AMD Zen Microarchiture Part 2: Extracting Instruction-Level Parallelism
by Ian Cutress on August 23, 2016 8:45 PM EST- Posted in
- CPUs
- AMD
- x86
- Zen
- Microarchitecture
Execution, Load/Store, INT and FP Scheduling
The execution of micro-ops get filters into the Integer (INT) and Floating Point (FP) parts of the core, which each have different pipes and execution ports. First up is the Integer pipe which affords a 168-entry register file which forwards into four arithmetic logic units and two address generation units. This allows the core to schedule six micro-ops/cycle, and each execution port has its own 14-entry schedule queue.
The INT unit can work on two branches per cycle, but it should be noted that not all the ALUs are equal. Only two ALUs are capable of branches, one of the ALUs can perform IMUL operations (signed multiply), and only one can do CRC operations. There are other limitations as well, but broadly we are told that the ALUs are symmetric except for a few focused operations. Exactly what operations will be disclosed closer to the launch date.
The INT pipe will keep track of branching instructions with differential checkpoints, to cut down on storing redundant data between branches (saves queue entries and power), but can also perform Move Elimination. This is where a simple mov command between two registers occurs – instead of inflicting a high energy loop around the core to physically move the single instruction, the core adjusts the pointers to the registers instead and essentially applies a new mapping table, which is a lower power operation.
Both INT and FP units have direct access to the retire queue, which is 192-entry and can retire 8 instructions per cycle. In some previous x86 CPU designs, the retire unit was a limiting factor for extracting peak performance, and so having it retire quicker than dispatch should keep the queue relatively empty and not near the limit.
The Load/Store Units are accessible from both AGUs simultaneously, and will support 72 out-of-order loads. Overall, as mentioned before, the core can perform two 16B loads (2x128-bit) and one 16B store per cycle, with the latter relying on a 44-entry Store queue. The TLB buffer for the L2 cache for already decoded addresses is two level here, with the L1 TLB supporting 64-entry at all page sizes and the L2 TLB going for 1.5K-entry with no 1G pages. The TLB and data pipes are split in this design, which relies on tags to determine if the data is in the cache or to start the data prefetch earlier in the pipeline.
The data cache here also has direct access to the main L2 cache at 32 Bytes/cycle, with the 512 KB 8-way L2 cache being private to the core and inclusive. When data resides back in L1 it can be processed back to either the INT or the FP pipes as required.
Moving onto the floating point part of the core, and the first thing to notice is that there are two scheduling queues here. These are listed as ‘schedulable’ and ‘non-schedulable’ queues with lower power operation when certain micro-ops are in play, but also allows the backup queue to sort out parts of the dispatch in advance via the LDCVT. The register file is 160 entry, with direct FP to INT transfers as required, as well as supporting accelerated recovery on flushes (when data is written to a cache further back in the hierarchy to make room).
The FP Unit uses four pipes rather than three on Excavator, and we are told that the latency in Zen is reduced as well for operations (though more information on this will come at a later date). We have two MUL and two ADD in the FP unit, capable of joining to form two 128-bit FMACs, but not one 256-bit AVX. In order to do AVX, the unit will split the operations accordingly. On the counter side each core will have 2 AES units for cryptography as well as decode support for SSE, AVX1/2, SHA and legacy mmx/x87 compliant code.
Facebook
Twitter
RSS
106 Comments
View All Comments
Bulat Ziganshin - Wednesday, August 24, 2016 - link
I think it's obvious from number of ALUs that 40% improvement is for scalar single-thread code that greartly bemnefits from access to all 4 integer ALUs. Of course, it will get the same benefit fro any code running up to 8 threads (for 8-core Zen). But anyway it should be slower than KabyLake since Intel spent much more time optimizng their CPUsFor m/t execution, improvements will be much smaller, 10-20%, i think. Plus, 8-core CPU will probably run at smaller frequency than 4-core Buldozers or 4-core KabyLake. AFAIK, even Intel 8c-ore CPUs run at 3.2 GHz only, and it's after many years of power optimization. We also know that *selected* Zen cpus run at 3.2 GHz in benchamrks. So, i expect either < 3 GHz frequency, or 200 Wt power budget
atomsymbol - Wednesday, August 24, 2016 - link
"For m/t execution, improvements will be much smaller, 10-20%, i think."There are bottlenecks in Bulldozer-family when a module is running two threads. An improvement of 40% for m/t Zen execution in respect to Bulldozer m/t execution is possible. It is a question of what the baseline of measurement is.
Bulat Ziganshin - Wednesday, August 24, 2016 - link
M/t execution in Bulldozer already can use all 4 INT alus, so i think that 40% IPC improvement is impossible. In other words, if s/t IPC improved by 40% by moving from 2 alu to 4 alu arrangement, m/t performance that keeps the same 4 alu arrangement, hardly can be improved by more than 20%looncraz - Wednesday, August 24, 2016 - link
IPC is NOT MT, it is ST only.IPC is per-core, per-thread, per-clock, instruction retire rate... which generally equates to performance per clock per core per thread.
Bulat Ziganshin - Thursday, August 25, 2016 - link
you can measure instruction per cycles for a thread, 3 threads, core, cpu or anything else. what's a problem??My point is that s/t speed on Zen is improved much more than m/t speed, compared to last in Bulldozer family. So, they advertized improvement in s/t speed, that is 40%. And m/t improvement is much less since it still the same 4 alus (although many other parts become wider).
atomsymbol - Thursday, August 25, 2016 - link
AMD presentation was comparing Zen to Broadwell in a m/t workload with all CPU cores utilized.From
http://www.cpu-world.com/Compare/528/AMD_A10-Serie...
you can compute that the Blender-specific speedup of Zen over a previous AMD design is about 100/38.8=2.57
Bulat Ziganshin - Thursday, August 25, 2016 - link
Can you compute IPC improvement, that we are discussing here?looncraz - Thursday, August 25, 2016 - link
Except you're absolutely wrong, the performance increases will be much higher for MT than ST.Bulldozer was hindered by the module design, so you had poor MT scaling - not an issue with Zen. On top of that, Zen has SMT, which should add another 20% or so more MT performance for the same number of cores.
A 40% ST improvement for Zen could easily mean a 100% performance improvement for MT.
Bulat Ziganshin - Saturday, August 27, 2016 - link
It's not "on top of that". Zen is pretty simple Bulldozer modification that fianlly allowed to use all 4 scalar ALUs in the module for the single thread. It's why scalar s/t perfroamnce should be 40% faster. OTOH, two threads in the module still share those 4 scalar ALUs as before, so m/t perfromance cannot improve much. On top of that, module was renamed to core. So, there are 2x more cores now and of course m/t performance of entire CPU will be 2x higherNenad - Thursday, September 8, 2016 - link
It is possible that AMD already count SMT (hyperthreading) into those 40%.Their slide which states "40% IPC Performance Uplift" also lists all things that AMD used to achieve those 40%...and first among those listed things is "Two threads per core". So if AMD already counted that into their 40% IPC uplift, then 'real' IPC improvement (for single thread) would be much lower.