Hi folks,
I talked to my crystal ball and here is what it said about AMD and what can we expect from them in order to remain competitive and even surpass Intel in both performance and power efficiency in certain CPU segments. Jokes aside, these are just my expectations and visions and are based on my imagination and knowledge only, but not on any AMD roadmap or insider information.
With the money earned in 2017, the good development and production partners, the right team motivated, AMD have great chances to surprise not only us, the customers, but the entire semiconductor industry. AMD had enough time to discover most of the Zen weaknesses and to find out easy-implementable solutions and fixes for most. It now depends mostly on their ambitions to compete with Intel. I believe that with a little effort they can now bring to the table much better Zen+ SKUs than Zen did. Like they said, Zen was only the beginning and it already is competing quite well in all segments. Step by step, AMD are going to fix most of the macro and micro-architectural issues of Zen with every silicon iteration and here I'm talking about the issues I expect AMD to fix with a silicon update in the following 2018.
1) AMD Ryzen 2000A series desktop APU
I expect the first Ryzen Desktop APUS on the sAM4 to be the same Ryzen 2xxxU mobile series APUs, tweaked for higher performance and power envelope. Some of these should include all 11 CUs enabled on the GPU part and higher GPU, CPU base/boost/xFR and DRAM clocks compared to the mobile variants. Since the same chips will be used, these desktop APUs will have all the power and efficiency improvements, like Precision Boost 2, present in the mobile Ryzen 2000U APUs.
The CPU part will have higher clocks and voltages compared to the mobile parts. It's performance, depending on the workload, will vary between the performance of a Ryzen 5 1500X/Core i3 8100 and a Ryzen 1600/Core i5 8400. The GPU performance on the, best performing, Ryzen 5 2400AX will vary between a dedicated RX550/GT1030 and a dedicated RX560/GTX1050 on resolutions up to 1080p.
The Ryzen 2000A desktop APUs, just like the 2000U mobile APUS, will have per core voltage and frequency regulation. This feature will enable the OC-ers to squeeze the maximal possible frequency at the lowest possible voltage for each CPU core and the GPU independently.
The CPU cores on most APUs will achieve 4GHz stable with some APUs hitting 4.1GHz on all cores at acceptable voltages and temperatures for a everyday use.
Most of the APUs regardless of the number of CUs on their GPU will achieve stable GPU clocks ranging from 1500 to 1600MHz.
Overclocking the DRAM controller and using a faster DDR4-3200 MHz (51.2 GB/s max bandwidth) will have small impact on the CPU performance, but great on the GPU and should bring almost linear ~20% framerate increase in GPU limited scenarios.
Such OC-ed APU will be powerfull enough to run most of todays DX12 titles at 1080p at aceptable framerates and visuals. The lowest performing R3 2100A combined with a mediocre B350 motherboard and a cheap 2x4GB DDR4-2666 will be the best low budget solution for multimedia, office and light gaming. Having very low total system power consumption will park these APUs in many mini-ITX mainboards and slim mini ITX w/o room for dedicated VGA(for example: https://www.in-win.com/en/computer-chassis/bq-series/USA), some of them even with fanless solutions.
2) AMD Ryzen 2000 desktop CPUs (12nm FinFET Zen+)
The Ryzen 2000 CPUs include the same macro-architectural improvements of the Ryzen 2000 series APUs and will bring some new, while the micro-architecture of the cores will remain same. The greatest improvement of all is going to be the new InfinityFabric 2 with Data Fabric 2 which will operate at double the clock speed of Data Fabric 1 on Infinity Fabric 1 on the Ryzen 1000/2000 CPUs/APUs or at the DRAM clock speed, providing twice the I/O bandwidth to everything connected with the DataFabric 2. This will result in:
- CCX to CCX and die to die absolute latencies to be cut to half
- CCX to CCX and die to die bandwidth to be doubled
- RAM access latency to be reduced by up to 50% depending of the CPU package (MCM or not) and configuration (UMA vs NUMA for ThreadRipper) and DRAM clock and latencies.
- RAM bandwidth to be slightly improved
These improvements will have a huge impact on gaming performanses, especially at 1080p and 1440p where Ryzen 1000 CPUs are bottlenecking the VGA in most modern games. Ryzen 2000 with the same number of cores/threads and clocked the same will perform 10-20% better in games on the very same system. At 4K it will not only match a same clocked, and same number of cores/threads, Coffe Lake/Skylake-X CPUs, but in some scenarios it will even outperform them.
In productivity and other kind of software IF2 will not bring any noticeable performance improvements. Ryzen TR and EPYC CPUS based on Zen+ might see more gains than Ryzen 3/5/7 in certain workloads which involve intensive data transfers across multiple dies and CCXs by significantly reducing the access time and doubling the transfer bandwidth between the CCX's of diferent dies and their IMCs.
Other than the CPU peroformance improvements, the new AMD 400 series chipsets will bring greater and faster connectivity thanks to the doubled bandwidth to the CPU by the PCIe 4.0 x4 connection.
The second, very important upgrade on the Ryzen 2000 CPUS is the improved, 12nm LPP FinFET, production process compared to the Ryzen 1000/2000A 14nm FinFET. The new process will bring 5-10% higher clocks(read performance) using the same power or 10-20% lower power consumption at the same clocks. I expect 8 core SKUs with 4.1GHz boost clock on all cores and 4.3GHz XFR boost on two cores.
ThreadRipper will get the best 5% of all dies, so 4.4GHz 4-core XFR boost will be common for all TR CPUs (except the 2900X which will have 2-core XFR boost). There might be even a 8-core SKU with 2 cores per CCX hitting 4.5GHz.
These upgraded Ryzen TR 2000 CPUs will give a very hard time to Intel Skylake-X CPUs with the same number of cores. 7960X in average will get outperformed by TR 2960X , while the 7920X will run toe-to-toe with the 2930X. All Skylake-X CPUs will get a massive pricecut inorder to remain competitive with the ThreadRipper 2000 Zen+ CPUs.
So folks, what do you think ?
I talked to my crystal ball and here is what it said about AMD and what can we expect from them in order to remain competitive and even surpass Intel in both performance and power efficiency in certain CPU segments. Jokes aside, these are just my expectations and visions and are based on my imagination and knowledge only, but not on any AMD roadmap or insider information.
With the money earned in 2017, the good development and production partners, the right team motivated, AMD have great chances to surprise not only us, the customers, but the entire semiconductor industry. AMD had enough time to discover most of the Zen weaknesses and to find out easy-implementable solutions and fixes for most. It now depends mostly on their ambitions to compete with Intel. I believe that with a little effort they can now bring to the table much better Zen+ SKUs than Zen did. Like they said, Zen was only the beginning and it already is competing quite well in all segments. Step by step, AMD are going to fix most of the macro and micro-architectural issues of Zen with every silicon iteration and here I'm talking about the issues I expect AMD to fix with a silicon update in the following 2018.
1) AMD Ryzen 2000A series desktop APU
I expect the first Ryzen Desktop APUS on the sAM4 to be the same Ryzen 2xxxU mobile series APUs, tweaked for higher performance and power envelope. Some of these should include all 11 CUs enabled on the GPU part and higher GPU, CPU base/boost/xFR and DRAM clocks compared to the mobile variants. Since the same chips will be used, these desktop APUs will have all the power and efficiency improvements, like Precision Boost 2, present in the mobile Ryzen 2000U APUs.
The CPU part will have higher clocks and voltages compared to the mobile parts. It's performance, depending on the workload, will vary between the performance of a Ryzen 5 1500X/Core i3 8100 and a Ryzen 1600/Core i5 8400. The GPU performance on the, best performing, Ryzen 5 2400AX will vary between a dedicated RX550/GT1030 and a dedicated RX560/GTX1050 on resolutions up to 1080p.
The Ryzen 2000A desktop APUs, just like the 2000U mobile APUS, will have per core voltage and frequency regulation. This feature will enable the OC-ers to squeeze the maximal possible frequency at the lowest possible voltage for each CPU core and the GPU independently.
The CPU cores on most APUs will achieve 4GHz stable with some APUs hitting 4.1GHz on all cores at acceptable voltages and temperatures for a everyday use.
Most of the APUs regardless of the number of CUs on their GPU will achieve stable GPU clocks ranging from 1500 to 1600MHz.
Overclocking the DRAM controller and using a faster DDR4-3200 MHz (51.2 GB/s max bandwidth) will have small impact on the CPU performance, but great on the GPU and should bring almost linear ~20% framerate increase in GPU limited scenarios.
Such OC-ed APU will be powerfull enough to run most of todays DX12 titles at 1080p at aceptable framerates and visuals. The lowest performing R3 2100A combined with a mediocre B350 motherboard and a cheap 2x4GB DDR4-2666 will be the best low budget solution for multimedia, office and light gaming. Having very low total system power consumption will park these APUs in many mini-ITX mainboards and slim mini ITX w/o room for dedicated VGA(for example: https://www.in-win.com/en/computer-chassis/bq-series/USA), some of them even with fanless solutions.
2) AMD Ryzen 2000 desktop CPUs (12nm FinFET Zen+)
The Ryzen 2000 CPUs include the same macro-architectural improvements of the Ryzen 2000 series APUs and will bring some new, while the micro-architecture of the cores will remain same. The greatest improvement of all is going to be the new InfinityFabric 2 with Data Fabric 2 which will operate at double the clock speed of Data Fabric 1 on Infinity Fabric 1 on the Ryzen 1000/2000 CPUs/APUs or at the DRAM clock speed, providing twice the I/O bandwidth to everything connected with the DataFabric 2. This will result in:
- CCX to CCX and die to die absolute latencies to be cut to half
- CCX to CCX and die to die bandwidth to be doubled
- RAM access latency to be reduced by up to 50% depending of the CPU package (MCM or not) and configuration (UMA vs NUMA for ThreadRipper) and DRAM clock and latencies.
- RAM bandwidth to be slightly improved
These improvements will have a huge impact on gaming performanses, especially at 1080p and 1440p where Ryzen 1000 CPUs are bottlenecking the VGA in most modern games. Ryzen 2000 with the same number of cores/threads and clocked the same will perform 10-20% better in games on the very same system. At 4K it will not only match a same clocked, and same number of cores/threads, Coffe Lake/Skylake-X CPUs, but in some scenarios it will even outperform them.
In productivity and other kind of software IF2 will not bring any noticeable performance improvements. Ryzen TR and EPYC CPUS based on Zen+ might see more gains than Ryzen 3/5/7 in certain workloads which involve intensive data transfers across multiple dies and CCXs by significantly reducing the access time and doubling the transfer bandwidth between the CCX's of diferent dies and their IMCs.
Other than the CPU peroformance improvements, the new AMD 400 series chipsets will bring greater and faster connectivity thanks to the doubled bandwidth to the CPU by the PCIe 4.0 x4 connection.
The second, very important upgrade on the Ryzen 2000 CPUS is the improved, 12nm LPP FinFET, production process compared to the Ryzen 1000/2000A 14nm FinFET. The new process will bring 5-10% higher clocks(read performance) using the same power or 10-20% lower power consumption at the same clocks. I expect 8 core SKUs with 4.1GHz boost clock on all cores and 4.3GHz XFR boost on two cores.
ThreadRipper will get the best 5% of all dies, so 4.4GHz 4-core XFR boost will be common for all TR CPUs (except the 2900X which will have 2-core XFR boost). There might be even a 8-core SKU with 2 cores per CCX hitting 4.5GHz.
These upgraded Ryzen TR 2000 CPUs will give a very hard time to Intel Skylake-X CPUs with the same number of cores. 7960X in average will get outperformed by TR 2960X , while the 7920X will run toe-to-toe with the 2930X. All Skylake-X CPUs will get a massive pricecut inorder to remain competitive with the ThreadRipper 2000 Zen+ CPUs.
So folks, what do you think ?
Last edited: