Computex 2018: AMD TR 32 core CPU , Intel 28 core 5ghz CPU announced

[NikosD]

A few comments regarding the 32C/64T release of TR2.

It seems that is a niche CPU, focused on specific workloads like rendering, encoding, compiling even more than 16C release.

Not only because of doubling the number of cores but also due to the specific implementation.

The compromise of drop-in compatibility to the existing X399 motherboards, led to reduced Infinity Fabric bandwidth and increased latency of compute dies with a performance impact on memory intensive workloads.

I saw a huge performance advantage over the new 2950X 16C CPU (72%) and 18C Intel's CPU using the new Blender Beta benchmark similar to Corona, POV-ray, Cinema4D etc but also saw a performance drop below 2950X (!) level in certain tasks.

Windows scheduler makes things worse but fortunately (and unfortunately because MS is involved) I read this paragraph in a review:
AMD continues working with Microsoft to route threads to the die with direct-attached memory first, and then spill remaining threads over to the compute dies.

Unfortunately, the scheduler currently treats all dies as equal, operating in Round Robin mode.

As a result, even moderately-threaded applications can suffer at the hands of high memory latency and low throughput.

This is further complicated by thread migration.

According to AMD, Microsoft has not committed to a timeline for updating its scheduler

32C/64T desktop performance for 1800$ is unforseen, but it needs to pay attention to the workload.

But who can beat 16C/32T of 2950X at 900$ ?

[Atak_Snajpera]

Is that x264 benchmark still around? Does it have multiple instances as well?
no just single instance. I'm planning to update x265 FHD Benchmark so each instance (except 5th) will run on specific NUMA node.

[NikosD]

Redmond, we've got a problem...
(unless the problem is the poor developer's NUMA implementation)

7zip compression benchmark running on 2990WX (32C/64T)

@Atak_Snajpera We definitely need this NUMA aware version of your x265 benchmark.

https://s22.postimg.cc/4yxav5g4x/7zip.png

[Cary Knoop]

Redmond, we've got a problem...
(unless the problem is the poor developer's NUMA implementation)

7zip compression benchmark running on 2990WX (32C/64T)

@Atak_Snajpera We definitely need this NUMA aware version of your x265 benchmark.

https://s22.postimg.cc/4yxav5g4x/7zip.png
The question is whether Redmond and other software vendors were notified in a timely fashion by AMD.

[NikosD]

The platform of Windows hasn't been called "Wintel" by accident.

Linux distributions didn't have previous notice in order to implement a proper NUMA aware OS.

And Ryzen processors based on Zen architecture have been released since March 2017 and Threadripper processors since August 2017.

It's been a year already.

[huhn]

what has windows to gain from intentionally slowing down AMD CPUs...

this CPU is a niche product with low priority on a consumer grade OS.

linux/unix OS are used on super computer and mainframes where windows has no market share to speak of and this is a typical CPU used there and NUMa is quite important there.
if this has anything todo with this result. because numa is supported since win7.

[NikosD]

what has windows to gain from intentionally slowing down AMD CPUs...

Money from Intel, I wrote it above.
Intel is trying to buy (literally) some time by all means (literally) in order to fix their 10nm broken process and to be competitive again with a new architecture.

It's not the first time for Intel and Nvidia, they are full in unethical moves in order to dominate their markets.

How do you think they become this big ?


if this has anything todo with this result. because numa is supported since win7.

The problem is obviously in the implementation of NUMA and Windows scheduler.

Windows are always ready for Intel processors, in a very strange way.

Like x265, in a strange way too.

[Atak_Snajpera]

It is not really a x265 fault that by default it uses all numa nodes. The issue is that AMD released another odd cpu after AMD FX.
I do not know if you remember but before patch for win7 FX-8150 was being detected as 8C/8T CPU. This meant that all integer units were treated as full cores. After patch AMD FX 8xxx were "downgraded" to 4C/8T.

https://www.anandtech.com/show/5448/the-bulldozer-scheduling-patch-tested

M$ should implement better way of detecting topology of those newer "non-standard" cpus. Currently scheduler is clearly not aware what NUMA has direct access to memory.

[amichaelt]

Money from Intel, I wrote it above.

Which makes no sense. Microsoft pulls in twice as much revenue as Intel as has at least 10 times as much cash on hand. They aren't exactly struggling to make ends meet.

[NikosD]

It is not really a x265 fault that by default it uses all numa nodes. The issue is that AMD released another odd cpu after AMD FX.
I do not know if you remember but before patch for win7 FX-8150 was being detected as 8C/8T CPU. This meant that all integer units were treated as full cores. After patch AMD FX 8xxx were "downgraded" to 4C/8T.

https://www.anandtech.com/show/5448/the-bulldozer-scheduling-patch-tested

M$ should implement better way of detecting topology of those newer "non-standard" cpus. Currently scheduler is clearly not aware what NUMA has direct access to memory.

Yes, AMD is the underdog and needs sometimes to make "odd" CPUs.

But, in a very strange way, every "odd" CPU from AMD is odd for MS and SW vendors but every odd CPU from Intel is like an everyday CPU.

Look at the collaboration of Intel and x265 developers.
They worked together to optimize it for AVX 512.

But the same piece of software is not optimized for Ryzen, we are still waiting for it one and a half year after the release.

With Threadripper and its NUMA architecture is worse of course.

The FX processor was once again a lot faster in Linux out of the box versus Windows.

Intel makes the rules for everything on CPUs, but if it wasn't AMD we would stay to Pentiums and x86 architecture another 10 years at least.

Like the scheme of Core i3 2C/4T, Core i5 4C/4T and Core i7 4C/8T.

That scheme lasted from 2011 until Ryzen.

Oh and the HEDT category.

You could buy a 10C/20T Core i7 6950X for 1800$ before Threadripper, which is the cost for the 32C/64T from AMD right now.

Think about it.

Intel is the most ridiculous tech company along with Nvidia of course.

[huhn]

what do you want intel to do not optimise for there architecture
?
what should x265 do ignore help should they ignore there biggest customer base?
ryzen supports AVX2 and it is slower than the intel version and that'S the newest instruction.

you are aware that intel cpu are not rarely run better on linux too just like amd CPUs...

there is a lot shady stuff happening and most can't be seen here. but this is not shady.

what new CPU feature should they focus on? AVX512 was already in CPUs in 2015. amd even retired there 3Dnow!.

[NikosD]

So...The mythical monster chip that Intel was hiding in the closet is just a poor old Skylake X/Cascade Lake (?) 28C/56T CPU, unlocked and overclocked in order to justify its skyrocketed price and power consumption.

Intel will release it on December though, so we have to wait a little more to see who is going to justify again Intel's decisions regarding new CPUs.

[huhn]

zen 2 looks like it is fixing quite some issues that zen has.

up to 8 core per die 256 bit AVX.

so it's no longer a dream that we my end up with a CPU with a single CCX with no ram latency limitation.

to bad they only showed rome.

[Atak_Snajpera]

But you will get 16C (2 dies + IO die) for the price of 9900k. Regarding AVX. Yes x265 will work noticeable faster now!

[huhn]

i'm pretty sure i'll be fine with 8 cores.

[nevcairiel]

Mainstream just doesn't need 16 cores at this time. It barely needs 8. It would be rather annoying if the only offering they have is two CPU dies, and lower core counts being half-disabled. I rather have an optimized 8-core in one die with no inter-die communication troubles.
IMHO Ryzen should stay 8-cores and anyone that wants 16 can get ThreadRipper.

[huhn]

if i understood it correctly PCIe is still on the cores so two dies are not welcome to me.

[Atak_Snajpera]

Mainstream just doesn't need 16 cores at this time. It barely needs 8. It would be rather annoying if the only offering they have is two CPU dies, and lower core counts being half-disabled. I rather have an optimized 8-core in one die with no inter-die communication troubles.
IMHO Ryzen should stay 8-cores and anyone that wants 16 can get ThreadRipper.

Who cares what mainstream needs! Streamers will love 16c ryzen.
Even overclocked 9900k is too slow for 1080p60fps with preset medium. BTW. Few years ago people were also saying mainstream does not need 16 thread cpu. Without AMD you would be stuck in never ending 4 core era. AMD may release 16c ryzen just to show middle finger to intels 9900k.

INTEL: Look AMD we also have 16 thread cpu!
AMD: HOLD MY BEER!

AMD did similar move with 2990WX (32C/64T) this year. In this case middle finger to Intels 18C/36T HEDT.

[nevcairiel]

Who cares what mainstream needs!

99% of everyone, thats who. We already have 16c CPUs, its called ThreadRipper.

High-end gaming CPUs is where AMD needs to catch up right now, and more cores is not going to do it anymore, not in 2018/19. A 16c Ryzen would be no competition to a 9900k, not for gaming, not unless they can also close the ST gap.
An 8c with higher IPC and higher clock, that would show up a 9900k. Cores is easy, and cores is also not everything.

I'm sure they could easily release a 16c Ryzen if they wanted to. I'm just saying thats not what I want to buy. I want a really fast 8c with no compromises (ie. no half-disabled dies, no inter-die latency problems, etc), with supreme ST performance as well as decent MT, because a lot of things I do are still ST bound.

I don't care about stupid fights who has the most cores or whatever. I care about a CPU that makes sense for my needs, which align with the majority of high-end gamers as it happens.

[Atak_Snajpera]

Threadripper has 4 channels while Ryzen only 2. So segmentation is maintained. I do not understand why you complain about performance in games. 2700x with improved turbo boost work very well in games. Performance is slower due to weaker 128bit avx unit.(many newer games use AVX). Zen2 fixes that.

[huhn]

if you bottleneck the GPU ryzen 2 is fine but if you want to run very high FPS like 240/144 hz stable it simply can't compete the latency issues are simply killing it.

and newer games don't use AVX in a way that it is the main speed concern the issue for AMD currently is ram latency and the clockspeed. pre sandy bridge CPU are still performing as expected in new games.

or do you really think OC ram in CPU limited scenarios are magical making AVX much faster on AMD but not on intel and just look at threadripper when it is used with all cores it so unbelievable slow in games.

[Atak_Snajpera]

and newer games don't use AVX in a way that it is the main speed concern the issue for AMD currently is ram latency and the clockspeed. pre sandy bridge CPU are still performing as expected in new games.

Frostbite engine (for example BFV strongly uses AVX) and Unreal Engine 4 use AVX.
https://software.intel.com/en-us/articles/intel-software-engineers-assist-with-unreal-engine-419-optimizations

[huhn]

this was march 2018 so at an even later date this engine is used in games and can you remember a game that updated there unity engine and got a major performance boost that can be pointed at AVX? me neither.

how can you claim BFV is using AVX >heavily< if it isn't even out yet.

[Atak_Snajpera]

this was march 2018 so at an even later date this engine is used in games and can you remember a game that updated there unity engine and got a major performance boost that can be pointed at AVX? me neither.

how can you claim BFV is using AVX >heavily< if it isn't even out yet.

Have you ever heard about open beta? Some guy analyzed BFV in Vtune (if I remember correctly) and noticed that AVX instructions are used a lot more than in BF1. I wouldn't be surprised if new tomb raider was also optimized for avx. Some claim that Project Cars 2 also uses avx.

[huhn]

the open beta start tomorrow.

and again i'm not saying AVX isn't used i'm just questioning if it a major concern.

battlefield one runs as expected on a sandy bridge CPU which doesn't have AVX2. i'm not aware of a any prove that performance differences between generations is from AVX differences not clockspeed/IPC.

[Atak_Snajpera]

the open beta start tomorrow.

and again i'm not saying AVX isn't used i'm just questioning if it a major concern.

battlefield one runs as expected on a sandy bridge CPU which doesn't have AVX2. i'm not aware of a any prove that performance differences between generations is from AVX differences not clockspeed/IPC.

You must be living in alternative universe
https://www.ea.com/games/battlefield/news/battlefield-5-open-beta-announcement

https://www.youtube.com/watch?v=cAsyo8gIyys&feature=em-comments

https://i.postimg.cc/y8HXdKs5/Untitled-1.png

[huhn]

ok sorry this game is not on my rader and never will be on it be i seriously see a big problem in this random comment

so let's assume everything is 100 % AVX code in new games.
because there is no word about AVX2 the problem for zen is AVX2 not AVX 128 bit is not good but fine or at least on par with jaguar.

the next thing... battlefield 5 is a console game what do consoles use as a CPU AMD jaguar what can an AMD jaguar do native AVX not AVX2.

and the PC gaming margin is only 10% maybe 20% it so meaningless compared to console and they will optimise for the console. they will look into the PC port in there spare time.

than we have tests like this and not random comments: https://www.youtube.com/watch?v=U4k_ErEg-FU

and they doesn't even try to get high framerates totally ignoring the possibility to play at lower graphic settings to get your target framerate.

[nevcairiel]

so let's assume everything is 100 % AVX code in new games.
because there is no word about AVX2 the problem for zen is AVX2 not AVX 128 bit is not good but fine or at least on par with jaguar.

Both AVX and AVX2 are 256-bit. AVX is floating point, AVX2 is integer and FMA. In simplified terms.

[Atak_Snajpera]

ok sorry this game is not on my rader and never will be on it be i seriously see a big problem in this random comment

so let's assume everything is 100 % AVX code in new games.
because there is no word about AVX2 the problem for zen is AVX2 not AVX 128 bit is not good but fine or at least on par with jaguar.

the next thing... battlefield 5 is a console game what do consoles use as a CPU AMD jaguar what can an AMD jaguar do native AVX not AVX2.

and the PC gaming margin is only 10% maybe 20% it so meaningless compared to console and they will optimise for the console. they will look into the PC port in there spare time.

than we have tests like this and not random comments: https://www.youtube.com/watch?v=U4k_ErEg-FU

and they doesn't even try to get high framerates totally ignoring the possibility to play at lower graphic settings to get your target framerate.

Even AVX1 code runs noticeable slower on zen
https://i.postimg.cc/4xNL0x99/Flops_CPU.png

[huhn]

Both AVX and AVX2 are 256-bit. AVX is floating point, AVX2 is integer and FMA. In simplified terms.

because you really know this stuff can you elevate the use of the 128 bit version of AVX which is unofficially called AVX 128.

or is there little to nothing you can gain from doing it this way?

[nevcairiel]

AVX with 128-bit registers has small minor benefits, usually because there happen to be some instrustrictons that may do the job better, and because AVX allows more efficient use of some other instructions (ie. distinct destination registers for some instructions, instead of an implied one). The differences to SSE1/2/3/4.1 versions of the same code (which also are 128-bit) is relatively small, maybe 5-10% (for a single given function), while doubling the register size to 256 can in an ideal case of course be up to twice as fast.

[NikosD]

Intel's strong PR and marketing teams are trying really hard with tons of money behind them to persuade people that more cores are useless or at least not necessary.

It's the same old story for the dark ages of 2011 - 2017 era, when the same teams were telling us that the Core i3 2C/4T - Core i5 4C/4T - Core i7 4C/8T scheme is the best that we could ever had.

No more than 4 cores for mainstream desktop, ordered Intel.

If you wanted more, you should pay 2000$ to buy 10 cores of Broadwell-E 6950X in 2017.

On March 2017 a world revolution took place.

The revolution of Zen architecture and the first implementation of RyZen processor for mainstream desktop, using 8C/16T in a very affordable price of 350$ - 500$ (initially)

Nowadays, you can buy 32C/64T for 1800$ from AMD, when Intel still sells 18C/36T for 2000$ - even this means 80% more cores for the same price just a year later for Intel.

Cascade-AP, a one-off processor from Intel, mimicks the Zen first generation architecture for servers (EPYC) and rises the sum of cores to 48 from 28 of Skylake-EP using GLUE (as Intel called Infinity Fabric of Zen architecture) to add two Skylake 24C processors to one Cascade 48C.

Intel is trying hard to be as less humiliated as possible compared to the absolute monster of EPYC 2, a multi-die (8+1) CPU of 64C/128T adding PCIe v4.0 and Infinity Fabric 2 to the equation.

Intel guys, please relax...

If you can't avoid something, just enjoy it.

It's coming!

[Racer]

Looks like there are some serious issues with scheduler in windows. 5 instances of x265 running and 2990WX just chokes
https://p.xfastest.com/~sinchen/GIGABYTE-X399-AORUS-XTREME/GIGABYTE-X399-AORUS-XTREME-66.jpg

Source -> https://www.xfastest.com/thread-221870-1-1.html

Is there actually an updated benchmark with adjusted numa pools or did anybody check if this gives an performance improvement for the 2990WX?

Instance 1 = --numa-pools "+,-,-,-"
Instance 2 = --numa-pools "-,+,-,-"
Instance 3 = --numa-pools "-,-,+,-"
Instance 4 = --numa-pools "-,-,-,+"

[Atak_Snajpera]

Do you have access to 2990wx?