Hot Chips 2020 Live Blog: AMD Ryzen 4000 APU (Noon PT)

02:49PM EDT - Next up is AMD's Ryzen 4000 Renoir talk.

03:02PM EDT - Sonu Arora from AMD, AMD Fellow

03:02PM EDT - SoC architect

03:03PM EDT - Best notebook processor selling today

03:03PM EDT - First 8-core x86 for ultrathin notebooks

03:03PM EDT - Beaten goals we set for ourselves

03:03PM EDT - Evolution based on desire to break linear trend of performance

03:04PM EDT - Initially planned with 6 cores, but SoC/Power curves enabled 8 cores

03:04PM EDT - Made room for ST, MT, GPU, and Battery Life

03:04PM EDT - Renoir vs Picasso

03:04PM EDT - +25% ST, +200% MT, +27% GPU, 20% less SoC Power

03:05PM EDT - Decide how much to push ST without compromising GPU or Power

03:06PM EDT - Renoir is a complete SoC

03:06PM EDT - 8C/16T with 8MB L3

03:06PM EDT - 2nd gen Multimedia, 3rd gen Audio

03:07PM EDT - Added more IO to keep up with modern notebook requirements

03:07PM EDT - PCIe up from 16 to 20 in mobile package, 20 to 24 in desktop package

03:08PM EDT - Covered Zen 2 at Hot Chips last year

03:08PM EDT - 512 KB L2/core

03:08PM EDT - 4 cores in a core complex

03:09PM EDT - 2 core complexes into Renoir

03:10PM EDT - +15% IPC ST on top of +10% FMax

03:10PM EDT - Low power design methodology, can turn off segments when not in use

03:10PM EDT - Better idle-to-turbo latency

03:11PM EDT - Performance vs Intel's 10th Gen (Comet) at 15W and 45W

03:11PM EDT - Improved gaming experience

03:12PM EDT - Reduce CUs from 11 to 8, but higher frequency, better IPC, optimized devices for throughput and power, 77% higher peak bandwidth

03:12PM EDT - 2 render backends and 1MB L2 cache

03:12PM EDT - Optimize for pJ/bit

03:13PM EDT - -61% area vs 11CUs on 14nm, but +27% perf

03:13PM EDT - total 225% perf density on graphics

03:14PM EDT - Support DDR4-3200 and LPDDR4-4267

03:14PM EDT - 75% better power efficiency for Infinity Fabric

03:14PM EDT - Lot of time on the floor plan to reduce memory latency

03:15PM EDT - IF power gating focused for GPU bandwidth

03:16PM EDT - Renoir delivers CPU perf and GPU perf

03:16PM EDT - Different power states in Renoir in order to optimize exit/enter latency

03:16PM EDT - OS can decide to go to deeper low power states on lower idle

03:17PM EDT - (deeper sleep takes longer to reach peak power)

03:17PM EDT - 20% reduced SoC power

03:17PM EDT - aggressive L3 power gating

03:17PM EDT - lower power IOs

03:19PM EDT - STAPM uses system temperature in order to budget the turbo and boost

03:20PM EDT - 4x increase in boost duration

03:20PM EDT - Also simplifies OEM design for thermal boosting

03:20PM EDT - 3rd gen Audio

03:21PM EDT - Supports Native USB-C

03:21PM EDT - DP 1.4 with HBR3

03:21PM EDT - Four QHD60, or three QHD144

03:22PM EDT - next gen multimedia

03:22PM EDT - AMD Integrated Device Translation (AIDA)

03:22PM EDT - Enables Microsoft PlayReady with reduced UMA dedicated memory frame buffer

03:23PM EDT - AMD 25x20 energy efficiency initiative. AMD achieved 31.7x, beating 25x target from 2014

03:24PM EDT - Renoir built across 8 different AMD R&D Centers and offices

03:24PM EDT - Q&A time

03:27PM EDT - Q: Melt, Spectre? A: Melt doesn't impact AMD. Zen 2 has hardware mitigations for Spec v2/v4. Added Guest mode execution, virtualization security, IOMMU throughput, TME

03:29PM EDT - Q: 8 cores in this size, trade offs? A: We wanted to provide the 8 core perf, and as we analyzed with 7nm that if we took care we could enable 8 cores. Lots of content creation takes advantage of 8 cores. These come in handy with high perf applications. On the GPU, we were also careful in perf/watt and perf/mm2 that we can drive. We figured out in 7nm we could get higher frequency, so we balanced resources and that's why we went for 8 CUs. Vega with mem bandwidth gives better UX in 15W

03:30PM EDT - Next rpesentation is Tiger Lake!