The document discusses how deep learning requires a large amount of computing power and how its requirements are driving the convergence of high performance computing and hyperscale data centers. It notes that deep neural networks require orders of magnitude more computations per data byte than traditional algorithms. This large computational need, coupled with the growth of data centers and data volumes, means deep learning workloads will require specialized hardware designs across multiple levels, from silicon design to data center architecture. Heterogeneous computing resources optimized for deep learning will be needed both within and across servers and data centers to efficiently process large neural networks on big data.
Gen AI in Business - Global Trends Report 2024.pdf
Deep Learning Convergence of HPC and Hyperscale
1. Deep Learning : Convergence of HPC and
Hyperscale
Sumit Sanyal
CEO
2. Technology Revolution :
Deep Neural Nets
Ø A new paradigm in programming
Ø DNNs remarkably effective at tackling many problems
Ø Designing new NN architectures as opposed to “programming”
Ø Training as opposed to “compiling”
Ø Trained weights as the “new” binaries
Ø Big Neural Nets required to process Big Data
Ø Videos, images, speech and text
Ø DNNs are significantly increasing recognition accuracies which have stagnated
for decades
Ø Used to structure Big Data
3. Make a
guess at NN
architecture
Run on
training
data, does
it perform
well?
No
Run on
test data,
does it
perform
well?
No
Yes
DONE!
Design Bigger
Network
Need more Training
Data
Yes
Neural network development
Iterate
Iterate
Train on
labeled
dataset
Wait 1-7+ days
3
4. Ø Deep Neural Nets require a lot of compute cycles!
Ø An example – image classification:
Ø Ratio of (Compute cycles : IO bandwidth) significantly higher than non AI algorithms
Ø Training AlexNet (for image classification) requires ~27,000 flops/input data byte
Ø Training VGG ~150,000 flops/ data byte
Ø R3 / R2 à Volume (compute) / Surface(IO BW)
Ø Significantly higher for Deep Nets
Ø Power dissipation challenges
Ø Compute density limited by DC cooling capacity
Ø At 1 ~uW / MHz (current state-of-art in 28 nm) requires 300 Watts!
Compute vs. IO
AI is no longer bored J
5. Neural Net Computations
Ø All Deep Neural Net implementations have the following properties
Ø Small set of non-linear transforms
Ø Small set of linear algebra primitives
Ø Relatively modest dynamic range of weight/data values
Ø Very regular/repetitive data flows
Ø Only persistent memory requirement is for weights
Ø Updated while learning, fixed for recognition
Ø Variance in the size of the net across applications is >105
Compute cycles will be commoditized; not computers!
6. Why minds.ai?
minds.ai was founded on the basis that Deep Learning is the next
frontier for HPC – it’s what we do
“…around 2008 my group at Stanford started advocating
shifting deep learning to GPUs (this was really
controversial at that time; but now everyone does it); and
I'm now advocating shifting to HPC (High
Performance Computing/Supercomputing) tactics for
scaling up deep learning. Machine learning should
embrace HPC. These methods will make researchers
more efficient and help accelerate the progress of our
whole field.” – Andrew Ng, Feb 2016
HPC Design
Expertise
Neural Network
Design Expertise
Image and Video
Recognition
Expertise
GPU Programming
Expertise
+ + +
7. Ø Instruction level – SIMD, VLIW etc.
Ø Thread level – warps
Ø Processor level – many cores
Ø Server level – many GPUs in a server
Ø Cluster level – many servers with high BW interconnect
Ø Data Center level
Ø Planet level J
7 levels of parallelism
8. Ø Exponential growth in Data Centers
Ø Commoditization of Enterprise silicon
Ø Traditional mobile players announcing ASICs for enterprise
compute
Ø Higher demands for compute density
Ø GPGPUs have won the first round
Ø Dennardian scaling is breaking down
Ø Power dissipation will emerge as major challenge
Ø Chip level, server level and DC level
ASICs for DNNs
9. Age of dark silicon
Transistor property Dennardian Post Dennardian
# of transistors
( Q )
S2 S2
Peak clock frequency
( F )
S S
Capacitance
( C )
1/S 1/S
Supply Voltage
( Vdd )
1 / S2 1
Dynamic Power
(QFCVdd
2)
1 S2
Active Silicon 1 1/S2
* S is the ratio of feature size between next generation processes
10. Ø Dark silicon will drive heterogeneity
Ø Multi core architectures with different Instruction Sets
Ø Power aware scheduling across cores
Ø Decreasing parts of the chip can run at full clock frequency
Ø Specialized silicon for server blades
Ø Bridges to intra server and inter server communications
Ø Last level caching support, caching across MPI
Ø Distributed compute in network interfaces
Heterogeneity in Enterprise silicon
11. Ø Instruction level – SIMD, VLIW etc.
Ø Thread level – warps
Ø Processor level – many cores
Ø Server level – many GPUs in a server
Ø Cluster level – many servers with high BW interconnect
Ø Data Center level
Ø Planet level J
7 levels of parallelism
12. Ø Deep Learning is driving the convergence of High Performance
Compute (HPC) and Hyperscale (Data Centers)
Ø Traditional HPC ecosystems : expensive and bleeding edge
Ø DC infrastructure : commodity and homogeneous
Ø Single or dual CPU servers common
Ø All of this is changing
Ø GPGPUs now common in DCs, initial resistance
Ø InfiniBand penetration has reached a tipping point
Ø Dense compute clusters require high bandwidth interconnects
Heterogeneity in Hyperscale
13. Ø Intra server vs. Inter server bandwidths
Ø Inter server bandwidths will grow faster than intra
server
Ø Lead to larger, denser servers
Ø 8 or more GPGPUs per server for DNN training jobs
Ø Will co-exist with CPU based servers for search and
database operations
Ø Many kinds of servers, one size fits all does not work
Server Architectures
14. Training Server Reference Design
CPU
GPU
GPU
GPU
GPU
GPU
GPU
GPU
GPU
CPU
IB NIC
IB NIC
Ethernet
PCIe Root Hub
SSD
SSD
15. GPU Training Cluster Architecture
Server Node
IB Switch
Server Node
Server Node
Server Node
Server Node
Server Node
Server Node
Ethernet Switch
RAID & Access
• Scalable up to 7 Server Nodes
• 8 GPUs per Node
• ~2.5kw per Server Node
• 100Gbps IB Interconnect
16. Ø Instruction level – SIMD, VLIW etc.
Ø Thread level – warps
Ø Processor level – many cores
Ø Server level – many GPUs in a server
Ø Cluster level – many servers with high BW interconnect
Ø Data Center level
Ø Planet level J
7 levels of parallelism
17. Ø Computing super clusters
Ø 105 variability in size of compute ‘jobs’
Ø Large number of collocated servers running the same job
Ø High BW, low latency interconnect
Ø Clusters could grow to significant fraction of a Data Center
Ø Architecture of clusters will be hierarchical and heterogeneous
Ø Edge servers for security and Data management
Ø Dedicated RAID servers
Ø Dedicated compute servers
Ø Control and management nodes
Ø Multi DC training clusters for big models are technically feasible
Ø Scientific community has performed trans continental simulations
Data Center Architectures