AMD Zen 2 (Ryzen 3000) - *** NO COMPETITOR HINTING ***

[DylRicho]

Hi all,

I have been following this thread ever since I was notified that one of my resources had been shared here. I've since read through 300 pages, before making the fatal error to take a couple of weeks break. Now I'm 300 pages behind, but I've decided to take the time to join in as I miss talking about microelectronics with fellow forum members (the SemiAccurate forum was removed without warning, apparently).

Anyway, yes. Nice to meet you all.

Presumably sunny cove has AVX512 - I could see that sort of increase being likely in a benchmark that uses those instructions heavily (I've no idea if this benchmark does) even if normal IPC increase is more modest.

I have spoken directly to the developers regarding CPU-Z, so I can fill in the gaps for you:

• The benchmark uses superscalar SSE2 instructions, but nothing higher than that. This is why Core2 Duo/Quad processors are able to keep up with AMD APUs and newer Celerons.
• The benchmark used by versions 1.78 and prior were small enough to fit entirely into the Ryzen L2 cache, "enhancing" the scores for those models.
• The benchmark used by versions 1.80 and newer fixed that.
• Version 1.87 introduced a new AVX2 benchmark.

Going by the scores from the image, they are basing their numbers on the SSE2 benchmark.

It really isn't a great benchmark in the grand scheme of things. You're specifically testing either SSE2 or AVX2 instruction throughput, and nothing else. The latter of which will show Zen 2 significantly stronger than the previous generations due to the modified FPU capabilities.

 

[Bluntwrapped]

Let’s all be realistic and sensible and buy the 3600.

 

[AthlonXP1800]

3229 on Cinebench R20 on Ryzen 5 3600 65W CPU that on 7nm?

Wow that awful, it is just has same performance as i7 8700 65W CPU on 14nm.

Not very impressive and slight slower than my 8700K.

I would be impressed if it was 35W CPU.

Let see games benchmarks but I think it will have poor performance.

 

[smilingcrow]

The benefits of pcie 4.0 really is just for storage, gpu's can barely even max out pcie 2.0.

Even for storage we are already deep into diminishing returns with PCIe 3.0 x4 for most consumer workloads so I'm not seeing PCIe 4.0 as being that relevant beyond being a tick box for those that like such things.

 

[DylRicho]

3229 on Cinebench R20 on Ryzen 5 3600 65W CPU that on 7nm?

Wow that awful, it is just has same performance as i7 8700 65W CPU on 14nm.

Not very impressive and slight slower than my 8700K.

I would be impressed if it was 35W CPU.

Let see games benchmarks but I think it will have poor performance.

There is something off about those results. The 2600 scores exactly the same at the same frequency.

 

[Sayso]

I'm 4k BOTW with my 8086k so ideally I'm kind of hoping for an upgrade on that or at least it matching that. Do you think it can happen?

only if going for the 12c part and then depending on usage you its still not worth it.
These CPUs get AMD back in the game (well Ryzen did) the only way they top Intel is with core count and pricing.

 

[ninezerofive]

 

[chrcoluk]

Even for storage we are already deep into diminishing returns with PCIe 3.0 x4 for most consumer workloads so I'm not seeing PCIe 4.0 as being that relevant beyond being a tick box for those that like such things.

I just watched gamers nexus run down of x570.

I dont see any gpus needing anything like 16 4.0 lanes anytime soon, and chipsets have always been starved of bandwidth and you guys also know I am not a fan of nvme having 4 direct lanes from the cpu to a m.2 slot.

So I would have done this.

8 lanes to gpu pcie slot (more than enough)
16 lanes to chipset, a massive improvement over previous gen and intel's mainstream chipset.

Imagine how awesome that would be.

Then use bridge chips that can control where the chipset bandwidth is allocated, whether thats m.2, pcie slots, usb, sata etc. So e.g. you could still have 4 lanes directed to m.2 slots, but they would be going via the chipset instead of direct from cpu.

I feel tho they always limit to 4 lanes so they can push higher priced enthusiast/server sku's as my idea would hurt the other markets.

Also I would rather have a beefy passive heatsink and not a fan cooling my chipset, as those fans probably will after some years of usage (maybe even months) probably start rattling or making other weird noises.

Also it seems board makers have a choice of either 2 m.2 + 4 sata from chipset or 1 m.2 + 8 sata from chipset, I deffo prefer the latter but I expect no boards to have that configuration as m.2 is the hype. To me just 4 sata ports on a high end board is crazy low. If we had my idea of 16 lanes to chipset tho i expect you could do something like 3 m.2 "and" 8 x sata.

 

[Zeed]

There is something off about those results. The 2600 scores exactly the same at the same frequency.

What I get on my system

 

[awaybreaktoday]

Zen2 big problem:

As of the current rumored pricing, a 9900k and Z390 Formula is “~$350” cheaper than a 3900X and an X570 Formula, also the gap is big when it comes to the Aorus xtreme motherboards also....

 

[DylRicho]

What I get on my system
https://i.ibb.co/JjWjvPy/cb20.jpg

Exactly. Ryzen shines well in Cinebench (!= favoring Ryzen), so you would expect Cinebench to be one of those utilities that shows the largest improvement for Zen 2, especially given that R20 is an AVX1/AVX2/AVX3.2 benchmark; the area that Zen 2 improves the most.

Even a 10% IPC improvement at the same frequency would net an 8C/16T SKU a score of 4,814 cb. Doing some cheaty napkin math and multiplying that by 75% then results in 3,610 cb for a 6C/12T at 4.25 GHz.

So yes. That figure is well off what is expected.

 

[Orange Nexus]

Zen2 big problem:

As of the current rumored pricing, a 9900k and Z390 Formula is “~$350” cheaper than a 3900X and an X570 Formula, also the gap is big when it comes to the Aorus xtreme motherboards also....

Get the 8 core and a cheaper X570 like the Asus TUF (180£). That low-end X570 is overbuilt for the price. It should match the i9 9900K easy.

Problem solved.

That expensive X570 is meant for 16 cores - IMO.

 

[Grim5]

Get the 8 core and a cheaper X570 like the Asus TUF (180£). That low-end X570 is overbuilt for the price. It should match the i9 9900K easy.

Problem solved.

That expensive X570 is meant for 16 cores - IMO.

The question then becomes do you lose any features by moving down to the Tuf from the Formula?

I don’t know all the differences between these boards but what I know about ASUS is how cheap they are. On their entry boards all you get is a 1g lan connection and then they only put the 5g on the high end boards. Meanwhile other brands use 10g lan in mid and high end.

 

[Grim5]

 

[Troezar]

I'm not really up on streaming but there seems to be a lot of people disputing their findings in the comments.

 

[Orange Nexus]

The question then becomes do you lose any features by moving down to the Tuf from the Formula?

I don’t know all the differences between these boards but what I know about ASUS is how cheap they are. On their entry boards all you get is a 1g lan connection and then they only put the 5g on the high end boards. Meanwhile other brands use 10g lan in mid and high end.

Not sure. What i do know both X570 boards will have PCIe4. Something the intel board does not have. Some will say, well that feature is not really needed blah blah. Fact is its there. I would not hesitate to even pair the the TUF with a 12 core goodness. Its just overbuilt.

 

[Grim5]

I wasn't meaning PCIE4 - they all have that. All of them are pcie4, they all have a x570 chipset, they all support DDR4 ram and they all support Ryzen 3000. It's the smaller things that divide them.

I had a quick run down - the Formula obviously has better VRM, it's also a board that needs a custom loop to cool the VRM.

As I suspected, the Tuf only does 1g LAN, the formula 5g. The formula has in built wifi, the Tuf it seems does not. The Formula has nearly triple the amount of USB ports and the Formula can do 2 pcie4 GPU's in SLI, while the Tuf can only do 1 gpu with pcie4. While I dont care about the next one, but the Formula has a higher model Audio controller and various overclocking features the Tuf doesn't it (not sure that matters).

Some of these things may be important to people - for instance the Tuf would never cross your mind if you are going to SLI

 

[DragonQ]

Hi all,

I have been following this thread ever since I was notified that one of my resources had been shared here. I've since read through 300 pages, before making the fatal error to take a couple of weeks break. Now I'm 300 pages behind, but I've decided to take the time to join in as I miss talking about microelectronics with fellow forum members (the SemiAccurate forum was removed without warning, apparently).

Anyway, yes. Nice to meet you all.



I have spoken directly to the developers regarding CPU-Z, so I can fill in the gaps for you:

• The benchmark uses superscalar SSE2 instructions, but nothing higher than that. This is why Core2 Duo/Quad processors are able to keep up with AMD APUs and newer Celerons.
• The benchmark used by versions 1.78 and prior were small enough to fit entirely into the Ryzen L2 cache, "enhancing" the scores for those models.
• The benchmark used by versions 1.80 and newer fixed that.
• Version 1.87 introduced a new AVX2 benchmark.

Going by the scores from the image, they are basing their numbers on the SSE2 benchmark.

It really isn't a great benchmark in the grand scheme of things. You're specifically testing either SSE2 or AVX2 instruction throughput, and nothing else. The latter of which will show Zen 2 significantly stronger than the previous generations due to the modified FPU capabilities.

Any benchmark that doesn't aim to replicate a real-world use case is purely synthetic and largely a waste of time. Things like Blender, Handbrake, Time Spy, etc. are all actually useful.

 

[DylRicho]

Any benchmark that doesn't aim to replicate a real-world use case is purely synthetic and largely a waste of time. Things like Blender, Handbrake, Time Spy, etc. are all actually useful.

Agreed.

 

[kitfit1]

Any benchmark that doesn't aim to replicate a real-world use case is purely synthetic and largely a waste of time. Things like Blender, Handbrake, Time Spy, etc. are all actually useful.

Along with Real Bench of course.