? A: You can use a single card. You can put a gaudi in a single PCIe slot.

09:48PM EDT - That's a wrap for today. More talks tomorrow!