Speculation: Ryzen 3000 series

[DrMrLordX]

20% seems to be too much, but I think the speculation based on some of the results we've seen is that AMD did hit double digit improvements in at least some areas. I

15% seems to be the most commonly-cited number.

 

[scannall]

15% seems to be the most commonly-cited number.

At the Epyc demo Lisa Su was saying %13 which is decent. That along with improved memory controllers and the next gen of IF should be really nice.

 

[Topweasel]

Just a remark: consumers arguing for higher prices, that is quite a remarkable thing. It raises eyebrows.

It's almost nerve wracking. I get them wanting AMD to succeed and some basic understanding the profit per sale is an important aspect to AMD staying profitable. The answer is in the middle somewhere. We aren't getting 16c CPU's for $200. But if you add sufficient value to higher end more expensive parts you can still raise ASP and Margin by increasing adoption at those higher price breaks. A 16c 4.3 GHz CPU for the upper range of consumer CPU pricing at $500 is going to convince a lot of guys originally looking at ~$300 8c CPU's to spend a little extra for the dramatic increase in performance. A 32c TR for $1k-1.2K again value shoots up even if the price is high barrier. EPYC if AMD prices it to be competitive with Intel you are looking at companies paying 10-15k per AMD CPU. Manufacturers would rather wait out Cascadelake and see what Intel's pricing is. But if they keep it around 6k all of a sudden it becomes a hard CPU for data-centers to avoid.

 

[PotatoWithEarsOnSide]

I'm not sure where this "Intel IPC is 10% ahead" comes from. They're barely 3% ahead in ST, but their benchmark leads come from higher clocks. AMD are expected to increase IPC and clocks, with IPC going ahead of Intel, and clocks almost reaching parity.

 

[Joe Braga]

Guys How can it work on Ryzen 3000 e Threadripper 3000:
Ryzen: 4CCXs or 2 chiplets w/ 8 cores each
Threadripper: 8CCXs or 6/8chiplets w/ 8 cores each

 

[JasonLD]

Regarding the possible 48 or 64 core Threadripper, wouldnt that be bottlenecked by quad-channel memory?

 

[piesquared]

Regarding the possible 48 or 64 core Threadripper, wouldnt that be bottlenecked by quad-channel memory?

It's pretty hard to know anything given this is a huge shift in how CPUs are built, it could just as easily not be. You would have to know at least something about the Zenx IO and how the memory system is even structured with this new design to make any kind of valid guess imo.

 

[H T C]

Regarding the possible 48 or 64 core Threadripper, wouldnt that be bottlenecked by quad-channel memory?

Apparently, what was thought to be memory bottlenecks because of quad channel memory on TR 24c / 32c were in fact because of Windows not behaving as it should:

Will this also happen if TR moves up to 48c / 64c? Maybe, maybe not. I'm expecting no memmory bottlenecks with Zen 2's TR version of 24c / 32c but increasing the core count may in fact introduce them, assuming ofc there will be 48c / 64c TR CPUs which isn't confirmed, AFAIK.

Supposedly, Zen 2 will work the exact same way, as far as software is concerned:

"The optimization that we worked with the industry as we first rolled out Ryzen was our core complex," Papermaster said, "We very successfully worked across the OS, with Windows and Linux, so there is a recognition of AMD’s core complex, and so you can really have your workloads leverage that organization. As we go forward into this next-generation with Zen 2-based products, we actually just make it easier because as you have cores going into a common I/O die, it is the same core complex approach that we had before, and you actually just have a very centralized path. In our server implementation all the way through the Ryzen implementation we showed today, it adds no complication whatsoever for the software providers. All the work we did with first-gen Ryzen will carry right over. All those optimizations carry right over."

The above quote was taken from here.

 

[Timorous]

Apparently, what was thought to be memory bottlenecks because of quad channel memory on TR 24c / 32c were in fact because of Windows not behaving as it should:

This was known as soon as Phoronix published their linux review, they even did a windows / linux comparison and linux blew windows out of the water.

 

[DrMrLordX]

Supposedly, Zen 2 will work the exact same way, as far as software is concerned:

Zen2 should be much friendlier to software developers that don't want to deal with NUMA-on-chip.

 

[exquisitechar]

Regarding the possible 48 or 64 core Threadripper, wouldnt that be bottlenecked by quad-channel memory?

Quad-channel memory wouldn't be enough for a 64c TR, for sure.

 

[Atari2600]

Quad-channel memory wouldn't be enough for a 64c TR, for sure.

In some scenarios it certainly would be the bottleneck and cripple performance.

But in many other workloads it may not be such a big deal. You'd really have to see benchmarks of the actual applications you intend to use.

 

[mattiasnyc]

My bet's on Threadripper topping out at 32 cores, but with a big performance uplift over the current version.

 

[Elfear]

I don't think it's overly optimistic to think that they hit performance parity with Intel, at least in a few areas. Maybe they've even managed to get more out of their SMT which was surprisingly better than what Intel had.

IPC parity is my bet as well. That would still be a pretty significant uptick in performance.

 

[PotatoWithEarsOnSide]

I'm not sure that is what he was saying. Performance parity =/= similar IPC; AMD will likely still lag in clocks, though an IPC lead would result in overall performance parity.

 

[moinmoin]

Quad-channel memory wouldn't be enough for a 64c TR, for sure.

Depends on the use case. E.g. on Intel Skylake-SP multithreaded y-chrucher AVX2 and AVX512 saturates the existing memory bandwidth. So does integrated graphics in Raven Ridge with 11CU when overclocking.

Zen 2 comes with double the L3$ size as victim cache which should alleviate part of such bottleneck cases.

But at some point SIMD (be it AVX something or bigger iGPUs. "Single Instruction, Multiple Data", the memory bandwidth starving is already in the name!) just don't make sense on a traditional CPU anymore compared to dedicated accelerator cards with dedicated high speed memory.

Other than that a 64c TR will still have valid use cases even with only quad channels.

 

[Elfear]

I'm not sure that is what he was saying. Performance parity =/= similar IPC; AMD will likely still lag in clocks, though an IPC lead would result in overall performance parity.

True enough, I should have been more clear in my answer. What I predict is that there will be IPC parity and a slight clock speed deficit resulting in essentially the same performance. Not sure I can last 4 more months to find out though.

 

[beginner99]

Other than that a 64c TR will still have valid use cases even with only quad channels.

Agree. In my field of work, such a chip would be very useful. heavy/long calculations on small amounts of data.

 

[HurleyBird]

I think there's a point where the IPC gap between dissimilar architectures is "close enough" that the noise of benchmark selection tends to overshadow any actual difference and makes comparison very difficult. A prime example would be Athlon versus Pentium 3 back in the day. This wasn't the case with, say, Bulldozer versus anything post-Conroe because the difference in IPC far exceeded the noise. But my feeling is that with Zen vs. Skylake we're pretty close, but aren't quite close enough for that to be the case with single-threaded IPC, while arguably we are with multi-threaded IPC.

The difference between Zen 2 and Skylake for at least single threaded IPC is looking like it will be at worst close enough so that we're in that area where an IPC comparison becomes difficult, and at best might be just good enough to have a visible, slight edge. Multi-threaded IPC is another matter altogether of course.

 

[TheGiant]

I hope it won't be the case, but today reviews of radeon VII 7nm don't show the 7nm in the better light
lets see if ryzen 3 (which is not a die shrink as the GPU) will perform better than just sometimes reach the performance of its competitor (2080) and be such power hungry and noisy
are there any leaks since AMD CES demo ?

 

[Markfw]

I hope it won't be the case, but today reviews of radeon VII 7nm don't show the 7nm in the better light
lets see if ryzen 3 (which is not a die shrink as the GPU) will perform better than just sometimes reach the performance of its competitor (2080) and be such power hungry and noisy
are there any leaks since AMD CES demo ?

Well, I don't know, but it has a lot more memory than the Vega 64, and more transistors in the GPU in total and its runs the same temps and power.Design is not much different like the 3000 series will be.

 

[Kocicak]

I hope it won't be the case, but today reviews of radeon VII 7nm don't show the 7nm in the better light

You seem to be mixing apples with bananas. BTW you could certainly build a space heater using a 7nm process, if you wanted to.

 

[TheGiant]

You seem to be mixing apples with bananas. BTW you could certainly build a space heater using a 7nm process, if you wanted to.

care to share your insides of 7nm GPU situation and possible connection/correlation to CPU performance?

 

[Kocicak]

What making old GPU architecture on a new production node and then overclocking it does have in common with builing new CPU architecture on the same production node???? 7 nm technology is just a "building brick" and you can make anything with any parameters you wish using it.

 

[Atari2600]

I hope it won't be the case, but today reviews of radeon VII 7nm don't show the 7nm in the better light

Shrink ratio of just under 0.7, clocks up between 12%-13% and power down approximately 10% (ranging between 20% and 5% depending on workload - c.f. Ananadtech with Techreport - could be driver induced variance?).

So without any architecture changes, and assuming no changes in bottlenecks/ceiling sources, AMD should be able to get Zen boost clocks up from 4.0 GHz (1800X) to 4.5 GHz while dropping power from from nominal TDP of 95W to 85W. Sounds alright to me - if not quite the paradigm shift that some were maybe hoping for.