Q2 2021 Competitive Landscape Report – Cambrian AI Research

by | Jun 9, 2021

Report Overview

Welcome to the second edition of the Cambrian-AI Research Competitive Landscape report.  We hope you find this service to be of value and look forward to your feedback so we can continue to improve the report. The 2nd edition contains significant enhancements, including the addition of our first Chinese AI provider, Baidu, as well as a newest chip and benchmark announcements.

All subscriptions come with a money-back guarantee. If you are not satisfied with the service, just let us know within 30 days, and we will promptly refund your fee in full.

Published Quarterly
Individual Report Price


for a limited time

Published Quarterly
Individual Report
Annual Subscription
Access to ALL Quarterly Reports
Q1 Notable Announcements
A Brief Introduction to AI Processing

  • What is AI and why is it attracting so much capital?
  • AI Silicon market size
  • Training and Inference: Building and Using AI Models
  • The Primary Buyers

Competitive Comparison Pitfalls

Public Silicon and Service Providers

  • AMD
  • AWS
  • Baidu
  • Google Cloud
  • Qualcomm
  • Xilinx

The Startups

  • Blaize
  • Cerebras
  • FlexLogix
  • Deep Vision
  • Graphcore
  • Groq
  • Mythic
  • SambaNova Systems
  • SimpleMachines
  • Tenstorrent

Qualitative Comparisons of AI Vendors

Performance and Power Comparisons

Considerations & Conclusions

Figure 1: Data Center and Edge AI Accelerator revenue projections.

Figure 2: ARK Investment group has published a thought-provoking report on Big Ideas in 2021. AI is driving…

Figure 3: AMD has significantly enlarged the L3 cache on the new 7103 to…

Figure 4: The new EPYC CPU.

Figure 5: AMD claims significant TCO savings versus Intel, delivering…

Figure 6: The AMD MI100 has excellent peak double- and single-precision IEEE floating-point performance but …

Figure 7: Using NVIDIA’s data, the resulting cost went from bad news for NVIDIA, to bad news…

Figure 8: The AWS inf1 instance supports the Inferentia inference platform, which offers excellent price/performance compared to…

Figure 9: AWS Inferentia latencies look excellent on YOLO V4.

Figure 10: The AWS stack of hardware choices and software tools is perhaps…

Figure 11: The Kunlun AI accelerator.

Figure 12: The Baidu software stack looks fairly complete.

Figure 13: Google submitted benchmark results for the upcoming TPUv4 and…

Figure 14: Google’s TPU program has produced the company’s supercomputer, the TPU Pod, which, like the NVIDIA Selene, is…

Figure 15: Ice Lake offers excellent AI inference performance for 8-bit representations.

Figure 16: Facebook published a summary of what chips it uses for which types of AI processing.

Figure 17: In some situations, Xeon Cascade Lake can outperform a GPU (V100) and cost less than…

Figure 18: The Habana Gaudi chip scales nearly linearly over a large fabric of nodes in training throughput. It scales 3.8 times better than…

Figure 19: NVIDIA DGX is getting a lot of attention.

Figure 20: NVIDIA revenue trends await Q4 results to provide apples-to-apples trend…

Figure 21: The 80GB version of the A100 can triple the performance of training large models such as…

Figure 22: NVIDIA now offers a wide range of platforms for autonomous vehicles…

Figure 23: NVIDIA MLPerf benchmark results for training.

Figure 24: NVIDIA MLPerf benchmarks for inference blow away all competitors. However…

Figure 25: NVIDIA has demonstrated its ability to increase performance year over year with both…

Figure 26: The Cloud AI 100 is now available from Gigabyte, who has built a…

Figure 27: Qualcomm (QTI) demonstrated excellent performance per watt in their MLPerf submission for ResNet50.

Figure 28: The Cloud AI100 also outperformed NVIDIA in terms of power efficiency for Imagenet, although…

Figure 29: The Qualcomm AI100 has higher throughput at much lower latency than the…

Figure 30: Qualcomm CloudAI100 models are all fast and extremely efficient inference platforms.

Figure 31: Qualcomm offers a comprehensive development stack for AI.

Figure 32: Gartner projects that Edge AI chipsets will rocket past datacenter AI revenues.

Figure 33: Xilinx claims the Versal AI Edge platform quadruples the efficiency of a GPU, and increase…

Figure 34: Xilinx showed how an AI Edge platform could support a wide variety of computational tasks, including…

Figure 35: The Blaize GSP is targeting low-power edge applications.

Figure 36: Blaize has launched a range of platforms, ranging from 16 to 64 TOPS, for edge AI processing.

Figure 37: One of Blaize’s pilot projects is for smart city intersection management and safety.

Figure 38: The WSE-2 has doubled every feature of its predecessor.

Figure 39: The Cerebras “chip” is composed of interconnected dies on a single wafer…

Figure 40: Cerebras will win any comparison with any other chip since it is composed of…

Figure 41: The Flex Logix PCIe card.

Figure 42: Flex Logix has published an array of YOLO V3 performance and latencies that…

Figure 43: The ARA-1 platform delivers good edge performance at a low price point.

Figure 44: This slide wins our “best slide so far this year” award.

Figure 45: Deep Vision is already shipping acceleration cards in various form factors.

Figure 46: The 2nd generation IPU is available in a four-node IPU-Machine for ethernet access to CPUs.

Figure 47: While most observers think of the IPU as a training platform, Graphcore released data that portends a…

Figure 48: The GROQ processor is a unique and novel design, acting as a single core with a high level of…

Figure 49: The Groq node design with two cards with 4 TSP’s each,  joined with 2 AMD Rome EPYC CPUs.

Figure 50: The Groq TSP moves control, planning, and cache management into the…

Figure 51: Mythic supports a variety of PCIe cards for different deployment environments.

Figure 52: SambaNova has published impressive benchmarks for DLRM, besting NVIDIA in both…

Figure 53: SambaNova claims dramatically lower latency and throughput across…

Figure 54: The SimpleMachines architecture is “recomposable,” allowing for optimization of many different models.

Figure 55: The Mozart and coming Bach chips from SimpleMachines have excellent power efficiency…

Figure 56: The Tenstorrent core design.

Figure 57: The Tenstorrent Grayskull chip enables conditional execution, which can have a…

Figure 58: Tenstorrent’s product roadmap.

Figure 59:  Public Vendors Spider diagram.

Figure 60: Private AI companies targeting data center applications.

Figure 61: Edge AI Startups qualitative assessment.

Figure 62: Purdue University graduate student Mahmoud Khairy has kindly permitted us to share…

Subscribe to

The Quarterly AI Research Report

Published Quarterly

Annual Subscription: $1500 for access to the current report and the three upcoming quarterly reports.

Renewals will be automatic, unless you opt out (you will receive email notification before renewal). 

In all cases, these reports are offered with a money-back guarantee.