Skip to main content

Intel Architecture Day 2021: Alder Lake Chips, Golden Cove and Gracemont Cores

Image 1 of 3

Intel Architecture

(Image credit: Intel)
Image 2 of 3

Intel Architecture

(Image credit: Intel)
Image 3 of 3

Intel Architecture

(Image credit: Intel)

Intel's goal with its Performance Core (P-Core), which comes with the Golden Cove microarchitecture, was to create the highest-performance core the company has ever built and deliver a step-function increase in CPU performance. However, the architecture also needed to be scalable with a wide dynamic range so it could scale from laptops to desktops to data centers — Intel uses this same core in its Sapphire Rapids server chips, too, albeit with modifications that tailor it for datacenter workloads.

For the client market, the company focused on optimizing the core for single-threaded performance and AI workloads. The company also wanted to expand its support for large data sets and large code footprint applications.  

As we've seen in the past, Intel followed the "Wider, Deeper, Smarter" ethos with the Golden Cove design. That includes a deeper out of order scheduler and buffers, more physical registers, a wider allocation window, and more execution ports. Intel paired those enhancements, which we'll cover below, with smarter branch prediction to better feed the execution engine, thus unlocking more parallelism and performance.

The company also ingrained a new autonomous power management microcontroller into each core that measures telemetry at microsecond granularity, instead of milliseconds. Intel says this allows them to better optimize power usage based on the actual behavior of an application, thus delivering higher average frequency for any workload. This is Intel's first per-core integrated power management microcontroller, and it is only present on the Golden Cove cores.

We surmise that this unit also feeds Intel's Thread Director tech with telemetry data. The power management controller also allows for faster frequency transitions, though Intel hasn't said just how fast. Alder Lake still adjusts frequency in 100 MHz steps, though, as opposed to Zen's 25 MHz granularity.

Here's the short list of improvements that we'll cover below.

UnitSunny CovePerformance-core
iTLB 4K pages128256
iTLB 2M/4M1632
uop Cache2.25K4K
uop cache BW68
Decoders46
Allocation56
OoO Window352512
Integer ALUs45
# of Loads2×5123×256, 2×512
L1 data Cache48KB48KB
L1 Fill Buffers1216
L1 DTLB6496
L2 cache512KB/1.25MB1.25MB / 2MB
Image 1 of 2

Intel Architecture

(Image credit: Intel)
Image 2 of 2

Intel Architecture

(Image credit: Intel)

We'll jump ahead here and show you Intel's claimed results, and then show you how Intel got there. Intel claims that Alder Lake's Golden Cove performance core offers an average of a 19% performance improvement over the Cypress Cove architecture found in Rocket Lake when both chips operate at the same frequency (ISO frequency). Intel derived this average from a wide distribution of workloads that includes SPEC CPU 2017, Sysmark 25, Crossmark, PCMark 10, WebXPRT3, and Geekbench, among other unlisted benchmarks.

As with all vendor-provided benchmarks, you'll have to take this with a grain of salt. But this could be promising if Intel can deliver near these levels of performance improvements, assuming it can squeeze high enough clock rates out of the Intel 7 silicon and workloads can be correctly targeted to these performance cores, of course.

Intel claims this jump is larger than the improvement from Skylake to Sunny Cove. That's impressive, if true. The second image in the album above comes from our lab results in our CPU Benchmark hierarchy. Intel's Rocket Lake Core i9-11900K, which has the Cypress Cove architecture that is very similar to Sunny Cove, currently leads the single-threaded performance hierarchy against AMD's chips. A quick glance at Intel's previous-gen Core i9-10900K shows that Intel's jump from Skylake to Sunny Cove represented a large improvement that helped it take the lead from AMD's Zen 3 architecture. That means Alder Lake's single-threaded performance could be well-positioned against AMD's Zen 4 chips if it can pull off a comparable advance.

Front End

(Image credit: Intel)

Intel improved the micro-op supply from both the decoder and the micro-op cache. That starts by doubling length decode to 32B per cycle and adding two additional decoders (6) to enable 6 decodes per cycle. The micro-op cache also now feeds 8 micro-ops per cycle instead of 6 and can hold 4K micro-ops, up from 2.25K before. This helps to feed the out of order engine quicker and increases the micro-op hit rate.

To support software with a large code footprint, Intel doubled the number of 4K pages and large pages stored in the iTLB (as listed in the table above) and enhanced code prefetch by increasing from tracking 5K branch targets to 12K branch targets.

Intel says it improved branch prediction to reduce jump mispredicts and significantly overhauled the Branch Target Buffer (BTB). The BTB serves as a cache for the prefetcher, and Intel more than doubled it to help with large code footprints. Additionally, the BTB now has a machine learning algorithm that allows it to expand or contract capacity based upon usage, thus fine-tuning power and performance.

Out of Order Engine

(Image credit: Intel)

Intel widened the Out of Order (OoO) engine from a 5- to 6-wide allocate unit that then feeds into the scheduler, which now feeds 12 execution ports, as opposed to 10 with Sunny Cove. The engine is also deeper with a 512-entry reorder-buffer, up from 352, and more physical registers.

Intel also says that it has enabled collapsing dependency chains by executing some simple instructions at the rename/allocation stage to save resources further down the pipeline. This allows other operations to run faster and save power by better utilizing the instruction resources. Intel is playing coy about this new technique, though, and hasn't answered follow up questions regarding its new technique.

Integer and Vector Execution Units

Image 1 of 2

Intel Architecture

(Image credit: Intel)
Image 2 of 2

Intel Architecture

(Image credit: Intel)

Intel added a fifth general integer execution port with an ALU and LEA. All five LEAs are single-cycle and can be used for additions and subtractions, or fast multiplications with fixed numbers. On the vector side, Intel infused two more fast adders (FADD) on ports 1 and 5. Intel also bulked up the FMAs with support for FP16 operations with AVX-512 instructions, but that is confined to the server versions of the chip. AVX-512 is disabled for Alder Lake processors.

L1 and L2 Cache and Memory Subsystems

Image 1 of 2

Intel Architecture

(Image credit: Intel)
Image 2 of 2

Intel Architecture

(Image credit: Intel)

The L1 cache is now wider with 3 load ports instead of 2, and deeper with larger Load and Store Buffers. Intel increased the L1 data TLB by 50% and the L1 data cache can fetch 25% more misses in parallel. The L1-D cache also has an enhanced prefetcher that can now service four page table walks instead of two, which is good for workloads with irregular data sets.

The consumer chips, like Alder Lake, will come with 1.25MB of cache per performance core, while data center chips get 2MB. The lower-capacity cache for client results in improved latency, while the higher-capacity cache for datacenter helps feed chips with higher core counts. Intel says its new prefetch engine observes running programs to identify potential future memory access patterns, and then can prefetch down multiple potential paths to service those potential requests.

Image 1 of 3

Intel Architecture

(Image credit: Intel)
Image 2 of 3

Intel Architecture

(Image credit: Intel)
Image 3 of 3

Intel Architecture

(Image credit: Intel)

Intel's Advanced Matrix Extension (AMX) technology is disabled in the Golden Cove cores present in Alder Lake but is active for the Sapphire Rapids datacenter processors. AMX is a next-gen AI accelerator that enables hardware-accelerated matrix multiplication operations that significantly enhance AI inference and training performance. Like AVX, AMX incurs a clock frequency penalty. To reduce jitter, Intel has worked to reduce the impact of the AMX license levels on clock rates, largely by adopting a more fine-grained per-core power control scheme. We'll dive deeper into this technology in our Sapphire Rapids coverage.

Paul Alcorn

Paul Alcorn is the Deputy Managing Editor for Tom's Hardware US. He writes news and reviews on CPUs, storage and enterprise hardware.