Sunny Cove Microarchitecture: A Peek At the Back End

One of the best parts when a new microarchitecture is announced is learning how it all works and how it has changed. Intel has been on close variants of the Skylake core since Skylake first launched in 2015, which has carried through Kaby Lake, Coffee Lake, and Coffee Lake refresh – and so getting an update to his old core design is more than welcome. While I applaud Intel for getting something new to put in front of our faces, the level of information was unfortunately less than complete. For reasons of ‘we’ll discuss more nearer the launch’, Intel was only set on discussing the back end of the microarchitecture design in any detail. In core design parlance, this means everything beyond the dispatch and order queues.

Intel framed its microarchitecture updates into two different sections: general purpose performance increases and special purpose performance increases. The company described the general purpose updates essentially as raw IPC (instructions per clock) throughput or an increase in frequency. An increase in either of these two leads to a performance increase no matter the code, at least as long as it is compute bound. Frequency is often a function of the implementation and process, whereas IPC increases can come from cores being wider (more executing instructions per clock), deeper (more parallelism per clock), and smarter (better data delivery through the front end).

The other type of performance increase is the special purpose kind, which means that certain workloads used in specific scenarios can be improved through additional acceleration methods, such as dedicated IP or specialized instructions.

Sunny Cove is set to have both.

On the back end of the microarchitecture design, Intel has done several things including increasing cache sizes, making the execution of the core wider, increasing L1 store bandwidth, and other things.

Starting with the cache improvements, the Sunny Cove core back end has a 48 KB L1 Data cache, upgraded from 32 KB. Normally cache misses decrease by a factor of a square root of the proportional size when the cache is increased, so in this case we should expect a 22% reduction in L1-Data cache misses. We are told that the L2 cache size has also been increased in both the Core and Xeon variants of Sunny Cove (256 KB and 1 MB respectively), although to exactly what size will be revealed at a later date.

While not back-end, the micro-op (uOp) cache has increased from a 2048-entry design, although exactly to what Intel is not saying at this point. The second level TLB size has also increased, which will help with machine address translation of recent addresses – normally this is increased when more lookups need to be kept and stored, which means that Intel identified a number of use cases where recent addresses had been evicted shortly before being needed.

From the diagram we can see some other changes, such as the increase in execution ports from eight to ten, allowing more instructions to come out of the schedulers at one time. Port 4 and Port 9 is linked to a cycling data store, doubling that bandwidth, but also the AGU store capabilities have doubled, for which the increased L1-D size will help. Previously on Skylake there was a potential bottleneck when all three AGUs could attempt a store, however only enough bandwidth was present for one per clock cycle. Load performance is unaffected.  There is also an allocation increase, moving from four to five. This means that the dispatch out of the reorder buffer can now hit five instructions per cycle, rather than four – however given that the ops/cycle out of the Skylake decoders and uOp cache was four-to-five (depending on certain factors), it will be interesting to see how the front end has changed as a result.

There are also substantial changes in the execution ports between Skylake and Sunny Cove, which is best represented in a diagram and a table:

Back-End Execution Resources
Skylake Port Sunny Cove
Integer
  JMP Shift ALU 0 ALU LEA Shift JMP
    LEA ALU 1 ALU LEA Mul iDIV
  Mul LEA ALU 5 ALU LEA MulHi  
  JMP Shift ALU 6 ALU LEA Shift JMP
Vector / FP
fpDIV Shift ALU FMA 0 FMA ALU Shift fpDIV
  Shift ALU FMA 1 FMA ALU Shift Shuffle
  Shuffle ALU FMA 5 FMA ALU Shuffle  

The parts highlighted in bold are different in the Sunny Cove core.

We see that Intel has equipped the Integer section of the core with more LEA units to help with memory addressing calculations, possibly to help improve performance loss with security mitigations that require frequent memory calculations or to help with high-performance array code with constant offsets. Port 1 gets the Mul unit from Skylake Port 5, probably for rebalancing purposes, but there is also an integer divider unit here. This addition is a bit of a tweak, as we know that Cannon Lake has a 64-bit IDIV unit in its design as well. In that case it brings 64-bit integer division down from 97 clocks (a mix of instructions) down to 18 clocks, and this is likely something similar. Also on the INT side is that Port 5’s multiply unit has become a ‘MulHi’ unit which in other implementations leave most-significant nibble of the result in a register for further use. I'm not sure exactly what its position is here in the Sunny Cove core though.

On the FP side, Intel has increased the shuffle resources, which we are told was because they had customers requesting this functionality to help remove bottlenecks in their code. Intel didn’t state the capabilities for the FMAs in the FP part of the core, however we do know that the core has an AVX-512 unit, so at least one of these FMAs should interface with it. Cannon Lake only has one 512-bit FMA, and that most likely carries over here, and the Xeon Scalable version might get two. I would like to get proved wrong though.

Other updates to the core as listed from Intel include improvements to the branch predictor and a reduced ‘effective load latency’ (due to the TLB and L1-D), although it was pointed out that these improvements won’t help everyone, and might require new algorithms in order to use specific parts of the core.

We’re waiting to see what changes Intel has made on the front-end, which is where a lot of low-hanging fruit often lies for performance.

On top of the microarchitecture differences, there are new instructions for the Sunny Cove core to help speed up specialist compute tasks. With the AVX-512 unit in play, the new core will support IFMA instructions for big arithmetic compute, and which are also useful in cryptography. Along the same lines, Vector-AES (supporting more AES concurrently) and Vector Carryless Multiply are also supported. SHA and SHA-NI instructrions are also added, along with Galois Field instructions, which are fundamental building blocks in some elements of cryptography as well.

For Sunny Cove, there is also built-in support for larger memory capacities. The paging table for main memory is now a five layer design, up from four layers, supporting a Linear Address space up to 57 bits and a physical address space up to 52 bits. This means, according to Intel, that the server processors could theoretically support 4 TB of memory per socket. We will see Sunny Cove hit the market in the server space with Ice Lake-SP in 2020 according to Intel’s previous Xeon roadmaps.   For security, Sunny Cove has access to Multi-Key Total Memory Encryption, and User Mode Instruction Prevention.

So where does this leave us in terms of Sunny Cove predicted performance compared to Skylake at a clock for clock level? Unfortunately we don’t have enough information. There is only enough to whet the appetite. Intel wants you (and me) to come back for more in 2019.

The CPU Core Roadmaps The Next Generation Gen11 Graphics: Playable Games and Adaptive Sync!
POST A COMMENT

149 Comments

View All Comments

  • prisonerX - Wednesday, December 12, 2018 - link

    And you're a special brand of pathetic. Congrats.

    Don't worry though, one day you'll hit puberty and things will improve.
    Reply
  • LogitechFan - Wednesday, December 12, 2018 - link

    Oh yes, amdturds are THE breed of stupid to behold... Reply
  • drunkenmaster - Wednesday, December 12, 2018 - link

    Would that be Marvell's chip from 3 years ago that suddenly people are talking about... which is just a chip on an interposer with an HBM chip. HBM which AMD co-developed, had prototype APU packages using it in 2011 and launched Fury X with HBM on an interposer 3 years ago.

    Right, but AMD didn't pioneer it, because Marvell made a dramatically less complex switch using the same concept at the same time AMD did, but you know, AMD co-developed the memory, they worked with the packaging plants for years to be able to mass produce interposer packaged products and Fury X was the reason HBM went into full production in the first place.... but sure, AMD definitely didn't pioneer this latest move.
    Reply
  • ajc9988 - Wednesday, December 12, 2018 - link

    http://www.eecg.toronto.edu/~enright/micro14-inter... http://www.eecg.toronto.edu/~enright/Kannan_MICRO4... https://youtu.be/G3kGSbWFig4 https://seal.ece.ucsb.edu/sites/seal.ece.ucsb.edu/... https://www.youtube.com/watch?v=d3RVwLa3EmM&t=... Reply
  • qap - Wednesday, December 12, 2018 - link

    Well ... yeah. Intel had multi-chip CPUs almost 15 years ago (look for Pentium D). And it was mocked by AMD and its fans.
    I thought that the mocking intel did last month was just to remind AMD, what they said then and I found it hilarious. But maybe not. People don't change and it's completely plausible, that it was said by someone, who does't know the history.
    Reply
  • Topweasel - Wednesday, December 12, 2018 - link

    Never Mocked by AMD. They just made sure to point out that their 2 cores with the X2 and quad with the Phenom where "real" multicore chips. It mattered more back then because Intel's boards had FSB's instead of direct connections and the separate chips had to talk through the FSB and chipset northbridge to talk to each other. Also AMD hadn't been rambling on thier glue tech for years (with no product yet on market using it (Kaby-G isn't actual EMIB)) when they made their statements.

    AMD promoted their products said they are better because they were real. There is a negative connotation with that. But that is different then calling out your competitor for using "glue".
    Reply
  • 29a - Wednesday, December 12, 2018 - link

    "Well ... yeah. Intel had multi-chip CPUs almost 15 years ago (look for Pentium D). And it was mocked by AMD and its fans."

    I think pretty much everyone mocked anything P4 related, you have to admit it was pretty bad.
    Reply
  • FreckledTrout - Wednesday, December 12, 2018 - link

    What's P4? To flush it down the toilet. Reply
  • tshoobs - Wednesday, December 12, 2018 - link

    Never heard that one before! So true, the P4 was such an embarrassment. Reply
  • JlHADJOE - Wednesday, December 12, 2018 - link

    lol this takes me right back to the days of the P4, and the awesome "x is good... for me to POOP on!" meme Reply

Log in

Don't have an account? Sign up now