Gary Smith EDA Consulting in Electronic Design
Gary Smith EDA (GSEDA) is the leading provider of market intelligence and advisory services for the global Electronic Design Automation (EDA), Electronic System Level (ESL) design, and related technology markets.
The New Microprocessor Architecture: There isn’t one!
The New Microprocessor Architecture: There isn’t one!It’s a Cluster
I think what most people are missing is that this isn’t a battle between RISC and CISC, nor is it a battle between a General Purpose CPU or a General Purpose GPU; this is a battle between SMP (Symmetric Multiprocessing) and AMP (Asymmetric Multiprocessing) computing. What is developing is a new heterogeneous microprocessor Cluster architecture. The processors in this Cluster I’m calling Domain (or Dwarf) Optimized Processors (DOPs1). I’m basing my theory on the Thirteen Dwarfs work that has been done at UC Berkeley.
Figure 1: The 13 Dwarfs
1. Dense Linear Algebra
2. Space Linear Algebra
3. Spectral Methods
4. N-Body Methods
5. Structured Grids
6. Unstructured Grids
7. Monte Carlo
8. Combinational Logic
9. Graph Transverse
10. Dynamic Program
11. Backtrack and Branch + Bound
12. Construct Graphical Models
13. Finite Sate Machine
The CPU will be retained but in an increasingly less important role; mainly used for legacy code. The GPU has already taken its place as one of the DOPs and there will be more to come. There will be DOPs that will combine more than one of the Dwarfs. The best example of that is UC Berkley’s work on a DOP optimized for EDA tools. Some of the Dwarfs will continue to be executed by the CPU as any optimization will be minor. Others will just not be economically viable. I originally thought that there would be four to six DOPs, but with the cluster environment I think that most of the Dwarfs will become economically viable and therefore I believe there will be five to nine DOPs.
1 Steve Leibson has suggested Dwarf Optimized Processor Environment (DOPE) which does go along with the Dwarf theme as one of the seven dwarves was dopey (never mind).
In the Cluster architecture some of the DOPs will be limited to four, possible sixteen cores due to Amdahl’s Law. Others, such as the GPU, will not and will appear in a many core configuration. Because of this the Cluster could easily exceed a thousand cores.
This is the architecture that is emerging out of the switch to parallel computing. Power is the prime consideration therefore optimization around the different types of algorithms is a must. We have a lot of work to do but at least now the future is in sight.
A Vision for Intel
As we see the heterogeneous Cluster emerge we also Intel’s future clearer. First of all Intel needs to take their position as one of the three Foundation Semiconductor vendors of this new era. I believe what we will see is Intel merging parts of the IBM strategy with parts of the Samsung strategy into their plan. There will be four synergistic parts of their organization.
The first is based on their uncontested manufacturing capability. They need to form a partnership, with other semiconductor vendors based on their advanced process R&D as IBM has. They then need to go into the foundry business not to support the x86 architecture, as they are doing now, but to support their semiconductor partners in manufacturing the home runs (IC production over 10 million a year) that they produce.
Second they need to advance their SoC business by adopting the ARM processor. It’s not that Intel can’t eventually match ARM’s Power/Performance; it’s that they have no chance of matching ARM’s ecosystem of partners. ARM has pulled a chapter out of the Intel playbook and captured the CPU position in the new heterogeneous processor Cluster. Keep in mind that Intel did not win control over the PC with a superior processor architecture; they won it with a superior programming environment- the Blue Box strategy.
Once they do this they can start developing platforms, as platforms have become the new driving force in SoC development. Today Intel’s major competition is from the platform vendors, most specifically NVIDIA. This is where they need to start competing, not in processors.
Third they need to get into the ASIC business to support their SoC platform program. And fourth they will have a legacy CPU group that will continue to be the manufacturing driver as Samsung has used their memory business to drive their manufacturing strategy.
A Small Suggestion: BUY NVIDIA
Intel is in a situation similar to the one it was in during the early 1980s. The Japanese semiconductor vendors were taking over the DRAM market and Intel had to make a drastic change in their strategy. That’s when Gordon Moore told them to get out of the DRAM business and concentrate on microprocessors. Today Intel must drop the x86 architecture as their strategic emphasis and concentrate on SoC Platforms and the heterogeneous processor Cluster that will be the heart of all future platforms. It wouldn’t be a bad idea for Intel to acquire NVIDIA and make Jen-Sen Huang their next president. Jen-Sen has the best concept of the Cluster architecture of anyone I know.
Cadence BlogsCombating System-Level Design Confusion
AccelleraAccellera Announces IP Tagging Effort, Calls for Participation
DAC Pavilion YoutubeGary Smith EDA at DAC 2011
DAC Pavilion YoutubeWhat's Hot at DAC 2011 DAC Pavilion
EDA CafeVIDEO: ESL, IP, ARM, Etc. and Prediction for 2011