Revision as of 14:51, 2 June 2018

Vulcan is a 16 nm high-performance 64-bit ARM microarchitecture designed by Broadcom and later Cavium for the server market.

Introduced in 2018, Vulcan-based microprocessors are branded as part of the ThunderX2 family.

History

Vulcan can trace its roots all the way back to Raza Microelectronics XLR family of MIPS processors from 2006. With the introduction of their XLR family in 2009, Raza (and later NetLogic) moved to a high-performance superscalar design with fine-grained 4-way multithreading support. In 2011, Broadcom acquired NetLogic Microsystems and integrated them Broadcom's Embedded Processor Group.

In 2013, Broadcom announced that they have licensed the ARMv7 and ARMv8 architectures, allowing them to develop their own microarchitectures based on the ISA. Vulcan is the outcome of this effort which involved adopting the ARM ISA instead of MIPS and enhancing the cores in various ways. Vulcan development started in early 2012 and has was expected to enter mass production in mid-2015.

In 2017 Cavium acquired Vulcan from broadcom which was introduced later that year. In early 2018, Vulcan-based microprocessor entered general availability under the ThunderX2 brand.

Architecture

Vulcan builds on the prior MIPS-based XLP II microarchitecture. The design has been substantially improved and changed to execute ARM (based on the ARMv8.1 ISA).

Key changes from XLP II

• Converted to ARM ISA (from MIPS)
  • Aarch64, Aarch32
• 16nm FinFET process (from 28 nm planar)
• 40% IPC improvement
• 25% higher clock (2.5 GHz, up from 2 GHz)
• Core
  • Longer pipeline (15 stages, up from 13)
  • Improved branch predictor
  • Double fetch throughput (4, up from 2)
  • New Decoder
    • Decodes ARMv8.1 (Instead of MIPS64 R5)
    • Decodes to micro-ops
      • Roughly 10-20% more µOPs
  • New loop buffer
  • Execution Engine
    • New scheduler
      • Unified schedule (from distributed)
        • 60 entries
    • Large ROB (180 entries, up from 100)
    • Execution Units
      • New FP Unit (2, up from 1)
      • Wider FP Units (128-bit, up from 64-bit)
  • Memory Subsystem
    • Double load bandwidth (128-bit, up from 64-bit)
    • New Store Data Unit
    • Half L2 Cache Size (256 KiB, down form 512 KiB)
• Memory Controller
  • DDR3 → DDR4
  • 4 → 8 channels
  • 1600 MT/s → 2666 MT/s
  • 47.68 GiB/s → 158.9 GiB/s

Block Diagram

Entire Chip

Individual Core

Memory Hierarchy

• Cache
  • L1I Cache
    • 32 KiB, 8-way set associative
  • L1D Cache
    • 32 KiB, 8-way set associative
  • L2 Cache
    • 256 KiB, 8-way set associative
  • L3 Cache
    • 1 MiB/core slice
    • Shared
  • System DRAM
    • 8 Channels
    • DDR4, up to 2666 MT/s
      • 1 DPC and 2 DPC support
      • RDIMM, LRDIMM, NVDIMM-P support
      • Single, Dual and Quad rank modules
    • 8 B/cycle/channel (@ memory clock)
    • ECC
• TLBs
  • ITLB
    • Dedicated instruction TLB
    • DTLB
      • TLB unit for each LSU
  • STLB
    • 2048-entry
    • 4 KiB - 16 GiB pages

Overview

Scaled up from prior architectures in all vectors (performance, area, and TDP), Vulcan was designed to be a Xeon-class ARM-based server microprocessor. Vulcan features 32 high-performance custom-designed ARM cores fully compliant with ARMv8.1 along with their accompanying 1 MiB of level 3 cache slice (for a total of 32 MiB of shared last level cache). Since each core supports up to four simultaneous threads, the full configuration can support up to 128 threads. Supporting a large number of cores, are eight DDR4 channels capable of data rates of up to 2,666 MT/s, allowing for 170.7 GB/s of aggregated bandwidth.

The processor comes with 14 fully-configurable PCIe Gen3 controllers with 56 available lanes. The chip also has 2 USB 3 and 2 SATA 3 ports.

Vulcan supports up to two-way multiprocessing through their second-generation Cavium Coherent Processor Interconnect (CCPI2) capable of providing 600 Gbps of aggregated bandwidth.

Core

Vulcan is an out-of-order superscalar with support for up to four simultaneous hardware threads. Vulcan features a 13-15-stage pipeline, slightly longer than the XLP II.

Front-end

Vulcan's front-end is tasked with fetching instructions from a ready thread instruction stream and feeding them into the decode in order to be delivered to the execution units. Since Vulcan supports up to four threads, a thread scheduler determines from which thread's instruction stream to operate on. This determination is done on each cycle with the help of the branch predictor with no added cost.

Vulcan has a 32 KiB L1 cache which is 8-way set associative and a dedicate L1 instruction TLB. It's worth noting that all the core caches on Vulcan are 8-way set associative to aid the branch predictor which works on cache line stride patterns. Under normal flow, the data to be fetched was already predicted and has made it to the L1 from the L2. The bandwidth to the L1 instruction cache is 64 bytes per cycle.

Fetch

Instruction fetch is done on a 32-byte window or 8 (4-byte) ARM instructions. This is twice the throughput of the previous architecture and is designed in order to better absorb bubbles in the pipeline. The instruction stream is decomposed into its constituent instructions where they are queued to go for the decoder. The queue is shared by all threads. The size of the queue has not been disclosed.

Decoding

Each cycle, up to four instructions are sent to the decoder. In prior design, Broadcom's products decoded MIPS instructions. With Vulcan, the switching to ARM meant the decoder had to be replaced with much more complex logic that decodes the original instruction and emits micro-ops. For the most part, there is a 1:1 mapping between instructions and µOP with an average of 15% more µOPs emitted from instructions. The extra complexity has added another pipeline stage to the decode.

Loop Buffer

Sitting between the decoder and the scheduler is a loop buffer. The loop buffer, in conjunction with the branch predictor, will queue recent tight loop operations. The buffer will play back the operations repeatedly until a branch take occurs. When this takes place, the front-end (instruction fetch, decode, etc..) is largely power-gated in order to save power. Although Broadcom originally told us the buffer was had 48-entries, when the product was re-released by Cavium in late 2018, WikiChip was unable to confirm this number.

Execution engine

Vulcan's back-end handles the execution of out-of-order operations. Vulcan's back-end has been substantially enhanced from prior designs including a complete redesign of the scheduler. Most of the improvements dealt with entirely reworking the scheduler in order to more efficiently extract additional instruction-level parallelism opportunities. From decode, instructions are sent to the Reorder Buffer (ROB) at the rate of up to 4 µOPs each cycle.

Renaming & Allocation

In the prior XLP II microarchitecture, NetLogic had a five-queue instruction distributed scheduler mechanism whereby each queue is associated with certain execution units. In Cavium, Broadcom got rid of the distributed scheduler and replaced it with a more efficient unified scheduler, similar in design to Intel's Skylake.

Vulcan's Reorder Buffer is 180-entry in size, an 80-entry increase over prior design. The ROB tracks all µOPs in flight. Vulcan renames and retires up to 4 µOPs per cycle.

Scheduler

Vulcan uses a 60-entry unified scheduler. Each cycle, up to six µOPs can be issued to the execution units, a much wider design from the four µOPs issue in prior design. It's worth noting that in order to support four threads, Vulcan duplicates most of the logic for each thread such as all the registers, architectural states, program counters, and interrupts. Ready µOPs from any thread are issued on each cycle.

Execution Units

Up to six µOPs can be sent into Vulcan's six execution units each cycle. As far as integer operations, up to three operations can be issued each cycle. One of the ALUs also handles branch instructions. In the XLP II, there were two simple integer ALUs and a single complex integer ALU unit. Only the complex integer ALU unit was able to perform operations such as multiplication and division. Though unconfirmed, it's suspected that both ALUs can now do complex integer operations as well.

Vulcan has doubled the number of floating point units to two and widened them to 128-bit to support ARM's NEON operations (prior design was only 64-bit wide). In theory, Vulcan's peak performance now stands at 8 FLOPS/cycle or 8 GFLOPS at 1 GHz.

Memory subsystem

Vulcan's memory subsystem deals with the loads and store requests and ordering. There are two load-store units each capable of moving 128-bit of data - double the bandwidth of the XLP II. The widening of the units was done in order to more efficiently support operations such as the Load Pair (LDP) and Store Pair (STP) instructions. In addition to the LSUs, there is a new dedicated Store Address unit. Similar to Intel's older architectures, the store operation is cracked into two distinct operations - a store address operation used to calculate the effective address and finally the store data operation. Vulcan can issue a store to the Store Address unit before the data is available where the address can be calculated and memory ordering conflicts can be detected. Once the data is ready, the operation will be reissued to the LSU. The store buffer is 36-entry deep with the load buffer at 64-entries for a total of 100 simultaneous memory operations in-flight or roughly 55% of all µOPs. Note that the store buffer is considerably smaller than the load buffer because Vulcan can only sustain a single load operation per cycle as most workloads do far more loads than stores.

Vulcan's L2 cache is 256 KiB, half that of prior design, and has an L2 to L1 bandwidth of 64 bytes per cycle in either direction. There is a 1 MiB L3 cache per core arranged as 2 MiB slices for a total of 32 MiB of cache shared by the entire chip.

System Architecture

Vulcan incorporates 32 cores on a single monolithic die. The cores are interconnected using a bidirectional ring bus, meaning the worst case path is halved. It appears that the cores consist of quad-core blocks with the L3 cache being the crossbar between the cores in the block. Eight of those cores are interconnected via the ring which operates at core frequency. The exact behavior has not been disclosed but in the XLP II, the ring was capable of moving 64 bytes of data each cycle per core.

Attached to the ring is 2nd-generation Cavium Coherent Processor Interconnect (CCPI2) used to support multiprocessing as well as the I/O subsystem. In its maximum configuration, the die feature 14 PCIe Gen 3 memory controllers supporting up to 56 lanes. Bifurcation is supported for up to 14 x1/x4, 7x8, or 3x16.

Scalability

Vulcan has multiprocessing support. Each die comes with 2nd-generation Cavium Coherent Processor Interconnect (CCPI2) which incorporates 24 25 Gpbs SerDes for a total aggregated bandwidth of 600 Gbps between the two sockets.

In this configuration, a max system consists of two chips for a total of 64 cores, 256 threads, 16 memory channels and 32 DIMMs, 112 PCIe Gen 3 lanes, 4 SATA3 ports, 4 USB 3 ports, and other various low speed I/O interfaces.

Die

• Broadcom's original die size was rumored to be around 600 mm². It's unknown how much the die has changed when it was modified by Cavium.
• TSMC's 16 nm process

All Vulcan Chips

This section is empty; you can help add the missing info by editing this page.

References

• Some information was obtained directly from Broadcom
• Some information was obtained directly from Cavium

See also

• Qualcomm's Falkor
• Intel's Skylake