Intel Core i7-11700K "Rocket Lake" CPU Outperforms AMD Ryzen 9 5950X in Single-Core Tests

[AnarchoPrimitiv]

Here comes the Ryzen 5000 XT series on 7nm EUV (improved node)...willing to bet on it

 

[dragontamer5788]

I have no issue with the possibility of increasing the decoder width or even adding more execution ports. But I question how likely it is, if Cypress Cove is basically a backport of Sunny Cove, since these kinds of changes usually require a total overhaul of the cache, register files and everything on the front-end.

Wouldn't it be more likely to backport something from the execution side from e.g. Sapphire Rapids, or to simply add more execution units on existing execution ports? (like one extra MUL unit?)

I think ARM has an advantage on decoder width. That's the only weak point of the x86 ISA I can think of.

x86 requires a byte-by-byte decoder, because you have 2-byte, 3-byte, 4-byte... 15-byte instructions (some of which are macro-op fused and/or micro-op split). ARM standardized upon 4-byte instructions with an occasional 8-byte macro-op fused. That means if you want to perform 4-wide decoding (and assume an average of 4-bytes per instruction), you need 64-parallel decoders: one for every byte (byte0, byte1, byte2) of the cache line.

ARM on the other hand is always 4-bytes or 8-bytes at a time (in the case of macro-op fused operations). Which means for a 64-byte decoder, ARM only need 16-parallel decoders: knowing there's no 2-byte or 3-byte instructions that could be "in-between". Just hypothetically speaking of course, I dunno really how these things are organized.

Anyway: Apple M1 shot a broadside at the x86 camp with their 8-wide decoder. I do think its relevant to bring up. However, ARM Neoverse is still only 4-wide decoding. It hasn't really been proven yet that an ultra-wide decoder (like Apple's M1) is really the best path forward.

 

[TumbleGeorge]

LoL I see that X86 is ok with better decoder. But isn't possible to make better decoder because has depencies how work ISA with information. This is same as ISA X86 is not ok itself.

 

[Fierce Guppy]

tbh the “next big thing” should be Alder Lake, not Rocket Lake.


Let’s be honest: it almost was a paper launch with skyrocket prices for Zen 3.
I don’t know about new zeland, but here in Europe it is very difficult to find one at a decent price.

Well, what's a decent price? This OK? Considering that you guys suffer more with higher taxes on stuff, it's probably a decent price.

 

[FinneousPJ]

Well... they hit 5Ghz on 14nm++++++ which we know they can do quite easily. Let's see if Alder Lake hits 5Ghz at 10nm.

Oh, silly me for thinking Intel was finally on a new process lmao

 

[Dave65]

Got to love that headline..

 

[Max(IT)]

Well, what's a decent price? This OK? Considering that you guys suffer more with higher taxes on stuff, it's probably a decent price.

It is one of the best prices in Europe... if you consider UK as part of the Europe (which is not...).
By the way that website is one of the best, but for some items they don’t ship outside UK

 

[Fierce Guppy]

It is one of the best prices in Europe... if you consider UK as part of the Europe (which is not...).
By the way that website is one of the best, but for some items they don’t ship outside UK

The UK is not part of the European Union. It will always be a part of Europe by dint of its geographical location. In any case, the offer is not availble to you if you're not a UK resident. But if you knew someone in the UK that might do you a favour....

 

[Max(IT)]

The UK is not part of the European Union. It will always be a part of Europe by dint of its geographical location. In any case, the offer is not availble to you if you're not a UK resident. But if you knew someone in the UK that might do you a favour....

yep... I was speaking about not being in Europe in a commercial way, not geographically

 

[Fluffmeister]

Here comes the Ryzen 5000 XT series on 7nm EUV (improved node)...willing to bet on it

Oh God, lets hope not, 2% more performance for way more money and they don't include a cooler as a bonus!

 

[SL2]

tbh the “next big thing” should be Alder Lake, not Rocket Lake.

You missed the "since 2015" part. Intel have been using the same Skylake design all these years, and Rocket Lake is the departure from that. Alder Lake may be better, but it's the next big thing after RL.

 

[TheoneandonlyMrK]

yep... I was speaking about not being in Europe in a commercial way, not geographically

Yep well 6800s most geforce's n in fact most GPU are scarce in the UK a rx580 sells for 220£ new still.
Plenty of intel CPU but few others in stock and favourite ones like 5900x aren't about anymore.
Oh to be rich.

 

[SL2]

Nevermind.

 

[dragontamer5788]

Yep well 6800s most geforce's n in fact most GPU are scarce in the UK a rx580 sells for 220£ new still.
Plenty of intel CPU but few others in stock and favourite ones like 5900x aren't about anymore.
Oh to be rich.

AMD is supply constrained. They were only planning to reach 10% or 15% marketshare around now, and didn't expect that their chips would be such a hit. If AMD produced too many chips, they could risk bankruptcy as well as damage to their brand.

The Xilinx purchase probably helps: since it will give them a stable source of revenue, allowing them to play a bit more aggressive in the months and years ahead.

 

[Vya Domus]

I think ARM has an advantage on decoder width. That's the only weak point of the x86 ISA I can think of.

x86 requires a byte-by-byte decoder, because you have 2-byte, 3-byte, 4-byte... 15-byte instructions (some of which are macro-op fused and/or micro-op split). ARM standardized upon 4-byte instructions with an occasional 8-byte macro-op fused. That means if you want to perform 4-wide decoding (and assume an average of 4-bytes per instruction), you need 64-parallel decoders: one for every byte (byte0, byte1, byte2) of the cache line.

ARM on the other hand is always 4-bytes or 8-bytes at a time (in the case of macro-op fused operations). Which means for a 64-byte decoder, ARM only need 16-parallel decoders: knowing there's no 2-byte or 3-byte instructions that could be "in-between". Just hypothetically speaking of course, I dunno really how these things are organized.

Anyway: Apple M1 shot a broadside at the x86 camp with their 8-wide decoder. I do think its relevant to bring up. However, ARM Neoverse is still only 4-wide decoding. It hasn't really been proven yet that an ultra-wide decoder (like Apple's M1) is really the best path forward.

Regardless, I do not think x86 designs are limited in any way by current decode width. Actually, they clearly can't be since the back end on these CPUs keeps getting wider and wider and there don't seem to be any problem feeding all of those execution ports. And after all x86 is still more compact when it comes down how much decode is necessary to get the same amount of work done.

Apple's obsession with a ultra wide front end (and ultra wide everything really) seems to be rather wasteful, there is no obvious reason why that's actually required at the moment, I bet you everything that with half the decode stage the performance regression would be marginal.

I mean most of the performance that's worth extracting through ILP sits in loops and those don't put pressure on the decode stage because you'll keep hitting the instruction cache anyway which is colossal on something like M1. Actually the more I think about it the more absurd Apple's design choices appear to me.

 

[dragontamer5788]

I mean most of the performance that's worth extracting through ILP sits in loops and those don't put pressure on the decode stage because you'll keep hitting the instruction cache anyway which is colossal on something like M1. Actually the more I think about it the more absurd Apple's design choices appear to me.

Well, the uOp cache could be used for other registers, or L1 cache instead. So I'm not sure of the existence of the uOp cache is favorable to your argument.

 

[Vya Domus]

Well, the uOp cache could be used for other registers, or L1 cache instead. So I'm not sure of the existence of the uOp cache is favorable to your argument.

It's really, really, unlikely for a real world single program to have it's instructions pushed out of the micro-op cache, or if there is some kind of weird uop count that makes the whole mechanism inefficient. I am not saying it doesn't happen but the micro-op cache is typically the least of your worries.

 

[dragontamer5788]

It's really, really, unlikely for a real world single program to have it's instructions pushed out of the micro-op cache. I am not saying it doesn't happen but the micro-op cache is typically the least of your worries.

Oh, what I'm saying is that the M1 has an advantage, because its decoder is 8-wide, while Intel / AMD has a disadvantage, because their uOp caches are only 6-wide. Instead of having a uOp cache, the M1 spent its transistors on more L1 cache and a larger register-file.

The M1 can fit 192kB into its L1 i-cache, which performs a bit faster than the Intel/AMD uOp cache thanks to 8-way decoding. Intel / AMD only have 48kB i-L1 (for Rocket Lake) or 32kB i-L1 (AMD Zen 3), and smaller than that for its uOp cache.

-------

EDIT: I should say "Seems to have an advantage". Its not very clear if Apple's big decoder is a good strategy yet IMO. But its interesting, and worth keeping an eye on. Especially because it seems like an area that may be harder to implement into x86.

 

[Vya Domus]

Oh, what I'm saying is that the M1 has an advantage, because its decoder is 8-wide, while Intel / AMD has a disadvantage, because their uOp caches are only 6-wide. Instead of having a uOp cache, the M1 spent its transistors on more L1 cache and a larger register-file.

I really doubt that a micro-op cache is that much of an expensive mechanism to add definitely not comparable to the size and power of a larger L1 I-cache, they probably didn't add one because it just wasn't required, you're really gonna tell me that Apple is that conscious about their transistor budget ? And after all a lot of CPUs out there don't have one either, it's a pretty recent addition.

What I am also saying it that I haven't actually seen any evidence that such a wide decoder is actually worth it. Wide decode means a lot of delay in the circuitry which means poor clock speed scaling.

 

[Steevo]

Oh, what I'm saying is that the M1 has an advantage, because its decoder is 8-wide, while Intel / AMD has a disadvantage, because their uOp caches are only 6-wide. Instead of having a uOp cache, the M1 spent its transistors on more L1 cache and a larger register-file.

The M1 can fit 192kB into its L1 i-cache, which performs a bit faster than the Intel/AMD uOp cache thanks to 8-way decoding. Intel / AMD only have 48kB i-L1 (for Rocket Lake) or 32kB i-L1 (AMD Zen 3), and smaller than that for its uOp cache.

-------

EDIT: I should say "Seems to have an advantage". Its not very clear if Apple's big decoder is a good strategy yet IMO. But its interesting, and worth keeping an eye on. Especially because it seems like an area that may be harder to implement into x86.

And the fact that ARM is essentially hardware based accelerators taped together with good power gating on the most advanced nodes. X86-64 has the advantage of is you want to do X in the future software and brute force will do it, ARM designs..... You need to buy a whole new device.

What happens when 8K or whatever is next becomes a thing? Apple products become obsolete and cheap, which is why a used Ipad Pros get sold for dirt cheap. 2-3 year old one is now $270 VS the initial price of $1k. Almost as bad as other ASIC hardware like GPU's, but 1K of X86 hardware will retain its value longer, and you can upgrade RAM and GPU's, increase storage and it just works.

 

[dragontamer5788]

I really doubt that a micro-op cache is that much of an expensive mechanism to add definitely not comparable to the size and power of a larger L1 I-cache, they probably didn't add one because it just wasn't required, you're really gonna tell me that Apple is that conscious about their transistor budget ? And after all a lot of CPUs out there don't have one either, it's a pretty recent addition.

What I am also saying it that I haven't actually seen any evidence that such a wide decoder is actually worth it. Wide decode means a lot of delay in the circuitry which means poor clock speed scaling.

I guess I feel like the uop cache in x86 (both Skylake and Zen3) is because of the decode width problem. In performance-critical sections, Skylake / Zen3 go from 4-uops / tick (from the decoder) to 6-uops/tick (from the uop cache). In effect: its a way for x86 to reach higher uops/tick... but only in select areas of code (the areas that fit inside a uop cache).

Apple has a superior decoder: just 8-uops/tick no matter what. Its the "more expensive transistor budget" compared to a uop cache. Apple can achieve 8uops/tick across the entire 192kB L1 instruction cache, while Intel Skylake / AMD Zen3 can only achieve 4-uops/tick across a 48kB L1 cache (Skylake) / 32kB L1 (Zen3) cache, and a 6-uop/tick across a smaller region inside of the uOp cache.

 

[Hossein Almet]

There goes AMD's brief lead in gaming.

But it was never a real lead since the Ryzen 5000 launch was a paper launch.

The 5950X has been sitting on my desk for about month now, as I have been awaiting the shipment of the Dark Hero. When I placed the order, I saw a lot of high end motherboards had been sold out, a few gaming monitors were also sold out, also.

 

[RandallFlagg]

More leaks :

Intel B560 chipset motherboards will support overclocking, Core i9-11900K blows away Ryzen 7 5800X - TechnoSports

Intel will be launching its 11th Gen Rocket Lake-S desktop CPUs next month as well as the new 500 series chipset motherboards. This will bring a lot of new features that were only found only on AMD motherboards until now like PCIe Gen 4.0 and DDR4 3200 MHz memory support. Apart from these...

technosports.co.in

 

[300BaudBob]

They should call them Turbo Rockets--they'd sell like hot cakes and maybe cook them too.
But seriously I'm starting to really feel, between this and AMD's new offerings, the upgrade itch. Just need a little patience for all these new goodies to become obtainable.

 

[EarthDog]

I started reading this thread and all I saw were people taking jabs at Intel... ridiculous. This forum man... I swear........

Anyway, who knows how true this is... but it's a good sign so far. Wondering what the power draw will be (more than AMD I'd guess), but if IPC is back up there along with clocks and they keep the more reasonable pricing... it sounds like a solid option in the market to me...