Nice graph, go to page 4 where the reviewer admitted himself that the amd part was pulling a lot more power than indicated in that graph, lol

This is what the equivalent intel cpus do (14900t)

fevgatosNice graph, go to page 4 where the reviewer admitted himself that the amd part was pulling a lot more power than indicated in that graph, lol

This is what the equivalent intel cpus do (14900t)

It would be really helpful if you stopped exaggerating things and stop posting 35W intel numbers when there are no 35W AMD numbers in that chart. Let me help you. In the Anandtech chart where you mention it pulls a lot more power, here's the lovely stats for you:

Intel @ 35W - 39.3 Watts for 12,370 points = 314Points/W
AMD @35W - 45.1 Watts for 18,947 points = 420 points/W

Seriously man..

edit: before you try to argue the power consumption isn't exact, do your research to figure out that intel are actually in a more advantageous power/performance curve here. Just saying

mkppoIt would be really helpful if you stopped exaggerating things and stop posting 35W intel numbers when there are no 35W AMD numbers in that chart. Let me help you. In the Anandtech chart where you mention it pulls a lot more power, here's the lovely stats for you:

Intel @ 35W - 39.3 Watts for 12,370 points = 314Points/W
AMD @35W - 45.1 Watts for 18,947 points = 420 points/W

Seriously man..

There are no 35w amd parts at all. Not at stock.

In the graph above the Intel part is scoring 583 pts / watt. Is it a world record for desktop CPUs or what?

fevgatosThere are no 35w amd parts at all. Not at stock.

In the graph above the Intel part is scoring 583 pts / watt. Is it a world record for desktop CPUs or what?

Same argument holds true for 65w as well FYI. And lol erm..no i'm sorry it isn't a world record as much as you want it to be. As I just showed AMD always scales better with lower power and I literally pointed out just how far ahead they are at 35W and you ask if the intel part holds a world record at 35W? Do you see where you're going wrong here?

fevgatosThere are no 35w amd parts at all. Not at stock.

www.techpowerup.com/cpu-specs/ryzen-5-pro-5650ge.c2704
Good little CPU, I have one.

mkppoSame argument holds true for 65w as well FYI. And lol erm..no i'm sorry it isn't a world record as much as you want it to be. As I just showed AMD always scales better with lower power and I literally pointed out just how far ahead they are at 35W and you ask if the intel part holds a world record at 35W? Do you see where you're going wrong here?

The TPU graph has the 14900k at 583.6 pts / watt. Way higher than the 7950x. Let's deal with the facts, okay?

ARFWithout HT, and before AMD even caring to increase the core count from the old 16-core figure to at least 24-core figure?
Err, good luck waiting that.

HT seems overplayed by a lot of people. I dont think it was ever great to begin with, but its hay day was probably in the quad core era when 4 cores had scheduling bottlenecks in some games. IPC has always been more important, and the new direction of "smaller" cores obsolete's HT. HT's performance comes from filling in stalls/gaps in the pipe line (usually io wait), it typically was 0-10% benefit on processing throughput with maybe a few isolated use case where it would be higher, at the expense of much more power consumption.

Now we have chips with much more physical cores the use case for HT is diminishing and its a security headache.

I expect AMD to follow suit and remove SMT within one to two gens of Arrow Lake.

chrcolukIPC has always been more important

And how does that affect HT?

chrcolukit typically was 0-10% benefit on processing throughput with maybe a few isolated use case where it would be higher

Nope, AMD always had higher returns even Zen 1 had higher scaling it was 20~30% in CB IIRC. Of course that changes with application to application but in general it's always been higher for Zen.

R0H1TAnd how does that affect HT?

Nope, AMD always had higher returns even Zen 1 had higher scaling it was 20~30% in CB IIRC. Of course that changes with application to application but in general it's always been higher for Zen.

CB is one of those isolated use cases, bear in mind CB doesnt represent real world very well. I would be very surprised if SMT is not gone by AM6.

Then Rar or 7zip? Although I'd have check the latest revisions how well they do.

R0H1TThen Rar or 7zip? Although I'd have check the latest revisions how well they do.

Think of it this way, if SMT takes up e.g. 10% of the die space, think how many zen4c cores could be added or even bigger cache, then you might appreciate why the trade off is made. I feel it is inevitable it will happen at some point, but of course I could be wrong, only time will tell.

Consumers dont typically spend lots of time uncompressing or compressing huge amounts of data or doing things like software based encoding.

chrcolukCB is one of those isolated use cases, bear in mind CB doesnt represent real world very well. I would be very surprised if SMT is not gone by AM6.

HT is garbage even for / especially for games, problem is you can't turn it off unless you have more than 8 cores, else performance will plummet even in gaming.

chrcolukI would be very surprised if SMT is not gone by AM6.

Or: It wouldn't be a surprise if AMD keeps SMT as a competitive advantage in a world where competing ARM/RISCV/x86 CPUs lack SMT.

chrcolukHT seems overplayed by a lot of people. I dont think it was ever great to begin with, but its hay day was probably in the quad core era when 4 cores had scheduling bottlenecks in some games. IPC has always been more important, and the new direction of "smaller" cores obsolete's HT. HT's performance comes from filling in stalls/gaps in the pipe line (usually io wait), it typically was 0-10% benefit on processing throughput with maybe a few isolated use case where it would be higher, at the expense of much more power consumption.

Now we have chips with much more physical cores the use case for HT is diminishing and its a security headache.

I expect AMD to follow suit and remove SMT within one to two gens of Arrow Lake.

Disabling the HT will be a fatal, gross mistake. I would advise to work on 4-way SMT for future CPUs, instead.

So comparing 9700k to 9900k (which is the same CPU with hyperthreading) they are basically equal in games but 9900k does much better in heavy multithreaded tasks.

For highly threaded work it'll give a boost of ~25% (0-50% is the most common range, it depends on the application).

intel/comments/ll2ol9

It isnt enough to be highly threaded, so e..g. a highly threaded game, may have a lot of threads e.g. 8 threads, but it wont be constantly filling the pipe line,, in that scenario HT doesnt achieve a lot, its the combination of having more threads than cores and a completely filled pipe line that does it.

In an era where we have 8-32 cores on consumer CPUs, HT isnt needed to resolve scheduling bottlenecks anymore.

If I am wrong on the AM6 thing I wont hide and people can remind me of this post when the time comes. :toast:

fevgatosThe TPU graph has the 14900k at 583.6 pts / watt. Way higher than the 7950x. Let's deal with the facts, okay?

I really tried to not go on and write a whole paragraph trying to explain where you're going wrong. Also, your argument is so far from the truth that you are probably just trolling at this point. Regardless, i'll do it anyway so please find the time to read it and hopefully you might learn a thing or two.

First off, the 'fact' is the review I linked to since it has data for both CPU's. I mean, I point to a review where you can calculate points points/W yourself for both intel and AMD where it's clear intel is behind at 65W and most certainly 35W. That's a fact. What's not a fact is you pointing to a different review where there's no data for AMD at 35W and comparing it to a calculated value from another review. Why don't you just use the one review with both sets of data? Your 'way higher' quote is comedic, I mean there's no data for AMD at 35W in your chart so uh..what?

There's literally a mountain load of evidence out there which points to the fact that Zen 4 is more power efficient. Everyone in the datacenter space that i'm in touch with have been super impressed by Zen 3 and 4, in fact the jump from 3>4 was absolutely staggering even though in the desktop it was a bit meh. Reason being, most of the gains were lower in the V/F curve and each Zen core puts out a stupid amount of performance with very little watts. There's no defying physics here because there's simply less to power on Zen 4 than Golden Cove/Raptor Lake. Fewer registers, less ALU's, less FPU's, smaller uop cache, narrower load, store and reorder window, denser packaging, not to mention the process node disadvantage. I mean it would be an absolute miracle if Intel consumes less power, but alas they do not. What's surprising is the sheer amount of performance those little(big) zen cores spit out even with those 'disadvantages'.

You must be aware that very few out there are buying Intel chips in the server space. Do you know what one of the primary issues in that space is? Power. The P cores have to clock so low to meet their power targets that they simply cannot even remotely compete with Zen. Plus the P cores are a lot bigger due to the much larger core so they can't really cram that many into one socket. No wonder they are switching to e-cores in that space now.

I'll attach a snip of a comparison between i9-13980HX vs R9-7945HX a while ago, so basically 13900k and 7950x in laptop form. Since many power levels are covered here, its pretty interesting. Nothing surprising though.

chrcolukHT seems overplayed by a lot of people. I dont think it was ever great to begin with, but its hay day was probably in the quad core era when 4 cores had scheduling bottlenecks in some games. IPC has always been more important, and the new direction of "smaller" cores obsolete's HT. HT's performance comes from filling in stalls/gaps in the pipe line (usually io wait), it typically was 0-10% benefit on processing throughput with maybe a few isolated use case where it would be higher, at the expense of much more power consumption.

Now we have chips with much more physical cores the use case for HT is diminishing and its a security headache.

I expect AMD to follow suit and remove SMT within one to two gens of Arrow Lake.

The discussion regarding HT is more complicated than it seems at first glance. Even though it increases threads/parallelism, it's used more so to keep the individual cores fed better and depends entirely on whether an architecture is designed with HT in mind or not. If you don't use SMT in a core which is designed around it, it's individual cores will just get underutilized and tests on Zen 4 show pretty good scaling on MT and negligible difference in ST.

Sure, the 'perfect' architecture is one without HT where every core gets pretty much fully utilized without SMT and that's a good thing. But reality isn't always quite that and there will always be bottlenecks and underutilization. Question is how much, and whether spending additional transistors on SMT will be worth it. Note that a well implemented SMT will have negligible losses in ST, but there's the argument to be made that maybe the extra transistors used for SMT can be used to increase core performance. It's all a trade off, which i'm sure Intel did when they decided to not have it for ARL but it might entirely be the case that Zen 6 will not be significantly different from Zen 5 and SMT will have a net benefit. We'll see.

Security headache for sure though, as intel learned the hard way.

R0H1TNope, AMD always had higher returns even Zen 1 had higher scaling it was 20~30% in CB IIRC. Of course that changes with application to application but in general it's always been higher for Zen.

Slightly higher, yup. But not by much. Of course very test dependent but when Intel tends to get 25-30% from HT, Zen seems to get 30-35%.

atomsymbolOr: It wouldn't be a surprise if AMD keeps SMT as a competitive advantage in a world where competing ARM/RISCV/x86 CPUs lack SMT.

What? ARM does have SMT. Wasn't that in ARM's own A65 or around that time? Some of the manycore ARM things making headlines had something like 4-way SMT.

mkppoThere's no defying physics here because there's simply less to power on Zen 4 than Golden Cove/Raptor Lake. Fewer registers, less ALU's, less FPU's, smaller uop cache, narrower load, store and reorder window, denser packaging, not to mention the process node disadvantage.

I bet the last one has by far most to do with the comparatively low power usage :D

Not sure about R23 but R20 seems to be about ~20% for Zen 3.

www.anandtech.com/show/16261/investigating-performance-of-multithreading-on-zen-3-and-amd-ryzen-5000/2

chrcolukHT seems overplayed by a lot of people. I dont think it was ever great to begin with, but its hay day was probably in the quad core era when 4 cores had scheduling bottlenecks in some games. IPC has always been more important, and the new direction of "smaller" cores obsolete's HT. HT's performance comes from filling in stalls/gaps in the pipe line (usually io wait), it typically was 0-10% benefit on processing throughput with maybe a few isolated use case where it would be higher, at the expense of much more power consumption.

Now we have chips with much more physical cores the use case for HT is diminishing and its a security headache.

I expect AMD to follow suit and remove SMT within one to two gens of Arrow Lake.

Try turning it off then!

I thought the similar to you until I disabled it and it cut about 20-30% off my games and benchmarks. It soon got turned back on! I have a 5950x.

mkppoI really tried to not go on and write a whole paragraph trying to explain where you're going wrong. Also, your argument is so far from the truth that you are probably just trolling at this point. Regardless, i'll do it anyway so please find the time to read it and hopefully you might learn a thing or two.

First off, the 'fact' is the review I linked to since it has data for both CPU's. I mean, I point to a review where you can calculate points points/W yourself for both intel and AMD where it's clear intel is behind at 65W and most certainly 35W. That's a fact. What's not a fact is you pointing to a different review where there's no data for AMD at 35W and comparing it to a calculated value from another review. Why don't you just use the one review with both sets of data? Your 'way higher' quote is comedic, I mean there's no data for AMD at 35W in your chart so uh..what?

There's literally a mountain load of evidence out there which points to the fact that Zen 4 is more power efficient. Everyone in the datacenter space that i'm in touch with have been super impressed by Zen 3 and 4, in fact the jump from 3>4 was absolutely staggering even though in the desktop it was a bit meh. Reason being, most of the gains were lower in the V/F curve and each Zen core puts out a stupid amount of performance with very little watts. There's no defying physics here because there's simply less to power on Zen 4 than Golden Cove/Raptor Lake. Fewer registers, less ALU's, less FPU's, smaller uop cache, narrower load, store and reorder window, denser packaging, not to mention the process node disadvantage. I mean it would be an absolute miracle if Intel consumes less power, but alas they do not. What's surprising is the sheer amount of performance those little(big) zen cores spit out even with those 'disadvantages'.

You must be aware that very few out there are buying Intel chips in the server space. Do you know what one of the primary issues in that space is? Power. The P cores have to clock so low to meet their power targets that they simply cannot even remotely compete with Zen. Plus the P cores are a lot bigger due to the much larger core so they can't really cram that many into one socket. No wonder they are switching to e-cores in that space now.

I'll attach a snip of a comparison between i9-13980HX vs R9-7945HX a while ago, so basically 13900k and 7950x in laptop form. Since many power levels are covered here, its pretty interesting. Nothing surprising though.

I agree my graph doesn't have a 35w AMD part but the fact of the matter is that doesn't exist at stock. There is no 7950x at 35w variant. There is one for 12/13/14 900k etc. So out of the box, Intel has a lot more efficient CPUs. That's just a reality, I don't know why we keep arguing about it.

Usually comparisons should include prices. Originally both the 7700x and the 7800x 3d were priced at the same / higher than a 13700k / 14700k. They are both a lot less efficient in MT @ same power, and so is the 7600x compared to the 13600k. It's only the 7950x that has a lead. That's why I said Intel is more efficient both at out of the box - since they have the 35w power cpus - and at ISO power in most segments.

stimpy88Try turning it off then!

I thought the similar to you until I disabled it and it cut about 20-30% off my games and benchmarks. It soon got turned back on! I have a 5950x.

That's because you have 8 core ccds. On Intel it really doesn't. A 14900k loses around 10% in MT performance and nothing in games (actually it runs faster in those).

chrcolukHT's performance comes from filling in stalls/gaps in the pipe line (usually io wait), it typically was 0-10% benefit on processing throughput with maybe a few isolated use case where it would be higher, at the expense of much more power consumption.

Notes:

• In an out-of-order CPU, SMT's main advantage is the fact that the CPU knows that the 2 instruction streams belonging to the 2 threads are by definition independent and the CPU does not need to check+track those two instruction streams for cross-dependencies
• In desktop PCs and notebooks, many people no longer actively benefit from SMT because the core count (there can be 6-24 cores in an average new desktop machine build, up to 32 threads with SMT) is so high that the probability of utilizing all the cores (even without SMT) has become relatively low. Nevertheless, the fact remains that SMT by its nature leads to non-negligible performance improvements in out-of-order CPUs that were designed with SMT in mind.
• SMT benefits from a larger per-thread register file, thus AMD/Intel CPUs with APX (32 user-visible GPR registers, whereas amd64/x86-64 has 16 such registers) would benefit from SMT slightly more than non-APX CPUs

CifuFor the record: the 2CCD solutions historically have had the upper hand in boost clock:

7950X: 5750Mhz
7700X: 5550Mhz
7700: 5350Mhz
7800X3D: 5050Mhz

Not to mention the AMD firstly let the Zen4 to boost even more with the base 100Mhz clock tuning. This door was closed after the AGESA 1.0.0.3, after that only the eCLK tuning remained.

Agreed. My 5900X boosted 5150MHz in 2022. Now it is capped at 4950MHz. 200MHz easily shaved off via AGESA updates.

mkppoNot sure about R23 but R20 seems to be about ~20% for Zen 3.

www.anandtech.com/show/16261/investigating-performance-of-multithreading-on-zen-3-and-amd-ryzen-5000/2

I think Zen3-over-Zen2 IPC jump will not be seen now as Zen5-over-Zen4. It will be similar but just shy of that result despite various reports od 40+%. Nevetheless, it will be good but not mindblowing. All eyes on Zen5 X3D models.

fevgatosI agree my graph doesn't have a 35w AMD part but the fact of the matter is that doesn't exist at stock. There is no 7950x at 35w variant. There is one for 12/13/14 900k etc. So out of the box, Intel has a lot more efficient CPUs. That's just a reality, I don't know why we keep arguing about it.

Usually comparisons should include prices. Originally both the 7700x and the 7800x 3d were priced at the same / higher than a 13700k / 14700k. They are both a lot less efficient in MT @ same power, and so is the 7600x compared to the 13600k. It's only the 7950x that has a lead. That's why I said Intel is more efficient both at out of the box - since they have the 35w power cpus - and at ISO power in most segments.

You can't just change your narrative as you go and that's why we keep arguing about FYI. So this is your initial comment and you were clearly implying Intel CPU's are more efficient at ISO power levels. Now suddenly you're talking prices, segments, out of box specs for OEM only CPU's etc. Really strange.

fevgatosAt ISO power limits most Intel CPUs are more efficient than their AMD counterparts in MT. By a lot.

Which is blatantly false because as we know, AMD is much more efficient at pretty much every realistic power limit. Now if your new argument is that Intel makes a relatively rare OEM only 35W 13x00T lineup that is more efficient out of the box than a comparable 65/105W AMD parts in CB R23 nT, sure. Let's ignore the fact that AMD can very easily be set at 35W, which will make it more efficient, or that the intel parts can't be readily bought without buying a whole system so your price comparison is also moot since you can't readily buy these separately anyway.

mkppoYou can't just change your narrative as you go and that's why we keep arguing about FYI. So this is your initial comment and you were clearly implying Intel CPU's are more efficient at ISO power levels. Now suddenly you're talking prices, segments, out of box specs for OEM only CPU's etc. Really strange.


Which is blatantly false because as we know, AMD is much more efficient at pretty much every realistic power limit. Now if your new argument is that Intel makes a relatively rare OEM only 35W 13x00T lineup that is more efficient out of the box than a comparable 65/105W AMD parts in CB R23 nT, sure. Let's ignore the fact that AMD can very easily be set at 35W, which will make it more efficient, or that the intel parts can't be readily bought without buying a whole system so your price comparison is also moot since you can't readily buy these separately anyway.

I made 2 different claims. Firstly, that out of the box they are more efficient, which is true even for no other reason than that there are 35 and 65w versions. That is a clear cut case no arguing required, it's just true.

And then yes, I also said that in most segments Intel are more efficient at ISO power- which is also correct. Segments - going by MSRP, i5 13600k vs r5 7600x, i7 13700k vs r7 7700x, i7 14700k vs r7 7800x 3d, i9 13900k / 14900k vs 7900x / 7950x. In most of those Intel wins in both ST and MT efficiency, it's only at the top end (7950x) that amd has the edge in MT.

Of course you can buy those parts separately, what are you talking about man?

fevgatosI made 2 different claims. Firstly, that out of the box they are more efficient, which is true even for no other reason than that there are 35 and 65w versions. That is a clear cut case no arguing required, it's just true.

Of course you can buy those parts separately, what are you talking about man?

You made the second claim after you realised you lost the first two arguments. You started off with a useless ST power test, then we shifted to MT and this is what you had to say:

fevgatosNo, it just has a lower power limit.

At ISO power limits most Intel CPUs are more efficient than their AMD counterparts in MT. By a lot.

Which is false. At the same ISO power limits AMD CPU's are more efficient. Note that you mentioned nothing about being 'out of the box' or anything to do with price or segment. You simply implied that AMD is inherently more inefficient. Which is, again, false. Your other dubious 'claims' were made later.

fevgatosAnd then yes, I also said that in most segments Intel are more efficient at ISO power- which is also correct. Segments - going by MSRP, i5 13600k vs r5 7600x, i7 13700k vs r7 7700x, i7 14700k vs r7 7800x 3d, i9 13900k / 14900k vs 7900x / 7950x. In most of those Intel wins in both ST and MT efficiency, it's only at the top end (7950x) that amd has the edge in MT.

So even in your own graph, 13600k is behind 7600x, 13700k is behind 7700x, 14700k is behind 7800x etc in efficiency. You probably meant non-k CPU's. In that case, we also need the non-x AMD CPU's in the graph. Since those aren't there, there's no real comparison.

Buy yes, your final argument is that Intel actually makes 35W CPU's which are more efficient out of the box than the higher power AMD parts since AMD doesn't make 35W CPU's. That's correct. But don't pretend like that's the argument you started off with. Also, those are OEM only parts, that's what I was talking about. You can buy these, but those are usually pulled from systems and not on the retail channel.