Speculation: Ryzen 3000 series

[Hans Gruber]

In CB r15 that is. Not what I would call ipc but anyways.

FMax rumours at at least 5GHz is sweet but I bet for 4.7. Perhaps we need a bet here as we had with zen ?

Isn't the 2700X built on a GloFlo fab? Zen2 is built on a TSMC 7nm fab. Since the silicon is completely different, it's possible higher ghz could result. Someone pointed out Nahalem vs. SandyBridge. Different silicon and that is why SandyBridge Oc'd so well. IvyBridge could never touch 5ghz like SandyBridge CPU's did.

It's basically a wait and see thing with Zen 2. AMD said the IPC was 15% improvement for Zen 2 over Zen+. With all the new memory controller bits. Probably faster ram and faster clocks could result in very impressive benchmarks.

 

[Asterox]

On CB R15 a 4.2GHz 8C/16T Zen 2 will be on par with a 5GHz clocked 9900K at 1/3 of the TDP...

Improvement over Zen + is 15% or so, sure that R15 is FP but there s no AVX/FMA and it s somewhat representative of legacy SSE2-4.2 apps performances, possibly that there wll be a nice improvement in pure Integer like 7 Zip, partly due to bigger caches than Zen + that hide the MC latencies.

We see that on AMD CES event, or i9 9900K 4.7ghz(no TDP limit) all core Turbo vs 8/16 Zen 2 let say 4.2ghz.

Rumors are that retail 8/16 Zen 2 will have 4.5ghz all core Turbo.

 

[EXCellR8]

Ryzen 7 3700X take my moneyyyy!!!

 

[Mockingbird]

Old Jim better hope that his "leaks" are correct, or he can find a new job.

A lot of people are sick and tired of his messiah complex.

 

[dnavas]

9900k with MCE off is 12% faster than 2700x.

My guess is that Zen2 7nm is near parity performance and maybe 1%-5% faster than Intel 14++.

Careful, though -- we don't know what kind of SMT improvements Zen2 brings over Zen1, and my guess is that we'll get a fair increase in SMT, not just because of front-end improvements, but informed by the Zen3 rumors of SMT4. There'd be no point in pursuing SMT4 unless MT is getting a fair amount of attention, and I assume some of it is happening this round.

 

[krumme]

Yes it is only "air IPC", or tell that to any "Maxon Cinema4D" user who paid 3000euro for software licence.

I am sure a lot of people will be motivated That fpu score is insane and well - the end product might perfectly eclipse 18c Intel hedt and then some.... Going to be lot of entertainment any way.

 

[krumme]

We see that on AMD CES event, or i9 9900K 4.7ghz(no TDP limit) all core Turbo vs 8/16 Zen 2 let say 4.2ghz.

Rumors are that retail 8/16 Zen 2 will have 4.5ghz all core Turbo.

Looking at the old vid what strikes me is Lisa comment about the IO die being optimised for desktop market.

Have we discussed that because that sounds pretty important since she choose to say it specifically? What could that mean?

 

[Markfw]

Looking at the old vid what strikes me is Lisa comment about the IO die being optimised for desktop market.

Have we discussed that because that sounds pretty important since she choose to say it specifically? What could that mean?

My take on that is that it will do for 12 and 16 core AM4 what 2990wx and 2970wx NEED for TR. Then the 3xxx successor comes out, I am sure the IO die will fix the problems in 24, 32 and up cores for TR.

 

[Fanatical Meat]

Smart observation from Toms Hardware

The asrock tiachi is shown in a retail box, likely indicates it will be release soon.

https://www.google.com/amp/s/amp.to...3000-motherboard-msi-asus-colorful,39370.html

 

[Ajay]

My take on that is that it will do for 12 and 16 core AM4 what 2990wx and 2970wx NEED for TR. Then the 3xxx successor comes out, I am sure the IO die will fix the problems in 24, 32 and up cores for TR.

I would think optimized for desktop means just that - AM4. I think that, as b4, Zen2 TR will be based on Rome’s platform. That would offer, obviously, Twice the bandwidth and memory capacity and many more PCIe lanes - perfect for lower priced workstations.

 

[krumme]

I would think optimized for desktop means just that - AM4. I think that, as b4, Zen2 TR will be based on Rome’s platform. That would offer, obviously, Twice the bandwidth and memory capacity and many more PCIe lanes - perfect for lower priced workstations.

Perhaps but then it's kind of weird remark unless your audience is less knowledgeable. You will not dump a 500mm2 io die on a am4 platform anyway.
A single io die is imo given as similar memory access to all dies is a huge advantage.
Not that I know what the difference should be as there is way to little space for eg cache.?

 

[itsmydamnation]

Careful, though -- we don't know what kind of SMT improvements Zen2 brings over Zen1, and my guess is that we'll get a fair increase in SMT, not just because of front-end improvements, but informed by the Zen3 rumors of SMT4. There'd be no point in pursuing SMT4 unless MT is getting a fair amount of attention, and I assume some of it is happening this round.

There is no point in SMT 4, 4 threads, 2load and 1 store port... GTFOH , only place it makes sense is for workloads that get lots of outstanding misses and then you still need your entire SOC/system built around that and thats no the code that you want to run.

 

[inf64]

As per AdoredTV's latest video, 16C @ 4.2Ghz scores ~4287pts in C15 (with optimzied DDR4 clocks I guess). This means that AMD's 8C part at CES demo that scored 2057 pts most likely ran at around 4Ghz. Compare to 2040pts that 9900K @ 4.7Ghz scores and we are seeing ~17.5% IPC difference in MT Cinebench R15 test. This is likely one of the best case scenarios for Zen2 (similar to Zen1/+) but bodes well for AMD's positioning in desktop/server space. They'll likely have parity or advantage in IPC for most workloads except AVX512 which is niche of niche . Clocks seem to be nice and TDP looks great too. The only downside is very hot X570 chipset, albeit with monstrous SSD performance if leaks are true.

 

[StinkyPinky]

The boards are leaking fast and furious now, looks like a July release is on the table. 90% certainty for me.

 

[Markfw]

The boards are leaking fast and furious now, looks like a July release is on the table. 90% certainty for me.

Or a week from today at CES

 

[StinkyPinky]

I think they would be announced formally then with all the details, but July availability. Hope I'm wrong though.

 

[Mockingbird]

probably an engineering sample of the Ryzen 5 3600 on the MSI MEG X570 ACE

https://ranker.sisoftware.co.uk/sho...d4ecdfe6deedd5f381bc8caacfaa97a781f2cff7&l=en

 

[amd6502]

There is no point in SMT 4, 4 threads, 2load and 1 store port... GTFOH , only place it makes sense is for workloads that get lots of outstanding misses and then you still need your entire SOC/system built around that and thats no the code that you want to run.

I think it's more likely to get non-SMT multithreading using heterogenous CCX, to achieve 4 thread per Zen core. (Different but not totally different from big.little approach.)

Zen 4, which we may expect to have a wider core, may get SMT4.

 

[itsmydamnation]

I think it's more likely to get non-SMT multithreading using heterogenous CCX, to achieve 4 thread per Zen core. (Different but not totally different from big.little approach.)

Zen 4, which we may expect to have a wider core, may get SMT4.

This is just meaningless , even look at how power does SMT4/8 they effective have two completely different sets of front end that are partitioned from each other. You still have to retire in order, just saying make things wider, more SMT, it doesn't buy you anything for a single thread, infact it makes single thread harder, better off with more cores. The advantage of SMT is over specing resources then you could otherwise justify for a single thread, but going past what a single thread can use makes no sense.

 

[Thunder 57]

probably an engineering sample of the Ryzen 5 3600 on the MSI MEG X570 ACE

https://ranker.sisoftware.co.uk/sho...d4ecdfe6deedd5f381bc8caacfaa97a781f2cff7&l=en

Could you or someone familiar with sisoft scores provide some context here? I think it looks good, but I'm not really sure?

 

[DownTheSky]

Could you or someone familiar with sisoft scores provide some context here? I think it looks good, but I'm not really sure?

The guys @ Reddit did the math and is 29% faster at same clock vs 2600x

https://www.reddit.com/r/Amd/comments/brkb0v

 

[TheELF]

The advantage of SMT is over specing resources then you could otherwise justify for a single thread, but going past what a single thread can use makes no sense.

So which is the golden standard thread?
If you measure the instructions used by all threads and take a middle case,the average,then you will have almost 50% of threads not having enough instructions because they use more than that and almost 50% of threads that will leave instructions on the table.
You have to put more IPC into a core than what the average thread uses to get good performance with as many different threads as possible,SMT is just a logical step in using the increased IPC if and when the main thread can't use it.

SMT4 isn't going to happen because there are not enough threads that use that small an amount of IPC and also it's way too complicated for the CPU to keep up with that many context switches on the same core.

 

[itsmydamnation]

So which is the golden standard thread?

That's all about your target workload power/complexity/performance trade off you want to make. But look across high performance CPU's intel/amd/apple/power as you see commonality while targeting very different workloads. 2 loads a cycle 2-4 int execution units , high 100's PRF, deep load store queues. Probably comes down to the inherent limitations of the Tomasulo algorithm.

 

[EXCellR8]

I'd like to stay around 100W TDP since current 1700 is around ~85W TDP I think, when overclocked. Been awhile since I've actually looked at the numbers...

To add 4 more cores to that, ramp up the speed to near 5Ghz, and only have ~25W more is a no-brainer for me, personally.

 

[krumme]

I'd like to stay around 100W TDP since current 1700 is around ~85W TDP I think, when overclocked. Been awhile since I've actually looked at the numbers...

To add 4 more cores to that, ramp up the speed to near 5Ghz, and only have ~25W more is a no-brainer for me, personally.

Yes but also completely unrealistic imo. A 1700 using 85w is max 3.7 all core. I expect the highest frequency 8c to come in at 95w tdp non oc and slot into all existing boards sans 320.