Intel Unveils 10nm Sunny Cove CPU Architecture With Gen11 Graphics, Major Performance Lift
First up is Sunny Cove, which is Intel’s next generation Core CPU platform. Interestingly, this is the first that we’d ever heard of this internal codename, but it is “deeper, wider, and smarter” than its predecessors, and will form the basis of both Xeon (server) and Core (consumer) processors starting in the second half of 2019. Intel says that Sunny Cove will allow for further increases in performance-per-clock (IPC) and will include dedicated registers and new instruction sets to accelerate artificial intelligence (AI), compression processing, and cryptography workloads (via AES and SHA-NI instructions).
Sunny Cove is based on Intel’s 10nm manufacturing process, which will be a welcome relief from multiple generations of 14nm++ process-built products. Other enhancements that will be of great importance to increasing overall performance will include greater parallelism, further increases to buffer, instruction, and data cache sizes, overall reductions in latency, and better branch prediction accuracy. Allocation units in the architecture have increased from 4 to 5, while execution ports have increased from 8 to 10 (Haswell had previously increased the numbers of ports from 6 to 8). Sunny Cove also can issue 2 loads and 2 stores per clock cycle (versus the previous gen 2 loads and 1 store).
On the memory side of things, linear address space has jumped from 48 bits to 57 bits, while physical address space has been upped to 52 bits. As you might expect, Sunny Cove will feature the hardware-level mitigations for Meltdown and Spectre that have already been implemented in Whiskey Lake and Cascade Lake, but will also include further hardening of the architecture in an effort to prevent future exploits.
Complementing Sunny Cove will be Intel’s next-generation integrated graphics (Gen11), which will begin shipping in 2019. Compared to existing Gen9 graphics, Gen11 increases the number of enhanced execution units from 24 to 64, while pushing compute performance to over 1 TFLOPS. While this performance might still pale in comparison to competing discrete graphics solutions, this is a huge leap in performance for systems that will rely on integrated graphics performance exclusively, from notebooks to desktops. Intel is obviously now keenly focused on better gaming performance and the company has really stepped up its driver game as well, with far more frequent updates, performance enhancements, increased compatibility, and Day 0 support for many of the latest games.
Intel Gen11 graphics also will include support for tile-based rendering as well as adaptive sync technology that will offer similar tear- and jitter-free gaming experiences to AMD FreeSync and NVIDIA G-Sync, which is important, especially when frame rates drop below the connected panel's refresh rates. The good news is, Intel's adaptive sync technology will not require any external hardware or incur and up-charge and will also be compatible with AMD FreesSync monitors. When you stop to think about it, with the wealth of low cost FreeSync capable panels on the market these days, NVIDIA could be out in the cold with its proprietary G-Sync solution.
Intel Sunny Cove Validation Platform
Intel Gen11 Sunny Cove (Left) Vs Gen9 (Right) was significantly faster at 1080p
Interestingly, it was noted that a Gen11 EU (Execution Unit) comprises only about 75% of the silicon area of a Gen9 EU. However, all told, with over 2X the graphics compute resources, Gen11 will deliver significant performance gains. Intel was demonstrating Tekken 7 on Intel Sunny Cove with Gen11 graphics, versus Gen9 Coffee Lake-based machines and it was clear frame rates were dramatically better. Though not specifically confirmed, if you do the math on the EU counts, you can expect at least 2X the performance versus previous Intel IGPs and Intel did note this is in a similar power envelope.
Also on board Gen11's more efficient EUs is enhanced mixed precision performance (Int8 and Int16), along with better FPU throughput, for improvements in AI workload processing, as well as further enhancements to Intel's VPU media engine for higher performance and more efficient HEVC video decode and encode in hardware.
And of course, a solid software development and support ecosystem behind all this new silicon is critical. A philosophy that SVP and GM of Intel's Core and Visual Computing Group, Raja Koduri focused on was the mantra of "no transistor left behind." Intel's current OpenVINO SDK is a good example of this strategy, but the slide above articulates the overreaching goal -- a single software suite to support the many different compute architectures now offered by Intel, from CPUs, to GPUs, FGPAs and specialized application accelerators.