Speculation: Ryzen 4000 series/Zen 3

[amrnuke]

How else would they fit in 12 ALUs so they can stop being lazy stupid x86 engineers?

If 12 is good, surely 24 is better
Heck, why not toss SMT8 in there for good measure
If I can write it so easily, I don't see why it's such a hard thing for AMD and Intel to implement

 

[NostaSeronx]

Do to AMD's focus with gaming they should just integrate an Integer core into the RDNA's workgroup. They already use the same physical design techniques, SRAM, register files, etc. Then, allow the integer core to push fixed/float arithmetic instruction loads via translation logic units on to the local RDNA CUs. Thus, giving a logical core access to 64 ALUs!!!

 

[soresu]

Do to AMD's focus with gaming they should just integrate an Integer core into the RDNA's workgroup.

It already does Integer computation.

Apparently it doesn't do so well with Int64 numbers though compared to nVidia, this could well be a focus with CDNA at some point.

 

[Thibsie]

Yeah, and replace L1 with HBM while you're at it.
Any sane suggestion though?

Edit: directed at Nosta of course.

 

[DrMrLordX]

Yeah, and replace L1 with HBM while you're at it.

BRILLIANT

Keep up the good ideas m8

 

[jamescox]

An otherwise great, well-written, clear and cohesive post, but your memory is off here - my Q9450 had 12MB of L2! While that wasn't an early C2Q, it wasn't an extreme edition either. It kept up admirably until I replaced it in 2017, after nine years of service.

My memory isn’t great these days, but I remember some of the QX extreme edition parts showing up on VR game requirement list. Also, I believe the Core 2 quad parts were actually MCMs, so it was 2x6 MB.

 

[NostaSeronx]

It already does Integer computation.

But can it do x86 control, branches, recovery, etc? If not it's not an x86 integer core.

Yeah, and replace L1 with HBM while you're at it.
Any sane suggestion though?

L4 has always been HBMx, I haven't seen any mention of L1 changing towards slow RAM. It is more likely to go 10T SRAM than anything else.

Zen3 won't increase issue ports. Zen3 will move to a smaller but smarter front-end/back-end/load-store unit HW/algorithms. <== Stuff they couldn't put in Zen2, but could in Zen3 and did.

I believe that trend will continue going forward Zen4/Zen5/etc. Intrinsically smaller cores that take up less area have less R and less C and other effects to worry about. Hence, why SMT is more likely to be dropped.

2015-1H2018 => Zen3 had brainiac ambitions.
2H2018-1H2020 => Zen3 has speed demon ambitions.

 

[amd6502]

Do to AMD's focus with gaming they should just integrate an Integer core into the RDNA's workgroup. They already use the same physical design techniques, SRAM, register files, etc. Then, allow the integer core to push fixed/float arithmetic instruction loads via translation logic units on to the local RDNA CUs. Thus, giving a logical core access to 64 ALUs!!!

For almost all INT operations this wouldn't make sense. However, the DIVIDE operation is complex and takes supposedly 76 cycles on PD/XV and 46 cycles on Zen1 and Zen2. Furthermore, only one of the four ALU's seems capable of it, and to tie this sole ALU within the core for a far ahead speculative operation would take away resources and be an iffy bet. Queueing such an operation to to the iGPU would be a useful feature. L1 and L2 could help coordinate that.

I'm not sure what the whole lot of popular complex FPU operations are which have long latencies (and how long they are), but I'm sure there are many complex and popular trig functions that aren't done in only a few cycles. I don't know how capable the iGPU's are as far as those math functions.

It wouldn't be a bad idea that such an accessory FPU becomes universal and gets integrated into a small iGPU. The way things are going it might be renamed a ML-PU, or ML & Graphics Accessory Processor unit.

 

[Veradun]

i'm looking forward to see if that holds true with zen 3 CCDs. technically 12LP was 10% denser yet zen+ SKUs have the same area than its zen 14LPP counterparts.

maybe it did shrink, but just CCX, cache and other parts except the I/O did, but it's not really noticeable.

https://www.globalfoundries.com/sites/default/files/product-briefs/pb-12lp-11-web.pdf

Zen+ was a direct port from GF14 to GF12, something you can't do from N7/N7P to N7+. You have to redesign the chip to use N7+, and even if it wasn't the case, they are redesigning the cache so redesign it is. Meaning they're leveraging the updated node's characteristics.

 

[FriedMoose]

I haven't been following this at all, but what would be a reasonable expectation for performance? Will Intel have to give up the gaming crown? I'm stuck at nearly 2 yr old 9900K @ 5GHz and haven't used AMD since A64, will or should the new lineup be able to kick Intel down from the specific best case they have at the moment? If I can get 10 or 12 cores and indisputable better performance in the best Intel case (gaming) then I may actually consider Zen 3 whereas I earlier had resigned to not be able to get a worthwhile upgrade for 1-2 more years. Or is Zen 3 targeting something completely different?

You're likely to be disappointed with Zen 3 if you're looking for a large upgrade over your 9900k. The 9900K already beats the 3700X by 5-10% per clock and your chip is overclocked 20% higher than the all core clocks of the 3700X. Even if Ryzen 4000 brought 30% higher gaming performance it would just match or slightly beat a highly overclocked 9900k.

4GHz CPU Battle: Ryzen 3900X vs. 3700X vs. Core i9-9900K

Expanding upon all the testing we performed in our day-one 3rd-gen Ryzen coverage, today we'll be running a clock-for-clock comparison benchmark. IPC can be a good indicator...

www.techspot.com

 

[lobz]

You're likely to be disappointed with Zen 3 if you're looking for a large upgrade over your 9900k. The 9900K already beats the 3700X by 5-10% per clock and your chip is overclocked 20% higher than the all core clocks of the 3700X. Even if Ryzen 4000 brought 30% higher gaming performance it would just match or slightly beat a highly overclocked 9900k.

4GHz CPU Battle: Ryzen 3900X vs. 3700X vs. Core i9-9900K

Expanding upon all the testing we performed in our day-one 3rd-gen Ryzen coverage, today we'll be running a clock-for-clock comparison benchmark. IPC can be a good indicator...

www.techspot.com

First I was wondering how you arrived to that supposed 5-10% number, then I saw techspot in your link. Wake me up when a reliable source also comes to that conclusion.

 

[Makaveli]

First I was wondering how you arrived to that supposed 5-10% number, then I saw techspot in your link. Wake me up when a reliable source also comes to that conclusion.

Tech spot is fine for a site, however this guy making these claims on an unreleased product with nothing to back his claims....

 

[tamz_msc]

First I was wondering how you arrived to that supposed 5-10% number, then I saw techspot in your link. Wake me up when a reliable source also comes to that conclusion.

Techspot = Hardware Unboxed, so they're as favourable for AMD as it can get. And yeah gaming performance is 5-10% higher at the same clocks on Skylake compared to Zen 2, even though Zen 2 has higher IPC.

This is common knowledge.

 

[dr1337]

You're likely to be disappointed with Zen 3 if you're looking for a large upgrade over your 9900k. The 9900K already beats the 3700X by 5-10% per clock and your chip is overclocked 20% higher than the all core clocks of the 3700X. Even if Ryzen 4000 brought 30% higher gaming performance it would just match or slightly beat a highly overclocked 9900k.

4GHz CPU Battle: Ryzen 3900X vs. 3700X vs. Core i9-9900K

Expanding upon all the testing we performed in our day-one 3rd-gen Ryzen coverage, today we'll be running a clock-for-clock comparison benchmark. IPC can be a good indicator...

www.techspot.com

Intel Core i9-9900KS Review - Impressive 5 GHz

The Core i9-9900KS is Intel's new consumer flagship processor. It runs at 5 GHz boost no matter how many cores are active, which translates into 10% application performance gained over the 9900K. Gaming performance is improved too, but pricing is high, especially compared to what AMD is offering.

www.techpowerup.com

The performance difference isn't as much as you claim it to be. Per TPU the 3900x is on average 9% slower at 1080p than an overclocked 9900ks. If zen 3 brought 30% higher gaming performance it would absolutely eclipse an overclocked 9900ks.

And don't just take their word for it, every site I've checked agrees that the gaming performance difference is on average less than or equal to 10%. heres a review aggregation from reddit, Guru3D has the 9900ks only 5% faster than the 3900x, and even in techspots review the 3900x is only 10% slower than the 9900k. The 9900k is 10% faster in games clock for clock, but its also only 10% faster at nominal clockspeeds too.

Whether or not zen 3 will be able to make up this performance gap is another story. But don't get it twisted, even just a 15% improvement in games would have zen 3 beating the overclocked 9900ks.

 

[Ajay]

My memory isn’t great these days, but I remember some of the QX extreme edition parts showing up on VR game requirement list. Also, I believe the Core 2 quad parts were actually MCMs, so it was 2x6 MB.

Two Core2 CPUs on one chip, with some interfacing to improve performance. My Q6600 (G0?) ran at a solid 4GHz 24x7 (8MB L2$).

 

[FriedMoose]

Intel Core i9-9900KS Review - Impressive 5 GHz

The Core i9-9900KS is Intel's new consumer flagship processor. It runs at 5 GHz boost no matter how many cores are active, which translates into 10% application performance gained over the 9900K. Gaming performance is improved too, but pricing is high, especially compared to what AMD is offering.

www.techpowerup.com

The performance difference isn't as much as you claim it to be. Per TPU the 3900x is on average 9% slower at 1080p than an overclocked 9900ks. If zen 3 brought 30% higher gaming performance it would absolutely eclipse an overclocked 9900ks.

And don't just take their word for it, every site I've checked agrees that the gaming performance difference is on average less than or equal to 10%. heres a review aggregation from reddit, Guru3D has the 9900ks only 5% faster than the 3900x, and even in techspots review the 3900x is only 10% slower than the 9900k. The 9900k is 10% faster in games clock for clock, but its also only 10% faster at nominal clockspeeds too.

Whether or not zen 3 will be able to make up this performance gap is another story. But don't get it twisted, even just a 15% improvement in games would have zen 3 beating the overclocked 9900ks.

The performance difference is only 10% because of GPU bottlenecks. The 9900K is very clearly 5-10% faster than an equivalently clocked 3700X.

 

[FriedMoose]

Tech spot is fine for a site, however this guy making these claims on an unreleased product with nothing to back his claims....

So Intel is 5-10% ahead per clock and 20% ahead on all core clocks. Seems like pretty simple math even with clocks not scaling with performance 1:1. The fact that people point to GPU bound benchmarks to "disprove" this gap is pretty hilarious.

 

[Markfw]

So Intel is 5-10% ahead per clock and 20% ahead on all core clocks. Seems like pretty simple math even with clocks not scaling with performance 1:1. The fact that people point to GPU bound benchmarks to "disprove" this gap is pretty hilarious.

I believe that the difference is 5-10% even with both at their highest clocks. I would post links to benchmarks, but they are all over.

Even in what you linked, its 1,4,7,13% depending on resolution. 13% was at 720p I think.

 

[DrMrLordX]

The performance difference is only 10% because of GPU bottlenecks. The 9900K is very clearly 5-10% faster than an equivalently clocked 3700X.

I hope you realize that if you replace a 3900X and 9900K with a 10900K (or Rocket Lake!) and a "4900X" or whatever AMD will call it, the GPU bottlenecking is just going to get more severe. CPU horsepower is increasing faster than GPU grunt right now. NV and AMD may change that soon, or they may not. But so long as the 2080Ti remains the gold standard, that won't change.

 

[inf64]

Computerbase did a test with frametimes/fps @ 720p, 9900KS is 2% faster than 3900XT when frametimes are in question and 4% faster than 3900XT in pure fps. They used 2080Ti FE for a GPU, so top of the line graphics card for a super low resolution testing.
https://www.computerbase.de/2020-07...-test/3/#abschnitt_fps_und_frametimes_in_720p

Zen3 will have zero problems outpacing 10900K and even Rocketlake, it just needs to get similar all core/ Turbo clocks as 3900XT and it's AMD's gaming crown soon.

 

[Rigg]

Techspot = Hardware Unboxed, so they're as favourable for AMD as it can get. And yeah gaming performance is 5-10% higher at the same clocks on Skylake compared to Zen 2, even though Zen 2 has higher IPC.

This is common knowledge.

Is it common knowledge? That strikes me as an incredibly difficult thing to quantify using a sweeping generalization. There are massive variables that effect the outcome of the data. IPC is extremely difficult to gauge for gaming because IPC is unique to every individual workload. Not to mention memory latency and GPU bottlenecks that influence results.

The 10400 and and 3600 have equal core/thread counts and roughly equal stock clock speeds. In this 25 game benchmark the 3600 wins the overall FPS averages.

Why does the 3600 win? Because the 10400 is running 2666 memory on a b460 board vs the 3600 running at 3200 on x570.

Unless you run testing across a huge swath of games, at multiple fixed clock speeds, with multiple memory clocks, with XMP vs optimized timings, at multiple resolutions, with multiple quality settings, the generalizations are mostly meaningless. All of these factors will vastly effect how performance scales in video games. Even a handful of games would take a ton of benchmarking to account for these variables. Without these variables factored in I wouldn't trust any broad conclusions.

I think it's a fools errand anyway. Who cares? Intel is faster at gaming in CPU bound scenarios. No need to muddy the waters with a much more difficult to quantify gaming IPC comparison.

 

[tamz_msc]

Is it common knowledge? That strikes me as an incredibly difficult thing to quantify using a sweeping generalization. There are massive variables that effect the outcome of the data. IPC is extremely difficult to gauge for gaming because IPC is unique to every individual workload. Not to mention memory latency and GPU bottlenecks that influence results.

The 10400 and and 3600 have equal core/thread counts and roughly equal stock clock speeds. In this 25 game benchmark the 3600 wins the overall FPS averages.

Why does the 3600 win? Because the 10400 is running 2666 memory on a b460 board vs the 3600 running at 3200 on x570.

Unless you run testing across a huge swath of games, at multiple fixed clock speeds, with multiple memory clocks, with XMP vs optimized timings, at multiple resolutions, with multiple quality settings, the generalizations are mostly meaningless. All of these factors will vastly effect how performance scales in video games. Even a handful of games would take a ton of benchmarking to account for these variables. Without these variables factored in I wouldn't trust any broad conclusions.

I think it's a fools errand anyway. Who cares? Intel is faster at gaming in CPU bound scenarios. No need to muddy the waters with a much more difficult to quantify gaming IPC comparison.

Gaming IPC is a stupid metric anyway, as we have seen examples in the past where drivers behave differently depending on the CPU architecture.

All other things equal Intel is 5-10% ahead in gaming at iso-clocks in GPU-bound scenarios as tested by Hardware Unboxed(they test at ultra settings). The moment you test CPU-bound settings like Gamers Nexus does, the gap increases to 20% in some games and even 30% in rare examples.

 

[Rigg]

Gaming IPC is a stupid metric anyway, as we have seen examples in the past where drivers behave differently depending on the CPU architecture.

I agree.

All other things equal Intel is 5-10% ahead in gaming at iso-clocks in GPU-bound scenarios as tested by Hardware Unboxed(they test at ultra settings). The moment you test CPU-bound settings like Gamers Nexus does, the gap increases to 20% in some games and even 30% in rare examples.

I just checked the GN 10400 review. I didn't see anywhere near 20-30 % gains when comparing the 10400 vs 3600. If there is better data to look at by all means link to it.

1080p benchmarks were .1 - 9.6% in favor of the 10400 all things being "equal". 3 Kingdoms campaign is 8.6%, 3 Kingdoms Battle is 3.1%, Hitman 2 is 6.2%, F1 2019 is 8.6%, Tomb Raider is 9.6%, GTA V is .1%, Assasins Creed Origins is 5.2%.

If you look at Ryzen 3000 with increased CPU frequency in the latest XT reviews from GN, it seems pretty clear that it doesn't scale FPS performance because its hitting a memory latency bottleneck and not a core clock bottleneck. We're not getting an iso clock comparison. We're getting a combination of isolated clock and memory latency comparison with varying degrees of GPU limitation thrown in for good measure.

The Intel CPU's scale with CPU frequency using the 3200 cl14 test RAM while the Ryzen doesn't. This doesn't mean that the Intel isn't a faster gaming CPU at the same frequency (it most likely is) it means we can't know that for sure without knowing how memory speed and timings scale performance with frequency.

 

[randomhero]

I am really tired of this gaming discussion.
Difference in this workload scenario is so negligible that people are arguing over height of graph bars.
I'll tell you, if you can indentify system in double blind test in HU 30 or so games benchmark testbed when using appropriate resolutions and game settings to hardware in systems, I'll lick yours pet behind, tape it and upload it to YouTube.

 

[JoeRambo]

I am really tired of this gaming discussion.
Difference in this workload scenario is so negligible that people are arguing over height of graph bars.
I'll tell you, if you can indentify system in double blind test in HU 30 or so games benchmark testbed when using appropriate resolutions and game settings to hardware in systems, I'll lick yours pet behind, tape it and upload it to YouTube.

If one agrees, that at certain resolution X, the difference is say 15-20%, and tomorrow with more powerful GPUs the difference will move to resolution X+1

then at certain point it will beome clear that said 15-20% is sometimes more than a step in GPU vendor cards ( like 2070 super => 2080 ).

Just food for thought in "FPS" department.

And FPS is not end to all, there are turn times, simulation speeds, emulation speeds and so on. In the end faster CPU is faster.

P.S. And i also own 3950x, not for gaming tho.