Tuesday, February 9th 2021

AMD Zen 4 Reportedly Features a 29% IPC Boost Over Zen 3

While AMD has only released a few Zen 3 processors which are still extremely hard to purchase for RRP we are already receiving leaks on their successors. Zen 3 Milan processors will likely be the final generation of AM4 processors before the switch to AM5. AMD appears to be preparing a bridging series of processors based on the Zen 3+ architecture before the release of Zen 4. Zen 3+ is expected to be AMD's first AM5 CPU design and should bring small IPC gains similar to the improvements from Zen to Zen+ in the range of 4% - 7%. The Zen 3+ processors will be manufactured on TSMC's refined N7 node with a potential announcement sometime later in 2021.

Zen 4 is expected to launch the next year in 2022 and will bring significant improvements potentially up to 40% over Zen 3 after clock boosts according to ChipsandChesse. A Zen 4 Genoa engineering sample reportedly performed 29% faster than an existing Zen 3 CPUs at the same clock speeds and core counts. The Zen 4 architecture will be manufactured on a 5 nm node and could potentially bring another core count increase. This would be one of the largest generational improvements for AMD since the launch of Ryzen if true. Take all this information with a heavy dose of skepticism as with any rumor.
Source: ChipsandCheese
Add your own comment

143 Comments on AMD Zen 4 Reportedly Features a 29% IPC Boost Over Zen 3

#101
lynx29
I'm not counting out AM4 yet, AMD has switched their tune a couple times on backwards compatibility with mobo's, a Zen 3 + on refined 7nm node probably will still work in AM4 with a bios update on 'select mobos' aka higher end ones to mid-upper tier ones. That's my guess anyway we will see.
Posted on Reply
#102
lexluthermiester
Vendor
Same can be said about 3600 too?
That is called binning. Everyone does it.
Night
Thought it's supposed to be graphene.
Possibly. The research I've read indicated that the concept needs work.

Back to the topic of the thread;
My question about Zen4 is what will the socket be? Will it be dual channel memory or are they going to finally step up to triple channel?
Posted on Reply
#103
Mats
lexluthermiester
My question about Zen4 is what will the socket be? Will it be dual channel memory or are they going to finally step up to triple channel?
With DDR5 in itself giving more bandwidth, I can't see why they'd go for more channels on mainstream.
Posted on Reply
#104
lexluthermiester
Mats
With DDR5 in itself giving more bandwidth, I can't see why they'd go for more channels on mainstream.
So people can do more work faster? Why else would you add channels of memory? It's the whole point of the concept. And yes we need to advance forward from dual channel configurations in the mainstream sectors of computing.
Posted on Reply
#105
Makaveli
R0H1T
What performance claims :confused:

Pretty sure Alderlake tops out at 8 "big cores" so there's no reason for AMD to counter them at MSDT top end, for HEDT Intel's not even in the frame. Having said that it'd be nice if zen4 or zen3+ launches in 2021, even if only to reduce the process on zen3 parts.
Performance as in IPC claims not core count.
Posted on Reply
#106
cueman
zen3 to zen4 29% more power, practical its 7nm to 5nm
hmm...we will see...looks clear percent calculate summum....lol


anyway,exiting,
well, lets see 1st 14nm rocket lake vs 7nm zen 3 and then...little bit but not yet same level process tech results...,
and i mean 10nm adler lake-s vs 7nm zen 3

then we go i guess 2022 and might be 5nm zen 4 get against 7nm meteor lake or even 3nm ones,bcoz looks sure that intel start trust also TSMC products.

is it there 2022/2023 both, amd and intel 3nm process ech..if so,then only engineer skills shows winner, not handicap advance like now aka 14nm cpu compare 7nm cpu...

naaaah.. useless.

sure is that intel get amd ,its only matter of time,interesting times....

well,anyway,competition is always good!!!
Posted on Reply
#107
LemmingOverlord
john_
How have we moved from "3-5% and you should be happy" every year, to the "if it offers less than 15% we are going to be disappointing"?
The problem is that sooner or later AMD becomes as greedy as Intel and realises it has such a performance/technology lead that it can squeeze every last buck from the consumer with minor performance bumps. When confronted with this fact, they will happily point to the fact that "Intel did it first" and promptly give consumers the middle finger. The student becomes the teacher, etc...
Posted on Reply
#108
Slizzo
lexluthermiester
So people can do more work faster? Why else would you add channels of memory? It's the whole point of the concept. And yes we need to advance forward from dual channel configurations in the mainstream sectors of computing.
Cost. That extra channel will add cost and complexity to the boards, and I don't see Intel nor AMD adding another memory channel when they'll already be realizing better bandwidth with just the switch from DDR4 to DDR5.
Posted on Reply
#109
lexluthermiester
Slizzo
Cost. That extra channel will add cost and complexity to the boards, and I don't see Intel nor AMD adding another memory channel when they'll already be realizing better bandwidth with just the switch from DDR4 to DDR5.
Nonsense. Currently, 4 DIMM slots are generally made for most boards. 3 would lower cost. 6 would be a minimal increase in cost. Complexity is a given.
Posted on Reply
#110
mechtech
"according to ChipsandChesse"

nough said
lol
Posted on Reply
#111
Mats
lexluthermiester
So people can do more work faster? Why else would you add channels of memory? It's the whole point of the concept. And yes we need to advance forward from dual channel configurations in the mainstream sectors of computing.
You clearly misunderstood my "why". :D

I didn't mean "I can't see what it would give consumers".
Posted on Reply
#112
InVasMani
lexluthermiester
Nonsense. Currently, 4 DIMM slots are generally made for most boards. 3 would lower cost. 6 would be a minimal increase in cost. Complexity is a given.
Triple channel DDR5 one less slot than DDR4 higher capacity and less voltage with the bandwidth of quad channel DDR4. Triple channel would have been fine even on DDR4 to have been more commonly used it would've been good on a ThreadRipper in fact if they'd want to lower the cost on it with more reasonable core count chips that have a bit better frequency scaling and a nice mixture of bandwidth. The Capacity wouldn't have been a issue to most 48GB to 96GB is still a lot.
Posted on Reply
#113
TechLurker
Mats
I don't see what we could expect here besides higher prices. 3000XT made absolutely no one happy. 100 MHz extra makes zero difference in the long run.
The difference between a theoretical 5000 XT and the 3000 XT is that the 5000 series itself is the end of the AM4 line, where it does make sense to put out the best, fully matured CPUs/APUs possible for the platform, whether they provide a bit more efficiency at the same speeds (effectively run a bit cooler), or provide a bit more punch (extra 100-200 Mhz). As well, pushing the non-XTs down 50-100 USD (depending on model or purpose) and putting the XTs at the old, original MSRP price point of the CPUs they're replacing, gives them another means to compete somewhat on price with Intel, while slotting in a few more CPUs at key tiers.

The other bit is that AMD, like Intel, will be going into a bit of a transition period from PCIe 4.0 and DDR4 to PCIe 5.0 and DDR5, and will have to somewhat start anew (more in I/O capability and making DDR5 truly work), so some new builders may still opt to buy into the last gen but also aim for the best of the last gen, as it's all been time-tested and proven.
Posted on Reply
#114
Alexa
If this is true, looks like me holding onto my 3900X till Zen 5 is paying off.
Posted on Reply
#115
DAWMan
Guess it’s time to get Windows 8.1 and retire my 4790k’s. Get me a W 1250P Xeon and in 2024 a Zen 3+.
Posted on Reply
#116
Mats
TechLurker
The difference between a theoretical 5000 XT and the 3000 XT is that the 5000 series itself is the end of the AM4 line, where it does make sense to put out the best, fully matured CPUs/APUs possible for the platform, whether they provide a bit more efficiency at the same speeds (effectively run a bit cooler), or provide a bit more punch (extra 100-200 Mhz).
Those minimal gains doesn't make sense for $100 extra. When 3000XT showed up, AMD had only squeezed out just a tiny bit extra clock speed. Next time, that same, now even more mature 7 nm might deliver even smaller improvements, assuming that the biggest improvements can be done in the beginning of a process.
TechLurker
As well, pushing the non-XTs down 50-100 USD (depending on model or purpose) and putting the XTs at the old, original MSRP price point of the CPUs they're replacing, gives them another means to compete somewhat on price with Intel, while slotting in a few more CPUs at key tiers.
That's not how it worked when 3000XT was launched. NOTHING happened with the prices for the older models. Why? Because they didn't have enough impact on the market to affect the pricing of existing models. The prices had already dropped to those lower points long before the XT launch.
TechLurker
The other bit is that AMD, like Intel, will be going into a bit of a transition period from PCIe 4.0 and DDR4 to PCIe 5.0 and DDR5, and will have to somewhat start anew (more in I/O capability and making DDR5 truly work), so some new builders may still opt to buy into the last gen but also aim for the best of the last gen, as it's all been time-tested and proven.
Maybe in theory, but again, judging by the 3000XT, none of those miniscule improvements makes sense in the long run. Show me ONE decent review that's praising the 3000XT's.
You have to weigh in cost here, people aren't going for ridiculously small improvements at any cost. For the same reason, the 3600 is much more popular than the 3600X.

Some people seems to believe that a 5000XT will bring like 500 MHz extra clock speed, or something like that. I'm not sure even that would justify a $100 price hike. Either way, that won't happen.
Posted on Reply
#117
Dredi
lexluthermiester
So people can do more work faster? Why else would you add channels of memory? It's the whole point of the concept. And yes we need to advance forward from dual channel configurations in the mainstream sectors of computing.
Just get an entry level epyc with 8 channels @ just $400 or so, one for each core. Epyc is the platform for doing more work, ryzen is just for gaming and browsing the internet. ;)

For real though, the bandwidth of ddr4 is hardly a bottleneck now and will definitely not be one with ddr5, save for some silly iGPU use cases.
Posted on Reply
#118
lexluthermiester
Dredi
ryzen is just for gaming and browsing the internet.
Bollocks! :roll:
Dredi
For real though, the bandwidth of ddr4 is hardly a bottleneck now
It can be for some tasks, thus the push for an expansion on the number of channels.
Dredi
will definitely not be one with ddr5
DDR5 is a bit of a solution without a problem. It's also not going to do mainstream any good during the early stages of release.
Dredi
save for some silly iGPU use cases.
That's not so silly anymore. With AMD and Intel iGPU quality and capability pushing into respectable ranges, the system RAM they use needs to step up in performance to match.
Posted on Reply
#119
InVasMani
It's not complicated I just want a quad channel DDR4I single CCX refreshment to the original Zen1 ThreadRipper line up more or less. It would be more cost effective and frequency scaling would have a higher ceiling and power more reasonable while having the upside to the improved IPC, latency, and frequency scaling. As far as bandwidth goes if it's not needed why bother with DDR5? If bandwidth doesn't matter why are people pushing infinity fabric far as they can!!? What is it for less bandwidth.
Posted on Reply
#120
RandallFlagg
lexluthermiester
Nonsense. Currently, 4 DIMM slots are generally made for most boards. 3 would lower cost. 6 would be a minimal increase in cost. Complexity is a given.
Isn't DDR5 inherently dual channel on a single DIMM? See here.

That should mean we can have two DIMM slots, and quad channel. As you say, that should be a net sum zero game as far as cost. Triple channel would probably be out though.
Posted on Reply
#121
lexluthermiester
RandallFlagg
Isn't DDR5 inherently dual channel on a single DIMM? See here.

That should mean we can have two DIMM slots, and quad channel. As you say, that should be a net sum zero game as far as cost. Triple channel would probably be out though.
While true, this does not bring a lot of tangible benefit as each channel operates at 32bits as opposed to the 64bit data bus current memory offers.
Interestingly, every DDR5 DIMM will work in dual-channel mode by itself, as the memory banks are now addressable on two independent 32-bit sub-channels (40-bit for ECC memory), which is a similar design to GDDR6 and LPDDR4 memory.
This means that the data bus is being split in half to make additional channels. While this scheme can have a few processing benefits, the actual performance increase will be minimal. As alluded to elsewhere in that same page, there are on-DIMM voltage regulators for power saving and it seems logical that the dual channel split is a part of that design.

In the end it's a trade off which will not do much to improve performance.
Posted on Reply
#122
RandallFlagg
lexluthermiester
While true, this does not bring a lot of tangible benefit as each channel operates at 32bits as oppose to the 64bit data bus current memory offers.
We don't know that yet, not until we have benchmarks, as there are other changes said to offset that :

Your point :
"Each DIMM can handle two 32-bit memory channels instead of only a single 64-bit channel."

The counterpoint:
"Since each bank operates independently of each other, the burst length can be doubled and greater efficiency can be achieved. That means, for instance, DDR5 SDRAM can perform two 64-byte operations in the same time it takes DDR4 SDRAM to perform just one operation."

Source: techxplore.com/news/2020-07-ddr5-sdram-standard-boost-dual-channel.html
Posted on Reply
#123
lexluthermiester
RandallFlagg
We don't know that yet, not until we have benchmarks, as there are other changes said to offset that :
We don't need the benchmarks. History and simple math tell us enough.
RandallFlagg
The counterpoint:
"Since each bank operates independently of each other
But at 32bits instead of 64bits.
RandallFlagg
the burst length can be doubled
Dubious claim. If true that would be great and a big step forward, but the proof is in the pudding and there have been not demonstrations as of yet. You'd think they'd want to show it off.
RandallFlagg
greater efficiency can be achieved
That part is the part that has some merit.
RandallFlagg
That means, for instance, DDR5 SDRAM can perform two 64-byte operations in the same time it takes DDR4 SDRAM to perform just one operation."
Again, the math does not support that claim based on what specifications have been stated.

It's not believable until they show it and explain in more detail how things work. So far the details do not support the claims stated.
Posted on Reply
#124
Dredi
lexluthermiester
It can be for some tasks, thus the push for an expansion on the number of channels.
Name one. And no silly iGPU gaming stuff. Focus on your ’do more work faster’ statement. 8 channel zen2 epyc is just $400 or so. Why don’t you buy that?
InVasMani
If bandwidth doesn't matter why are people pushing infinity fabric far as they can!!?
Infinity fabric makes the CCX itself do stuff faster, and also lowers the memory latency (which has nothing to do with bandwidth and cannot be lowered by adding channels).
Posted on Reply
Add your own comment