Ryzen 3000 series is out

[cma6]

To Dann & mwyoung:
Are you guys concerned about potential NUMA issues that have been raised about Threadripper with A/B engines like SF
at https://www.tomshardware.com/news/amd-c ... 37895.html ?

[mwyoung]

To Dann & mwyoung:
Are you guys concerned about potential NUMA issues that have been raised about Threadripper with A/B engines like SF
at https://www.tomshardware.com/news/amd-c ... 37895.html ?

I don't think any of the Threadripper's 16, 24, and 32 core had any issues running A/B engines. I can speak to the fact that the 2950x has no issues running engines in dynamic or Numa.

It was my understanding that most of the issues were resolved. This was due to a windows scheduling issue. When running on Linux the 24, and 32 core Threadripper's performed much better.

[Dann Corbit]

To Dann & mwyoung:
Are you guys concerned about potential NUMA issues that have been raised about Threadripper with A/B engines like SF
at https://www.tomshardware.com/news/amd-c ... 37895.html ?

http://www.talkchess.com/forum3/viewtop ... =2&t=71256
The threadrippers already do very well at chess,
The issues with threadrippers were only connected to the very high end parts at the very high end core counts,
The memory issues may have been addressed according to the link above,
If true, then the threadrippers will be the undisputed kings of alpha-beta unless you go over $20K

[Gian-Carlo Pascutto]

AnandTech included SPECint2017 benchmarks, which include Deep Sjeng:

https://images.anandtech.com/graphs/gra ... 111160.png

The results were...surprisingly bad, i.e. clearly worse than Intel cores. I wonder why.

I guess they compiled for BMI, but SSE3 builds will be 25% faster.

What you are saying makes no sense.

Deep Sjeng doesn't use BMI, BMI2 nor SSE3. The SPEC version doesn't even use popcount because there's no 100% portable way to express the intrinsic.

It's an architecture issue, but I don't know what. Intel is 17% faster clock for clock, and this is a lot. And to make it even more confusing, for Leela AMD is faster.

[Gian-Carlo Pascutto]

AnandTech included SPECint2017 benchmarks, which include Deep Sjeng:

https://images.anandtech.com/graphs/gra ... 111160.png

The results were...surprisingly bad, i.e. clearly worse than Intel cores. I wonder why.

This is one thread benchmark?

It's actually a bit confusing:

"Moving on to the 2017 suite, we have to clarify that we’re using the Rate benchmark variations. The 2017 suite’s speed and rate benchmarks differ from each other in terms of workloads. The speed tests were designed for single-threaded testing and have large memory demands of up to 11GB, while the rate tests were meant for multi-process tests. We’re using the rate variations of the benchmarks"

But I do believe they are running those single-threaded, as otherwise the scores don't make sense. They say later that the runtime and memory demands of the rate benchmarks are less and this makes the tests easier to run. This matches the expectation that they are running the (faster/lower memory) rate benchmarks single-threaded.

For Deep Sjeng, the only difference is the size of the hash table.

[Laskos]

AnandTech included SPECint2017 benchmarks, which include Deep Sjeng:

https://images.anandtech.com/graphs/gra ... 111160.png

The results were...surprisingly bad, i.e. clearly worse than Intel cores. I wonder why.

This is one thread benchmark?

It's actually a bit confusing:

"Moving on to the 2017 suite, we have to clarify that we’re using the Rate benchmark variations. The 2017 suite’s speed and rate benchmarks differ from each other in terms of workloads. The speed tests were designed for single-threaded testing and have large memory demands of up to 11GB, while the rate tests were meant for multi-process tests. We’re using the rate variations of the benchmarks"

But I do believe they are running those single-threaded, as otherwise the scores don't make sense. They say later that the runtime and memory demands of the rate benchmarks are less and this makes the tests easier to run. This matches the expectation that they are running the (faster/lower memory) rate benchmarks single-threaded.

For Deep Sjeng, the only difference is the size of the hash table.

Yes, otherwise I don't understand the scores. Do they say if they allowed the Boost? i9 9900k on 1-2 cores can go to 5.0GHz, while Ryzen 3900X 1 core to 4.6GHz. If they allowed the boost, many single-thread benchmarks will look bad for Ryzen in this comparison. Leela is a curiosity in this benchmark (all three CPUs look weird, I am not sure whether to trust the Leela bench, Leela speeds are a bit tricky to measure).

[dragontamer5788]

AnandTech included SPECint2017 benchmarks, which include Deep Sjeng:

https://images.anandtech.com/graphs/gra ... 111160.png

The results were...surprisingly bad, i.e. clearly worse than Intel cores. I wonder why.

I guess they compiled for BMI, but SSE3 builds will be 25% faster.

What you are saying makes no sense.

Deep Sjeng doesn't use BMI, BMI2 nor SSE3. The SPEC version doesn't even use popcount because there's no 100% portable way to express the intrinsic.

It's an architecture issue, but I don't know what. Intel is 17% faster clock for clock, and this is a lot. And to make it even more confusing, for Leela AMD is faster.

https://www.anandtech.com/show/14605/th ... -the-bar/5

We ran our original review numbers with the latest available firmware for the MSI MEG X570 ACE motherboard last week (Version 7C35v11). On Saturday the 6th MSI had shared with us a notice about a new version coming out, which became available to download to us on Sunday the 7th, the launch day and date of publication of the review.

We’ve had more time to investigate the new firmware, and have discovered extremely large changes in the behaviour of the frequency boosting algorithm. The new firmware (Version 7C35v12) for the motherboard contains AMD’s new ComboPI1.0.0.3.a (AGESA) firmware.

We notice a significant change in the CPU’s boosting behaviour, now boosting to higher frequencies, and particularly at a faster rate from idle, more correctly matching AMD’s described intended boost behaviour and latency.

We’re currently in the process of re-running all our suite numbers and updating the article where necessary to reflect the new frequency behaviour.

Anandtech tested all the benchmarks, and then later noticed that the boosting behavior was weird. New firmware versions seem to properly boost the clocks. So it seems like early motherboard firmware was just buggy for Ryzen 3000... buggy in a way that hampered performance.

There was a later article which summarizes the improvements: https://www.anandtech.com/show/14632/am ... date-recap

[Gian-Carlo Pascutto]

I'm already talking about the revised numbers. The original ones (now deleted) showed that the 3700X was the same speed as the 2700X, which obviously couldn't be true.

In the new data it clearly gains, but not enough to catch Intel.

[Gian-Carlo Pascutto]

Yes, otherwise I don't understand the scores. Do they say if they allowed the Boost? i9 9900k on 1-2 cores can go to 5.0GHz, while Ryzen 3900X 1 core to 4.6GHz. If they allowed the boost, many single-thread benchmarks will look bad for Ryzen in this comparison. Leela is a curiosity in this benchmark (all three CPUs look weird, I am not sure whether to trust the Leela bench, Leela speeds are a bit tricky to measure).

Why would they disallow boost? I'm pretty sure that's enabled.

Most of these benchmarks show the Zen 2 as very close to the Intel cores, so no, Leela's numbers are more normal. Besides Deep Sjeng, only perl and omnet seem to be worse on AMD. But obviously I care more about Deep Sjeng.

There is nothing "tricky" about measuring the speed of the Leela in SPECint. It was *specifically* *designed* to be a repeatable, deterministic benchmark. It runs the exact same workload, you check whether the end results are fully correct and count the seconds to get the answer. The ratio of speedup compared to the reference platform is the result. That's how SPEC works.

They're not just fetching a github release and compiling it. They run a specific version for benchmarking that I submitted to them a while ago. Same for all other benchmarks in the SPEC suite.

You are right about the boost difference, when I said 17% clock for clock that was just wrong because there's no indication clocks were normalized.

[Laskos]

Yes, otherwise I don't understand the scores. Do they say if they allowed the Boost? i9 9900k on 1-2 cores can go to 5.0GHz, while Ryzen 3900X 1 core to 4.6GHz. If they allowed the boost, many single-thread benchmarks will look bad for Ryzen in this comparison. Leela is a curiosity in this benchmark (all three CPUs look weird, I am not sure whether to trust the Leela bench, Leela speeds are a bit tricky to measure).

Why would they disallow boost? I'm pretty sure that's enabled.

Most of these benchmarks show the Zen 2 as very close to the Intel cores, so no, Leela's numbers are more normal. Besides Deep Sjeng, only perl and omnet seem to be worse on AMD. But obviously I care more about Deep Sjeng.

There is nothing "tricky" about measuring the speed of the Leela in SPECint. It was *specifically* *designed* to be a repeatable, deterministic benchmark. It runs the exact same workload, you check whether the end results are fully correct and count the seconds to get the answer. The ratio of speedup compared to the reference platform is the result. That's how SPEC works.

They're not just fetching a github release and compiling it. They run a specific version for benchmarking that I submitted to them a while ago. Same for all other benchmarks in the SPEC suite.

You are right about the boost difference, when I said 17% clock for clock that was just wrong because there's no indication clocks were normalized.

Ah, ok. I was thinking that they improvised something hard to control with Leela, as even people in Lc0 or Leela Go community often compare not very well defined results. Such issues as the network used, time to fixed task or task to fixed time, tree size, the compile, etc. can mess up the results. But it is not the case with you sending the benchmarking.

All in all, I am not that disappointed in the results. Yes, Deep Sjeng result for 3900X is bit off even compared to 2700X (single threaded I expected 7% increase in clock and an additional 10%+ from IPC increase, but it comes at a meager total of 8%). But a well parallelized AB engine running on 12 AMD cores will still run some 30% faster than 8 core Intel, and Leela will be at least on par. The price is the same, so AMD wins pretty convincingly. And let's see the full load multi-thread performance, there the Boost advantage of Intel might pale when boosting all the cores. Not bad, and for example 2700X starts to go these days for $260, half the price of Intel i9-9900k for 80-85% performance.