Semi dynamics high bandwidth vector capable RISC-V cores
0 likes274 views
SemiDynamics introduced two new RISC-V cores, AVISPADO 220 and ATREVIDO 220, both supporting the upcoming RISC-V Vector spec version 1.0. AVISPADO 220 is an in-order core with a technique called "Gazzillion Misses" that allows a high number of outstanding memory requests. ATREVIDO 220 is an out-of-order core also utilizing Gazzillion Misses. SemiDynamics also provides a customizable RISC-V Vector Processing Unit that implements the vector spec and can be integrated with the cores. Both cores and the vector unit are available for licensing.
![semidynamics
RISC-V Summit 2020
semidynamics
Example: Gathering Sparse Data with 8 outstanding requests
7
while ( condition )
load a[index[i]]
load b[i]
compute
store c[i]
Original Program
top:
load index[i] -> x5
load (x5) -> x6
load b[i] -> x7
compute x6, x7 -> x8
store x8 -> c[i]
loop to top
Stylized Assembly Code
Wait “L” cycles, where L=memory latency, L > 100
Time →
After the 8th miss, stop until first miss comes back
(loop will be unrolled ~4 times)
Now send the 9th miss to
the memory system](https://image.slidesharecdn.com/semidynamicshighbandwidthvector-capablerisc-vcores-210419141759/85/Semi-dynamics-high-bandwidth-vector-capable-RISC-V-cores-7-320.jpg)
![semidynamics
RISC-V Summit 2020
semidynamics
Example: Gathering Sparse Data with Gazzillion Misses™
8
while ( condition )
load a[index[i]]
load b[i]
compute
store c[i]
Original Program
top:
load index[i] -> x5
load (x5) -> x6
load b[i] -> x7
compute x6, x7 -> x8
store x8 -> c[i]
loop to top
Stylized Assembly Code
Wait time greatly reduced
Time →
After the 8th miss, continue sending requests!](https://image.slidesharecdn.com/semidynamicshighbandwidthvector-capablerisc-vcores-210419141759/85/Semi-dynamics-high-bandwidth-vector-capable-RISC-V-cores-8-320.jpg)
Semi dynamics high bandwidth vector capable RISC-V cores
- 1. semidynamics RISC-V Summit 2020 SemiDynamics High Bandwidth Vector-capable RISC-V Cores Roger Espasa, PhD CEO 1
- 2. semidynamics RISC-V Summit 2020 semidynamics SemiDynamics • European Based RISC-V IP Provider • HQ in Barcelona • Founded 2016 • Proud participant in the European Processor Initiative (EPI) • RISC-V Acceleration core supplied by SemiDynamics 2
- 3. semidynamics RISC-V Summit 2020 semidynamics Introducing 2 new RISC-V IP Cores November 11, 2020 3 Available for licensing 2-wide In-Order 2-wide Out-of-Order Gazzillion Misses™ AVISPADO 220 ATREVIDO 220 Gazzillion Misses™ RISC-V Vector Support RISC-V Vector Support
- 4. semidynamics RISC-V Summit 2020 semidynamics What’s a high bandwidth core? 4
- 5. semidynamics RISC-V Summit 2020 semidynamics A core with lots of outstanding memory requests 5
- 6. semidynamics RISC-V Summit 2020 semidynamics 6 A core with Gazzillion Misses™
- 7. semidynamics RISC-V Summit 2020 semidynamics Example: Gathering Sparse Data with 8 outstanding requests 7 while ( condition ) load a[index[i]] load b[i] compute store c[i] Original Program top: load index[i] -> x5 load (x5) -> x6 load b[i] -> x7 compute x6, x7 -> x8 store x8 -> c[i] loop to top Stylized Assembly Code Wait “L” cycles, where L=memory latency, L > 100 Time → After the 8th miss, stop until first miss comes back (loop will be unrolled ~4 times) Now send the 9th miss to the memory system
- 8. semidynamics RISC-V Summit 2020 semidynamics Example: Gathering Sparse Data with Gazzillion Misses™ 8 while ( condition ) load a[index[i]] load b[i] compute store c[i] Original Program top: load index[i] -> x5 load (x5) -> x6 load b[i] -> x7 compute x6, x7 -> x8 store x8 -> c[i] loop to top Stylized Assembly Code Wait time greatly reduced Time → After the 8th miss, continue sending requests!
- 9. semidynamics RISC-V Summit 2020 semidynamics Another Example: Vector Gather with Gazzillion Misses™ • Consider the RISC-V “vlxei32.v” gather instruction • Assume VLEN=512b • Instruction will request 16 addresses (assume to different cache lines) 9 Time → After the 8th miss, continue sending requests! At larger VLEN or smaller indices, even larger gains
- 10. semidynamics RISC-V Summit 2020 semidynamics Comparison to other cores 10 1 2 3 4 5 6 Retirement/Issue Width 20 40 60 80 100 120 140 POWER9 SMT4 16 12 IceLake 64+ Avispado 128+ Atrevido A76 20 RISC-V Boom 10 Outstanding Misses U8 SCR7 AX25
- 11. semidynamics RISC-V Summit 2020 semidynamics Gazzillion Misses™ good for… 11 Recommendation Systems MLP AE CNN RNN LSTM GRU DRL Key-Value Stores Sparse Data / HPC Machine Learning
- 12. semidynamics RISC-V Summit 2020 semidynamics Gazzillion Misses™ good for SoCs… 12 With Limited SRAM/Cache Core Bus Ctrl D D R Accel High Bandwidth/ Streaming With Vector Units Vector Unit Vector Unit D D R
- 13. semidynamics RISC-V Summit 2020 semidynamics Avispado Core Details 13
- 14. semidynamics RISC-V Summit 2020 semidynamics AVISPADO 220 RISCV64GCV 14 INT FP MIQ GIQ VFIQ 64B 8B PC I$ ITLB BP Queue Dec 2i D$ DTLB Align vAGU Gazzillion Coherent Interface • SV48 • 16KB I$ • Decodes upcoming V1.0 vector spec • 32KB D$ • Full hardware support for unaligned accesses • Coherent (TL, ACE, CHI) Available for licensing
- 15. semidynamics RISC-V Summit 2020 semidynamics Customizable upon request • Cache Size / Associativity • Coherent Interface • In-flight instructions 15
- 16. semidynamics RISC-V Summit 2020 Introducing SemiDynamics’ RISC-V Vector Processing Unit 16 November 11, 2020 SemiDynamics Confidential
- 17. semidynamics RISC-V Summit 2020 semidynamics SemiDynamics’ VPU • Implements the RISC-V Vector 1.0 Specification • Including lmul, segmented loads • No vector atomics currently • Ready for OOO support • Renaming • Customizable settings • VLEN = vreg size = from 128b to 4096b • Data path width = from 128b to 512b • Fast cross-lane network for slide/rgather/compress/expand November 11, 2020 17 Available for licensing 3 months after V-spec freeze SemiDynamics Confidential
- 18. semidynamics RISC-V Summit 2020 semidynamics VPU Block Diagram • Lane based organization • Full cache-line bus from D-cache • Units per lane • FMA • INT • DIV • XL: Cross-lane (rgather, …) • Full masking support November 11, 2020 18 From D-cache SemiDynamics Confidential Lane0 XL Network VREG0 A B C Bypass Lane1 LaneN-1 FMA INT XL DIV Available for licensing 3 months after V-spec freeze
- 19. semidynamics RISC-V Summit 2020 semidynamics AVISPADO 220 with VPU RISCV64GCV 19 INT FP MIQ GIQ VFIQ 64B 8B PC I$ ITLB BP Queue Dec 2i D$ DTLB Align vAGU Gazzillion Coherent Interface • SV48 • 16KB I$ • Decodes upcoming V1.0 vector spec • 32KB D$ • Full hardware support for unaligned accesses • Coherent (TL, ACE, CHI) • Vector Memory (vle, vlse, vlxe, vse, …) processed by MIQ/LSU VPU November 11, 2020 SemiDynamics Confidential Available for licensing 3 months after V-spec freeze
- 20. semidynamics RISC-V Summit 2020 semidynamics Atrevido Core Details 20
- 21. semidynamics RISC-V Summit 2020 semidynamics ATREVIDO 220 RISCV64GCV 21 INT FP MIQ GIQ VFIQ 64B 8B PC I$ ITLB BP Queue Ren 2i D$ DTLB Align vAGU Gazzillion Coherent Interface • SV48 • 16KB I$ • Decodes upcoming V1.0 vector spec • 32KB D$ • Full hardware support for unaligned accesses • Coherent (TL, ACE, CHI) Dec Available for licensing • Larger BP • Renaming • Retirement Logic
- 22. semidynamics RISC-V Summit 2020 semidynamics ATREVIDO 220 with VPU RISCV64GCV 22 INT FP MIQ GIQ VFIQ 64B 8B PC I$ ITLB BP Queue Ren 2i D$ DTLB Align vAGU Gazzillion Coherent Interface • SV48 • 16KB I$ • Decodes upcoming V1.0 vector spec • 32KB D$ • Full hardware support for unaligned accesses • Coherent (TL, ACE, CHI) Dec • Larger BP • Renaming • Retirement Logic Available for licensing 3 months after V-spec freeze VPU
- 23. semidynamics RISC-V Summit 2020 semidynamics Summary 23 IP cores available for licensing 2-wide In-Order 2-wide Out-of-Order Gazzillion Misses™ AVISPADO 220 ATREVIDO 220 Gazzillion Misses™ VPU 1.0 VPU 1.0
- 24. semidynamics RISC-V Summit 2020 semidynamics Thank you! www.semidynamics.com [email protected] 24