Dr. Lloyd Carney, CEO of Brocade predicts the end of Switch gear as we know it.

Confidential © 2015, Vavni Inc. All rights reservedRonald GD Davis

Future Compute Architectures

8/3/15

The Next Server / Switch Modality

Confidential © 2015, Vavni Inc. All rights reservedRonald GD Davis

Article by Dr. Lloyd Carney

Brocade CEO: Specialized Data Center Network Gear on Its Way Out

QUOTED FROMDATACENTERKNOWLEDGE.COM BYYEVGENIY SVERDLIK ON JULY 30, 2015

The days of having x86 servers by Intel or AMD running business apps on one side of the data center and specialized network chips by Cisco, Juniper, or Brocade running networking apps on the other are numbered. That’s according to Brocade CEO Lloyd Carney, who believes x86 servers will replace most if not all specialized networking hardware in data centers of the future.

“Ten years from now, in a traditional data center, it’ll all be x86-based machines,” he says. “They will be running your business apps in virtual machines, your networking apps in virtual machines, but it’ll all be x86-based.”

Confidential © 2015, Vavni Inc. All rights reservedRonald GD Davis

Analysis of Dr. Lloyd Carney Assertions

I belive Dr. Carney is correct in these assertions

1. The server / switch paradigm days are numbered

2. The switch function will be largely found within servers

My Assertions as to Future Compute Architectures

3. True SoC on one die with the following components

a) CPU, GPU

b) AI network fabric processor which optimizes itself as per the type of job, serial, parallel, combination serial / parallel or by network aware programs.

c) 3D Resistive Ram Extreme Long word cells (mux / demux & bank switching to achieve DDR speeds). Many individual cells supporting hybrid Harvard buss architecture to each Processor Core and to the fabric processor to support a new disaggregated hardware architecture design approach.

4. Removal of legacy busways and Ethernet connections.

Confidential © 2015, Vavni Inc. All rights reservedRonald GD Davis

The Functional Areas of the NPU Die

Hardware disaggregation is an essential functional aspect of the Network Processor design. The Fabric Processor is the Harvard bus arbiter of the memory to the CPU / GPU cores / and External fabric mesh. This allows only the configured NPU elements to be functional while other resources are suspended to save power. All legacy IO will be removed.

GPU CPU

Fabric Processor

3D-RRAM

Confidential © 2015, Vavni Inc. All rights reservedRonald GD Davis

3D-RRAM segment NPU Die

3D-RRAM Memory Scale Disaggregated Architecture.

The Multiple Memory busses from the 3D-RRAM are designed to support Independent memory channels to CPU / GPU cores with the Network Processor as the arbiter or bridge. Each core functions independently and can be provisioned as required.

Von Neumann buss is replaced by a hybrid Harvard bus architecture.

When the CPU / GPU cores are suspended, full bandwidth to the 3D-RRAM allows direct high speed access to the 3D-RRAM through the Fabric Arbiter / Mesh processor.

When either GPU or CPU is suspended additional bandwidth may be reallocated to the CPU or GPU cores for enhanced bandwidth performance.

Additionally, External additional 3D-RRAM may be employed to enhance CPU / GPU performance by expanding memory bandwidth through the fabric processor.

Confidential © 2015, Vavni Inc. All rights reservedRonald GD Davis

Mesh Fabric Processor segment NPU Die

Fabric Arbiter / Mesh Processor.

From functional aspect, “0” or more SerDes will be assigned to each leg which will support 2D / 3D mesh operation with the ability to improve serialism or parallelism on any leg. Flexible fabric, In my humble opinion, has not been given much due consideration of late. From this we garner self-configuring and self-optimizing networks will replace the hardwired networks of today. Programs will become more network aware.

1

2

3

4

5

6

Confidential © 2015, Vavni Inc. All rights reservedRonald GD Davis

NPU Closing Thoughts

3D-RRAM Biggest Disruptive Hardware change since the transition from the tube to the transistor. What you see below is a 1U server / meshed fabric switch reduced to roughly the size of a credit card. With a power envelope of 3-15 Watts approximately 3,000 to 5,000 servers like this one will fit in a standard 42U data center cabinet. A 50:1 data center shrink with a vastly superior polymorphic networking stratagem. This represents an order of magnitude power envelope decrease is due in large part by removing all the Legacy IO.

Confidential © 2015, Vavni Inc. All rights reservedRonald GD Davis

Future Compute Architectures 8/3/15

THANK YOU