Accelerating ML Recommendation with over a Thousand

RISC-V/Tensor Processors on Esperanto’s ET-SoC-1 Chip

Dave Ditzel, email: [email protected], Founder and Executive Chairman, Esperanto Technologies,

Roger Espasa, Nivard Aymerich, Allen Baum, Tom Berg, Jim Burr, Eric Hao, Jayesh Iyer, Miquel Izquierdo, Shankar Jayaratnam,
Darren Jones, Chris Klingner, Jin Kim, Stephen Lee, Marc Lupon, Grigorios Magklis, Bojan Maric, Rajib Nath, Mike Neilly, Duane
Northcutt, Bill Orner, Jose Renau, Gerard Reves, Xavier Reves, Tom Riordan, Pedro Sanchez, Sri Samudrala, Guillem Sole,
Raymond Tang, Tommy Thorn, Francisco Torres, Sebastia Tortella, Daniel Yau

Esperanto Technologies Hot Chips 33 Presentation August 24, 2021

The ET-SoC-1 is the highest performance commercial RISC-V chip

The ET-SoC-1 has over a thousand RISC-V processors on a single TSMC 7nm chip, including:
• 1088 energy-efficient ET-Minion 64-bit RISC-V in-order cores each with a vector/tensor unit
• 4 high-performance ET-Maxion 64-bit RISC-V out-of-order cores
• >160 million bytes of on-chip SRAM
• Interfaces for large external memory with low-power LPDDR4x DRAM and eMMC FLASH
• PCIe x8 Gen4 and other common I/O interfaces
• Innovative low-power architecture and circuit techniques allows entire chip to
• Compute at peak rates of 100 to 200 TOPS
• Operate using under 20 watts for ML recommendation workloads

This general-purpose parallel-processing system on a chip can be used for many parallelizable workloads

But today, we want to show why it is a compelling solution for Machine Learning Recommendation (inference)
• ML Recommendation is one of the hardest and most important problems for many hyperscale data centers

2
Requirements and challenges for ML Recommendation in large datacenters

Most inferencing workloads for recommendation systems in large data centers are run on x86 servers

Often these servers have an available slot for an accelerator card, but it needs to meet key requirements:
• 100 TOPS to 1000 TOPS peak rates to provide better performance than the x86 host CPU alone
• Limited power budget per card, perhaps 75 to 120 watts, and must be air-cooled[1]
• Strong support for Int8, but must also support FP16 and FP32 data types[1,2]
• ~100 GB of memory capacity on the accelerator card to hold most embeddings, weights and activations[3]
• ~100 MB of on-die memory[5]
• Handle both dense and sparse compute workloads. Embedding look-up is sparse matrix by dense matrix
multiplication[5]
• Be programmable to deal with rapidly evolving workloads[1], rather than depending on overly-specialized
hardware[4,5]

3
Esperanto’s approach is different... and we think better for ML Recommendation

Other ML Chip approaches Esperanto’s better approach

DRAM DRAM DRAM DRAM DRAM DRAM
10-20 CPU cores
DRAM
and
> 1000 > 1000 > 1000 > 1000 > 1000 > 1000
Systolic Array Multipliers RISC-V/Tensor RISC-V/Tensor RISC-V/Tensor RISC-V/Tensor RISC-V/Tensor RISC-V/Tensor
Cores Cores Cores Cores Cores Cores

One Giant Hot Chip uses up power budget Use multiple low-power chips that still fit within power budget
Limited I/O pin budget limits memory BW Performance, pins, memory, bandwidth scale up with more chips
Dependence on systolic array multipliers Thousands of general-purpose RISC-V/tensor cores
• Great for high ResNet50 score • Far more programmable than overly-specialized (eg systolic) hw
• Not so good with large sparse memory • Thousands of threads help with large sparse memory latency
Only a handful (10-20) of CPU cores Full parallelism of thousands of cores always available
• Limited parallelism with CPU cores when Low-voltage operation of transistors is more energy-efficient
problem doesn’t fit onto array multiplier • Lower voltage operation also reduces power
Standard voltage: Not energy efficient • Requires both circuit and architecture innovations
Challenge: How to put the highest ML Recommendation performance
onto a single accelerator card with a 120-watt limit?
4
Could fit six chips on 120W card, if each took less than 20 watts

Assumed half of 20W power for 1K RISC-V cores, so only 10 mW per core!

Power (Watts) = Cdynamic x Voltage2 x Frequency + Leakage

Power/core Frequency Voltage Cdynamic

Generic x86 Server core (165W for 24 cores) 7W 3 GHz 0.850v 2.2nF
10mW ET-Minion core (~10W for 1K cores) 0.01 W 1 GHz 0.425v 0.04nF
Reductions needed to hit goals ~700x 3x 4x 58x
Easy Hard Very Hard
Circuit/SRAM Architecture
5
 Study of energy-efficiency and number of chips to get best ML Performance in 120 watts[6]

20x better Energy-Efficiency

of ET-Minion cores (Inferences/Sec/Watt)

8.5 W 2.5x better performance than the 118W chip

by using lowest voltage instead of
Recommendation Energy Efficiency

6 chips highest voltage for our

recommendation benchmark
Esperanto’s
sweet-spot for
best performance
4x better ML
20 W recommendation
6 chips performance

275 W
1 chip
0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
Operating voltage for the 1K ET-Minion RISC-V/Tensor cores
6
 ET-Minion is an Energy-Efficient RISC-V CPU with a Vector/Tensor Unit

256b Floating Point 512b Int8 Vector RF

ET-MINION IS A CUSTOM BUILT 64-BIT RISC-V PROCESSOR 32-bit & 16-bit FMA Bypass TIMA TIMA VPU RF T0/T1 Trans
ROMs
• In-order pipeline with low gates/stage to improve MHz at low voltages 32-bit & 16-bit FMA Bypass TIMA TIMA VPU RF T0/T1

Vector/Tensor Unit
• Architecture and circuits optimized to enable low-voltage operation 32-bit & 16-bit FMA Bypass TIMA TIMA VPU RF T0/T1 Trans
• Two hardware threads of execution ROMs
32-bit & 16-bit FMA Bypass TIMA VPU RF T0/T1
•
TIMA
Software configurable L1 data-cache and/or scratchpad
32-bit & 16-bit FMA Bypass TIMA TIMA VPU RF T0/T1 Trans
ROMs
ML OPTIMIZED VECTOR/TENSOR UNIT 32-bit & 16-bit FMA Bypass TIMA TIMA VPU RF T0/T1

• 512-bit wide integer per cycle 32-bit & 16-bit FMA Bypass TIMA TIMA VPU RF T0/T1 Trans
• 128 8-bit integer operations per cycle, accumulates to 32-bit Int 32-bit & 16-bit FMA Bypass TIMA TIMA VPU RF T0/T1
ROMs

• 256-bit wide floating point per cycle

RISC-V Integer
• 16 32-bit single precision operations per cycle RISC-V Integer Pipeline
• 32 16-bit half precision operations per cycle Data Cache Control
• New multi-cycle Tensor Instructions (including D-tags)
L1 Data-Cache/Scratchpad
Front
End
• Can run for up to 512 cycles (up to 64K operations) with one tensor instruction
DC bank 0 DC bank 1 DC bank 2 DC bank 3
• Reduces instruction fetch bandwidth and reduces power
• RISC-V integer pipeline put to sleep during tensor instructions ET-Minion RISC-V Core and Tensor/Vector unit
• Vector transcendental instructions optimized for low-voltage operation
to improve energy-efficiency
OPERATING RANGE: 300 MHz TO 2 GHz

Optimized for energy-efficient ML operations. Each ET-Minion can deliver peak of 128 Int8 GOPS per GHz

7
8 ET-Minions form a “Neighborhood”

NEIGHBORHOOD CORES WORK CLOSELY TOGETHER

ET-Minion 0 ET-Minion 4
• Architecture improvements capitalize on physical proximity of 8 cores

Instruction Cache
32 KB shared
• Take advantage that almost always running highly parallel code
ET-Minion 1 ET-Minion 5

OPTIMIZATIONS FROM CORES RUNNNING THE SAME CODE ET-Minion 2 ET-Minion 6

• 8 ET-Minions share single large instruction cache,
this is more energy efficient than many separate instruction caches. ET-Minion 3 ET-Minion 7
• “Cooperative loads” substantially reduce memory traffic to L2 cache

NEW INSTRUCTIONS MAKE COOPERATION MORE EFFICIENT

• New Tensor instructions dramatically cut back on instruction fetch bandwidth
• New instructions for fast local synchronization within group
• New Send-to-Neighbor instructions
• New Receive-from-Neighbor instructions

8
32 ET-Minion CPUs and 4 MB Memory form a “Minion Shire”

Four 8-Core Mesh

Neighborhoods Interconnect 32 ET-MINION RISC-V CORES PER MINION SHIRE
Nominal Voltage
• Arranged in four 8-core neighborhoods

m3 m0 Bank0
m7 m4 (1MB) SOFTWARE CONFIGURABLE MEMORY HIERARCHY
m11 m8 Bank1 • L1 data cache can also be configured as scratchpad
m15 m12
4x4
(1MB) Mesh • Four 1MB SRAM banks can be partitioned as private L2,
stop shared L3 and scratchpad
m19 m16 xbar Bank2
m23 m20 (1MB)
m27 m24 Minion
m31 m28
Bank3
Shire SHIRES CONNECTED WITH MESH NETWORK
(1MB)

Low Voltage Low Voltage NEW SYNCHRONIZATION PRIMITIVES

4MB Banked SRAM • Fast local atomics
Cache/Scratchpad • Fast local barriers
Local Sync Primitives • Fast local credit counter
• IPI support

9
Shires are connected to each other and to external memory through Mesh Network

Minion Shire Minion Shire

m3 m0 m3 m0
Bank0 Bank0
m7 m4 (1MB) m7 m4 (1MB)
m11 m8 m11 m8
Bank1 Bank1
m15 m12 (1MB) Mesh m15 m12 (1MB) Mesh
stop stop
m19 m16 m19 m16
Bank2 Bank2
m23 m20
4x4 (1MB) m23 m20 4x4 (1MB)
xbar xbar
m27 m24 m27 m24
Bank3 Bank3
m31 m28 (1MB) m31 m28 (1MB)

Minion Shire Minion Shire

m3 m0 m3 m0
Bank0 Bank0
m7 m4 (1MB) m7 m4 (1MB)
m11 m8 Bank1 m11 m8
Bank1
m15 m12 (1MB) Mesh m15 m12 (1MB) Mesh
stop stop
m19 m16 Bank2 m19 m16
4x4 4x4 Bank2
m23 m20 (1MB) m23 m20 (1MB)
xbar xbar
m27 m24 m27 m24
Bank3 Bank3
m31 m28 (1MB) m31 m28 (1MB)

10
 ET-SoC-1: Full chip internal block diagram

MINION MINION MINION MINION RISC-V Service

SHIRE SHIRE SHIRE PCIe SHIRE MAXION / IO SHIRE
SHIRE Processor

34 MINION SHIRES
PCIe logic DFT/eFuses
ET-Maxion ET-Maxion
• 1088 ET-Minions
ET-Maxion ET-Maxion

8 MEMORY SHIRES
LPDDR4x
• LPDDR4x DRAM
LPDDR4x
16-bits 16-bits
controller controller

16-bits
LPDDR4x LPDDR4x
16-bits
controllers
controller controller

MEMORY SHIRE MEMORY SHIRE

1 MAXION / IO SHIRE
16-bits
LPDDR4x LPDDR4x
16-bits
• 4 ET-Maxions
controller controller
• 1 RISC-V Service
LPDDR4x
16-bits
controller
LPDDR4x
controller
16-bits Processor
MEMORY SHIRE MEMORY SHIRE

PCIe SHIRE
LPDDR4x LPDDR4x
16-bits 16-bits
controller controller

LPDDR4x LPDDR4x
160 million bytes of
16-bits 16-bits
controller controller
on-die SRAM
MEMORY SHIRE MEMORY SHIRE

LPDDR4x LPDDR4x
x8 PCIe Gen4
16-bits 16-bits
controller controller

16-bits
LPDDR4x
controller
LPDDR4x
controller
16-bits Secure Root of Trust
MEMORY SHIRE MEMORY SHIRE

MINION MINION MINION MINION MINION MINION

SHIRE SHIRE SHIRE SHIRE SHIRE SHIRE

Block diagram of Esperanto’s energy-efficient ET-SoC-1 chip

11
 ET-SoC-1 External Chip Interfaces

8-bit PCIe Gen4

• Root/endpoint/both

256-bit wide LPDDR4x

• 4267 MT/s
• 137 GB/s
• ECC support

RISC-V SERVICE PROCESSOR

• Secure Boot
• System Management
• Watchdog timers
• eFuse

EXTERNAL IO
• SMBus
• Serial – I2C/SPI/UART
• GPIO
• FLASH

12
Card with six ET-SOC-1 chips for large sparse ML Recommendation models
• Esperanto’s low-power technology allows six Esperanto chips and 24 DRAM chips to fit into120-Watt power budget of customer’s PCIe card
• A single ML model on one accelerator card can use up to 192 GB of low-cost LPDDR4x DRAM with up to 822 GB/s of memory bandwidth
• Over 6K cores with 12K threads handles memory latency on 96 memory channels and performs well for ML Recommendation (and other) tasks

1536-bit wide memory system

24 DRAM chips

4 16-bit channels

4 16-bit channels
4 16-bit channels

4 16-bit channels
4 16-bit channels

4 16-bit channels

4 16-bit channels

4 16-bit channels

4 16-bit channels
4 16-bit channels

4 16-bit channels
4 16-bit channels

4 16-bit channels

4 16-bit channels

4 16-bit channels
4 16-bit channels

4 16-bit channels

4 16-bit channels
4 16-bit channels

4 16-bit channels

4 16-bit channels

4 16-bit channels
4 16-bit channels

4 16-bit channels
LPDDR4x

LPDDR4x
LPDDR4x
LPDDR4x

LPDDR4x

LPDDR4x

LPDDR4x

LPDDR4x
LPDDR4x

LPDDR4x
LPDDR4x

LPDDR4x

LPDDR4x

LPDDR4x
LPDDR4x

LPDDR4x
LPDDR4x

LPDDR4x
LPDDR4x
LPDDR4x
LPDDR4x

LPDDR4x

LPDDR4x
LPDDR4x
96 16- bit channels

Up to 192 GB of accelerator memory with 24 DRAMs each 64-bits wide.

Esperanto can handle large memory requirements for ML Recommendation.

ET-SoC-1 ET-SoC-1 ET-SoC-1 ET-SoC-1 ET-SoC-1 ET-SoC-1

1088 ET-Minions 1088 ET-Minions 1088 ET-Minions 1088 ET-Minions 1088 ET-Minions 1088 ET-Minions
4 ET-Maxions 4 ET-Maxions 4 ET-Maxions 4 ET-Maxions 4 ET-Maxions 4 ET-Maxions
160 MB SRAM 160 MB SRAM 160 MB SRAM 160 MB SRAM 160 MB SRAM 160 MB SRAM

PCIe switch
PCIe card interface

13
Six ET-SoC-1 chips fit on an OCP Glacier Point v2 Card

GLACIER POINT V2 CARD SHOWN:

• 6,558 RISC-V cores
• Up to 192 GB of DRAM
• Up to 822 GB/s DRAM bandwidth
• ~120 W maximum power consumption
Based on the open-source Glacier Point v2 board
design from the Open Compute Project.
Three Esperanto Dual M.2 modules can mount on the
top side and three on the bottom side.
ON OTHER PCI EXPRESS CARDS:
• One ET-SoC-1 fits easily on a Low-Profile
(half-height, half-length) PCIe add-in card
• ET-SoC-1 power budget increases to ~60W
• Many PCIe cards can be operated in parallel

Peak performance > 800 TOPS8 when all ET-Minions on six chips are operating at 1 GHz

14
 ET-SoC-1 can be deployed at scale in existing OCP Data Centers

x2 x4 x8
Top of Rack Switch

Yosemite v2
Power Shelf

Power Shelf

OCP Glacier Point v2 Yosemite v2 Sled Yosemite v2 Cubby Rack with Yosemite v2 Example OCP Data Center
accelerator card holds: holds: holds: holds:
holds: • 2 Glacier Point v2 • 4 Yosemite Sleds  8 Yosemite v2 Cubbies  @ 30 sq. ft. per OCP rack [7]
• 6 Esperanto accelerator cards • 8 accelerator cards  64 accelerator cards  Estimated 4K-20K racks / data center
ET-SoC-1 chips • 12 ET-SoC-1 chips • 48 ET-SoC-1 chips  384 ET-SoC-1 chips  Millions of ET-SoC-1 chips

15
Software: Esperanto Supports C++ / PyTorch and Common ML Frameworks

…..

MS CNTK Development Tools

ML Model Frameworks
…..
C++ ONNX Models Console / Performance
Debugger Monitor
GLOW/TVM Frontend:
 GLOW = Graph LOWering
GLOW TVM  Open source graph compilers
 Hardware-independent optimizations
 Divides work across N chips

ET-GLOW Backend LLVM GLOW/TVM Backend: Management Utilities

 Does hardware-dependent optimizations
 Backend modified by ET to generate
instructions for ET-SoC-1 chip Firmware
ET RISC-V Code Gen
Diagnostics
Updater
ET-Runtime
ET PCIe Driver GLOW/TVM driver runs on x86 host

... ML Models run across multiple

ET-SoC-1 chips

16
 ML Recommendation performance per card comparisons
Based on MLPerf Deep Learning Recommendation Model benchmark [8]
123
125
= Relative Performance (Samples/Second)

100 = Relative Energy Efficiency (Samples/Second/Watt)

or Relative Performance at equal Power

75
123x better 59
performance per watt 52
50
39
31
25
11
1 1
0
Intel Xeon Platinum
Intel 8380H8380H 8S NVIDIA
NVIDIA T4T4 NVIDIA A10
NVIDIA A10 Esperanto ET-SoC-1
ET-SOC-1 x6

24,630 Samples/sec with 8 die [9a.1] 665,646 with 20 T4 cards [9b] 772,378 with 8 A10 cards[9c] 182,418 with 1 card (6 chips) [9d]
2000 watts for 8 die 1400 watts for 20 T4 1200 watts for 8 A10 120 watts
3,079 Samples/sec/die 33,282 Samples/sec/T4 card 96,547 Samples/sec/A10 card 182,418 Samples/sec/card
250 watts per die [9a.2] 70 watts per T4 card 150 watts per A10 card 120 watts per card
MLPerf DLRM results for Intel and NVIDIA from MLCommons website[8], power numbers from data sheets. Esperanto estimates[9d] 17
 Image Classification performance per card comparisons
30 Based on ResNet-50 benchmark [10]
25.7
25 = Relative Performance (Inferences/Second)

= Relative Energy Efficiency

20
(Inferences/Second/Watt)
or Relative Performance at equal Power
15.2 15.4
15
25.7x better
performance per watt
10
7.6
5.7 5.3
5 3.9
2.0
1.0 1.0
0
Intel Xeon Platinum 9282 2S NVIDIA T4 NVIDIA A10 Habana Goya Esperanto ET-SoC-1
(4 die @200W TDP each[10a.2]) (1 die @70W TDP) (1 die @150W TDP) (1 die @100W) 6 die (@20W each)
7,878 Inf/sec batch size=11[10a.1] batch size=8 batch size=8 batch size=10 batch size=8
1,970 Inferences/sec/die 3,920 Inferences/sec[10b] 7,755 Inferences/sec[10b] 15,012 Inferences/sec[10c] 30,369 Inferences/sec[10d]
200 watts per die 70 watts per card 150 watts per card 100 watts per card 120 watts per card

ResNet-50 numbers taken from respective company websites, power numbers from data sheets. Esperanto estimates
18
Four ET-Maxions: High-Performance Out-of-Order RISC-V Processors

FULL RISC-V RV64GC ISA SUPPORT

• Support for compressed ISA
• Privileged ISA
• Fully respects relaxed consistency model
SUPPORTS 4 CORE SMP LINUX OS
• Allows chip to run in standalone configuration
OUT-OF-ORDER EXECUTION
• 64-entry distributed scheduler
• 128-entry ROB
• 32-entry load queue, 32-entry store queue
• 8R/4W 128-entry integer physical register file
• 3R/2W 64 floating-point physical register file
• Execution units:
• 1 load/store, 2 simple ALU, 1 complex ALU/Branch, 1 FP
MEMORY SUB-SYSTEM
• 32 entry data and instruction TLBs and 512 entry unified L2 TLB
• Fully coherent 64KB data cache backed by a unified 4MB L2 cache
• ECC for both data and L2 caches
For details on ET-Maxion,
• Aggressive stride prefetchers for L1 and L2 data caches
see references [11,12]
OPERATING RANGE: 500 MHz to 2 GHz
19
Summary Statistics of ET-SoC-1

The ET-SoC-1 is fabricated in TSMC 7nm

• 24 billion transistors
• Die-area: 570 mm2
• 89 Mask Layers
1088 ET-Minion energy-efficient 64-bit RISC-V processors
• Each with an attached vector/tensor unit
• Typical operation 500 MHz to 1.5 GHz expected
4 ET-Maxion 64-bit high-performance RISC-V out-of-order processors
• Typical operation 500 MHz to 2 GHz expected
1 RISC-V service processor
ET-SoC-1 Die Plot
Over 160 million bytes of on-die SRAM used for caches and scratchpad memory
Root of trust for secure boot
Power typically < 20 watts, can be adjusted for 10 to 60+ watts under SW control
Package: 45x45mm with 2494 balls to PCB, over 30,000 bumps to die
• Each Minion Shire has independent low voltage power supply inputs that can be
finely adjusted to mitigate Vt variation effects and enable DVFS
Status: Silicon currently undergoing bring-up and characterization
ET-SoC-1 Package

20
Summary

The Esperanto ET-SoC-1 is the highest performance commercial RISC-V chip to date
• More RISC-V cores on a single chip
• More RISC-V aggregate instructions per second on a single chip
• Highest TOPS driven by RISC-V cores

Esperanto’s low-voltage technology provides differentiated RISC-V processors with the best performance per watt
• Energy efficiency matters!
• Best performance per watt delivers the best performance in a fixed number of watts
• Solution delivers energy efficient acceleration for datacenter inference workloads, especially recommendation

The hard part was making all the tradeoffs combining

• Processor and memory system architecture
• Circuits and techniques for low voltage operation

Esperanto now has a highly scalable design

• Efficient for ML recommendation
• Thousands of general-purpose RISC-V cores can be applied to many other highly parallel computing tasks
• Modular approach allows design to scale up and down, and to other semiconductor processes

Early Access Program for qualified customers beginning later in 2021 (for info, contact: [email protected])

21
Thanks to our Key Development Partners

Thanks to all our partners for their help in bringing our vision into reality! Sorry we can’t name
everyone!
22
Footnotes and References

[1] N. Jouppi, et al., Ten Lessons from Three Generations Shaped Google’s TPUv4i, 2021 ACM/IEEE 48th Annual International Symposium on Computer Architecture. Page 4 lesson 5 concludes “Inference
DSAs need air cooling for global scale”.
[2] J. Park, et al., Deep Learning Inference in Facebook Data Centers: Characterization, Performance Optimizations and Hardware Implications, arXiv:1811.09886v2, 29 November 2018.
[3] M. Anderson, et al., First Generation Inference Accelerator Deployment at Facebook, arXiv: 2107.04140v1, 8 Jul 2021
[4] M. Smelyanskiy, Facebook, Linley Fall Processor Conference 2019 “Challenges and Opportunities of Architecting AI Systems at Datacenter Scale”
[5] M. Smelyanskiy, AI System Co-Design: How to Balance Performance & Flexibility, AI Hardware Summit, September 17, 2019. Slide 12 indicated desired Inference characteristics with 100 TOPs of INT8,
100 MB of SRAM. Slide 19 talks about the need for programmability over fixed function hardware.
[6] Note that a core optimized for high voltage and high frequency (2-3GHz) operation will require higher power gate drive strengths to overcome wire delays and hence will have higher Cdyn than a processor
optimized for low-voltage operation. Each of the power/frequency points shown on this energy efficiency curve therefore represents a different physical design, i.e. not the same silicon, to take this changing
Cdyn into account. Designs were synthesized at high and low voltages to estimate potentially achievable frequencies. Performance at each frequency was estimated using our internal ML Recommendation
benchmark, based on running this benchmark on a full chip hardware emulation engine (Synopsys Zebu system) providing clock level accuracy at a few points and interpolating the other points. The goal was
to understand the shape of the energy efficiency curve to find voltages for best energy efficiency (Inferences/second/watt). Different benchmarks would likely have different curves, though we would expect the
overall shape to be similar. Repeating, this was a design study and does not represent any specific silicon results or design, each point on the curve is a differently synthesized design, though with the same
architecture, i.e., we used the full ET-Minion as the input to be synthesized.
[7] Estimate of 30 square feet per rack comes from “The Case for the Infinite Data Center” – Gartner, Source: Gartner, Data Center Frontier
[8] MLPerf DLRM Inference Data Center v0.7 & v1.0: https://mlcommons.org/en/
[9] Measured by MLPerf DLRM Samples / Second; FP32, Offline scores
Additional source information:
• a.1. Submitter: Intel; MLPerf DLRM score 24,630: Inference Data Center v0.7 ID 0.7-126; Hardware used (1-node-8S-CPX-PyTorch-BF16); BF16; https://mlcommons.org/en/inference-datacenter-07/
• a.2 Intel 8380H Processor TDP Power of 250W from: https://ark.intel.com/content/www/us/en/ark/products/204087/intel-xeon-platinum-8380h-processor-38-5m-cache-2-90-ghz.html
• b. Submitter: NVIDIA; T4 MLPerf DLRM score 665,646: Inference Data Center v0.7 ID 0.7-115; Hardware used (Supermicro 6049GP-TRT-OTO-29 (20x T4, TensorRT)); INT8; https://mlcommons.org/en/inference-datacenter-07/
• c. Submitter: NVIDIA; A10 MLPerf DLRM score 772,378: Inference Data Center v1.0 ID 1.0-54; Hardware used (Supermicro 4029GP-TRT-OTO-28 (8x A10, TensorRT)); INT8; https://mlcommons.org/en/inference-datacenter-10/
• d. Internal estimates by Esperanto for MLPerf DLRM: Inference Data Center v0.7; ET-SOC-1; Unverified result is from Emulated/Simulated pre-silicon projections; INT8; Result not verified by MLCommons™ Association.

[10] Measured by ResNet-50 Images per second (Esperanto INT8 Batch 8, NVIDIA INT8 Batch 8, Habana INT8 Batch 10, Intel INT8 Batch 11)
Additional measurement source information:
• a.1. Intel ResNet-50: https://software.intel.com/content/www/us/en/develop/articles/intel-cpu-outperforms-nvidia-gpu-on-resnet-50-deep-learning-inference.html
• a.2. Intel 9282 has 2 die in the package, CPU TDP power for both die from: https://ark.intel.com/content/www/us/en/ark/products/194146/intel-xeon-platinum-9282-processor-77m-cache-2-60-ghz.html
• b. NVIDIA (T4, A10) ResNet-50: https://developer.nvidia.com/deep-learning-performance-training-inference
• c. Habana ResNet-50: https://habana.ai/wp-content/uploads/2018/09/Goya-Datasheet-HL-10x-Nov14-2018.pdf
• d. Esperanto ResNet-50: Emulated/Simulated projections; INT8

[11] P. Xekalakis and C. Celio, The Esperanto ET-Maxion High Performance Out-of-Order RISC-V Processor, 2018 RISC-V Summit, presentation at https://www.youtube.com/watch?v=NjEslX_-t0Q
[12] Maxion is described in “Esperanto Maxes out RISC-V - High-End Maxion CPU Raises RISC-V Performance Bar”, Microprocessor Report, December 10, 2018.

23