Question Raptor Lake - Official Thread

[Hulk]

Since we already have the first Raptor Lake leak I'm thinking it should have it's own thread.
What do we know so far?
From Anandtech's Intel Process Roadmap articles from July:

Built on Intel 7 with upgraded FinFET
10-15% PPW (performance-per-watt)
Last non-tiled consumer CPU as Meteor Lake will be tiled

I'm guessing this will be a minor update to ADL with just a few microarchitecture changes to the cores. The larger change will be the new process refinement allowing 8+16 at the top of the stack.

Will it work with current z690 motherboards? If yes then that could be a major selling point for people to move to ADL rather than wait.

 

[pakotlar]

Interesting compare to the M1 Ultra: 17.7k ST 24k MT in Geekbench 5. Apple + ARM have done amazing things. Looks like this gen 13900K and 7950X will compete at more power with Apple’s M2 Ultra.

Mac Studio Benchmarks - Geekbench

 

[tamz_msc]

Now, how many people publishing GB5 scores know all that? And have you seen the actual source code?

How many people who publish Cinebench scores know what renderer it uses, how it works? Have they seen the actual source code? What does any of those questions have anything to do with the validity of a benchmark?

Why do so many of the MT tests not even make full use of the CPU? What the hell is it doing in there?

That is like, demonstrably false?

Here, using Det0x's run, barring AES, we have the following MT scaling:



So yeah, it is straightaway misleading to say that Geekbench doesn't make full use of the CPU in MT tests.

https://browser.geekbench.com/v5/cpu/16241068

 

[Gideon]

PBO does not affect ST performance as far as I am aware; and no, it will take the 7950X, assuming it actually has a 5.7 GHz boost according to the WCCFTech rumor, to match this Raptor Lake score in GB5 ST. That assumes Geekbench scales perfectly with frequency.

I'll try to do some rough napkin-calculations to guess the Geekbench ST score for a Ryzen 9 7950X.

Here are my presumptions and though process:

1. A decent 5950X gets around 1700 points (+- 25p) in GB5 single thread (under windows) with an average clock speed just below 5Ghz (one has to be logged in to see the JSON at the link).
2. 7950X seems to have a single-threaded clock speed of 5.7 Ghz. If it is similar to Zen 3, it will temporarily boost 100-150Mhz higher, so even if it's slightly lower, the ballpark is probably there.
3. Zen 4 has AVX-512 support. For Intel 11th series this only affected the AES scores and provided no difference in MT but a slight gain of 2% in ST.
4. Zen 4 also has DDR5, which seems to have no effect on ST score, but helps the MT score by at least 1000 points (probably more, if faster DDR5 is used)
5. AMD has claimed 8-10% extra IPC for Zen 4.

Here and here are some to a Ryzen 5 5600 (non-x) running at 4.4 Ghz compared to a Ryzen 5950X running at about 5 Ghz. Such comparisons are tricky, because the uplift is all over the place (6% - 15%). This is somewhat relevant, as the clock gain is almost identical to the hypothetical gain of the 7950X 5.0 Ghz to 5.7 Ghz(13% vs 14% uplift).

So, when extrapolating all this, we get:

Optimistic scenario:
1700 * 1.14 (clock-speed scaling) * 1.10 (IPC) * 1.02 (AVX-512) = 2175 points ST

Very pessimistic scenario:
1700 * 1.08 (worse clock-speed scaling) * 1.08 (IPC lower bound) * 1.00 (AVX-512 ends up not helping or counted into IPC) = 1982 points ST

The reality is almost certainly in between these two extremes.

I think that a score over 2000 points is all but guaranteed with decent ram. I'm even relatively confident it's over 2100 but I don't think it will quite reach 2200 points stock, bar some outliers.

TL;DR:

According to the napkin calculations above, my best guess is for a Ryzen 7950X Geekbench ST score (stock-config) is about 2100 points.

It might do better, but I'd rather be slightly conservative and surprised, than overhyped and disappointed.

EDIT: spelling

 

[Zucker2k]

5950X Cinebench R23 scores at 4.6 GHz and 5.0 GHz:

4.6 GHz: 31306
5.0 GHZ: 34745

View attachment 65546

https://twitter.com/x/status/1556319963492671488

A Zen 3 at 355w is most likely using a chiller, and even that couldn't prevent it from crashing.

 

[Det0x]

How many people who publish Cinebench scores know what renderer it uses, how it works? Have they seen the actual source code? What does any of those questions have anything to do with the validity of a benchmark?

That is like, demonstrably false?

Here, using Det0x's run, barring AES, we have the following MT scaling:

View attachment 65547

So yeah, it is straightaway misleading to say that Geekbench doesn't make full use of the CPU in MT tests.

https://browser.geekbench.com/v5/cpu/16241068

Do note the memory frequency needed to feed these cores at static 4900mhz.. (4466MT/s)
MT scale very much with memory bandwidth in GB5 while ST actually prefers low latency (synced 1900:3800 memory: infinity fabric scores better a tiny bit better on ST)


Here is a more normal run with PBO and synced 1900:3800:

Micro-Star International Co., Ltd. MS-7D51 - Geekbench

Benchmark results for a Micro-Star International Co., Ltd. MS-7D51 with an AMD Ryzen 9 5950X processor.

browser.geekbench.com

A Zen 3 at 355w is most likely using a chiller, and even that couldn't prevent it from crashing.

View attachment 65549

Those power numbers seem alittle off.. here is mine:

 

[DrMrLordX]

How many people who publish Cinebench scores know what renderer it uses, how it works? Have they seen the actual source code?

Practically nobody, and nobody, which is why I went out of my way to bring up R15 being a poor representation of Cinema4D performance back when it was the latest Cinebench version available. For all we know, R23 may not be good for profiling Maxon's software performance either.

That is like, demonstrably false?

Wrong! I've run GB5 and found that it leaves cores at sub-100% utilization on several tests on my 3900X. The only ones that actually heat up the CPU are Ray Tracing and Structure from Motion.

 

[pakotlar]

It's click bait article. The 3995WX Stomps that 13900K in MT

View attachment 65461

I wouldnt call that a stomping. 2k faster for 2.5x the cores, and half the single threaded performance, and more power.

 

[pakotlar]

Do note the memory frequency needed to feed these cores at static 4900mhz.. (4466MT/s)
MT scale very much with memory bandwidth in GB5 while ST actually prefers low latency (synced 1900:3800 memory: infinity fabric scores better a tiny bit better on ST)
View attachment 65550

Here is a more normal run with PBO and synced 1900:3800:
View attachment 65553

Micro-Star International Co., Ltd. MS-7D51 - Geekbench

Benchmark results for a Micro-Star International Co., Ltd. MS-7D51 with an AMD Ryzen 9 5950X processor.

browser.geekbench.com

View attachment 65552


Those power numbers seem alittle off.. here is mine:
View attachment 65554

Edit: You’re at 303W at 57C max temp, suggesting your room was close to 0C (good water will do 57C over ambient at 300W). At 70-80C you will use more power + need more voltage to remain stable. I would guess that you wouldn’t hit much under 350W, curious if you can try.

In any case, all of this is academic until we get retail hands on. 13900K looks to be overclockable as well, and people will try to run those with a combination of undervolting, freezing ambient temps, and ridiculous high clocks for 40K+ at sub 350W, we’ll see.

For a refresh coming 1 year after ADL, Raptor Lake is excellent. The fact that we’re even comparing it to Zen 4, which is a full generation over the best CPU line ever made is pretty remarkable. Zen 4 will probably be better for my use case (all big cores with AVX512), esp if they ever bump core count, but Raptor Lake is looking impressive, and it makes me excited for Intel’s future lineup.

 

[Abwx]

Edit: You’re at 303W at 57C max temp, suggesting your room was close to 0C (good water will do 57C over ambient at 300W). At 70-80C you will use more power + need more voltage to remain stable. I would guess that you wouldn’t hit much under 350W, curious if you can try.

At 57C you need more voltage than at 70-80C as counter intuitive as it may sound, mosfets conduct better as temp increase up to a given level at wich point the behaviour goes the other way, the crossover point should be at 100-120C or so.

 

[pakotlar]

At 57C you need more voltage than at 70-80C as counter intuitive as it may sound, mosfets conduct better as temp increase up to a given level at wich point the behaviour goes the other way, the crossover point should be at 100-120C or so.

So much I don’t know. Could you point me to some reading on the subject that you like?

 

[tamz_msc]

Wrong! I've run GB5 and found that it leaves cores at sub-100% utilization on several tests on my 3900X. The only ones that actually heat up the CPU are Ray Tracing and Structure from Motion.

Then explain how Det0x gets the scores that he does.

 

[Timmah!]

I wouldnt call that a stomping. 2k faster for 2.5x the cores, and half the single threaded performance, and more power.

More importantly, 2x faster for how many times more money? 5 times?
Obviously, if one has the need for it, its worth it. But for someone who can get away with 13900k (7950x), those provide "lot of music" for the money.

 

[shady28]

So, how many actually plan to get a 13900K or 7950X, given that they will likely be $800 MSRP and require a $300++ motherboard and $350 RAM to effectively use?

I'm thinking the 13600K and maybe 13700 / 13700K will be more interesting anyway.

 

[Markfw]

So, how many actually plan to get a 13900K or 7950X, given that they will likely be $800 MSRP and require a $300++ motherboard and $350 RAM to effectively use?

I'm thinking the 13600K and maybe 13700 / 13700K will be more interesting anyway.

I have A 7950x planned, already have the ram, the best you can buy.

 

[pakotlar]

More importantly, 2x faster for how many times more money? 5 times?
Obviously, if one has the need for it, its worth it. But for someone who can get away with 13900k (7950x), those provide "lot of music" for the money.

Yeah thats a good point. I kind of dont care about perf/watt as much as perf/$, where pain is $, how much the CPU and any mobo/ram/case changes cost, whether it makes my room hot requiring updates AC (so $ too), and what the failure rate is on the cooling solution I need is (can I air cool). Small weight on fun factor: how easily can I overclock to meaningful results? My 5950X is not that fun, because my ram holds me back (try finding 128GB of tight timings 3600+mhz DDR4), and because 5950X is so power dense my NH-D15 in a case with crazy air flow can’t cool effectively beyond 200W. And at 200W it still thermally throttles over time (even at 80c and below). Otherwise the CPU is fantastic.

Circling back to Raptor Lake, Raichu mentioned it can be cooled on normal water at 350W to ~80s temp, which got my attention. Better heat spreader? Larger die the reason? BS?

 

[shady28]

<snip>

Circling back to Raptor Lake, Raichu mentioned it can be cooled on normal water at 350W to ~80s temp, which got my attention. Better heat spreader? Larger die the reason? BS?

I think what Raichu said was water cooled, not specifically an AIO. You can make a custom loop with a 360mm and do 350W. You're talking about $300 to get started and probably have to add $200+ including an upgraded water block.

250W is somewhat reasonable for typical enthusiasts, you can get there with the right case and AIO and not having your own lab for this kind of thing. 350W is really hard core.

Honestly if everyone is talking about max OC, you can pretty much discard everything read on any mainstream site. You have to look deeper than they do.

 

[desrever]

This simply isn't true. Before my present laptop that has an i7-11370H, I used a desktop with an i7-3770. I can definitely tell the difference between the two when using a web browser. Stronger cores, bigger cache, faster RAM - they all add up.

Obviously if you compare to a 10 year old CPU that new one is noticeably faster. I was talking about comparing new high end CPUs to each other.

 

[Mopetar]

I think we're in agreement, the term IPC is being used by the community to imply how many instructions an entire package can do - not just the CPU. It includes core, cache, memory controller, and memory - including all those latencies and bandwidth.

Since a CPU is essentially useless without being connected to memory, IPC naturally must consider caches and the rest of the memory system. Thinking it's just the core is a misguided as thinking performance is just the clock speed.

IPC, performance/clock, these things are irrelevant to any kind of real investigation of how well a particular platform and CPU perform.

I'm not sure how you can claim that. IPC in a vacuum may be meaningless, but the IPC and the clock speed effectively determine the performance for a given program. The program has some number and type of instructions, the IPC (as a general number, but it could be more specific for a given program) will indicate the number of cycles it will take that processor to execute the program, and the clock speed will let you determine the amount of time it will take for a CPU to complete those cycles.

I think you've made your claim in the basis that the memory subsystem is not connected to or related to IPC, when that's clearly not the case. Consider Zen 3D which reduced clock speed, but saw increased performance in games and a small handful of other applications due to having more cache, even if the latency of that cache increased slightly as well. Clearly the CPU must have been able to execute more instructions within some number of clock cycles in order to achieve greater performance given a decrease (or even the same) clock speed.

Really the only potential issue is that there's no one true way to measure IPC. You can talk about IPC improvements strictly in terms of a single benchmark like Geekbench or even just a particular single SPEC benchmark. Or you can talk about it as an average across a broad set of applications and benchmarks. At worst you might mislead yourself into making bad conclusions that won't be borne out in reality by selecting a highly specific way of measuring IPC and trying to apply that in ways where it doesn't generalize or have much relevance to the way IPC was defined.

 

[shady28]

[

Since a CPU is essentially useless without being connected to memory, IPC naturally must consider caches and the rest of the memory system. Thinking it's just the core is a misguided as thinking performance is just the clock speed.



I'm not sure how you can claim that. IPC in a vacuum may be meaningless, but the IPC and the clock speed effectively determine the performance for a given program. The program has some number and type of instructions, the IPC (as a general number, but it could be more specific for a given program) will indicate the number of cycles it will take that processor to execute the program, and the clock speed will let you determine the amount of time it will take for a CPU to complete those cycles.

I think you've made your claim in the basis that the memory subsystem is not connected to or related to IPC, when that's clearly not the case. Consider Zen 3D which reduced clock speed, but saw increased performance in games and a small handful of other applications due to having more cache, even if the latency of that cache increased slightly as well. Clearly the CPU must have been able to execute more instructions within some number of clock cycles in order to achieve greater performance given a decrease (or even the same) clock speed.

Really the only potential issue is that there's no one true way to measure IPC. You can talk about IPC improvements strictly in terms of a single benchmark like Geekbench or even just a particular single SPEC benchmark. Or you can talk about it as an average across a broad set of applications and benchmarks. At worst you might mislead yourself into making bad conclusions that won't be borne out in reality by selecting a highly specific way of measuring IPC and trying to apply that in ways where it doesn't generalize or have much relevance to the way IPC was defined.

IPC isn't what you think it is. People are just re-defining the meaning of IPC to suit their view of how things work.

If you want to see real IPC, you don't need to run a benchmark. You can look in the technical docs for the CPU and see the number of clocks any particular instruction will take, provided that the memory subsystem can deliver the data.

So in your own words, you are also testing memory latency / bandwidth. In other words you are doing test of the platform, not the CPU's IPC potential.

 

[dullard]

IPC isn't what you think it is. People are just re-defining the meaning of IPC to suit their view of how things work.

If you want to see real IPC, you don't need to run a benchmark. You can look in the technical docs for the CPU and see the number of clocks any particular instruction will take, provided that the memory subsystem can deliver the data.

So in your own words, you are also testing memory latency / bandwidth. In other words you are doing test of the platform, not the CPU's IPC potential.

Other than academic debates, does IPC potential have any use whatsoever? All that matters in real life is how a particular system does on a user's applications.

• Performance = (Max IPC from technical docs) * (Efficiency of system for given benchmark) * (Clock)

Is that really any different for the user than:

• Performance = (Effective IPC) * (Clock)

The system matters and that will impact the effective IPC.

 

[fleshconsumed]

So, how many actually plan to get a 13900K or 7950X, given that they will likely be $800 MSRP and require a $300++ motherboard and $350 RAM to effectively use?

I'm thinking the 13600K and maybe 13700 / 13700K will be more interesting anyway.

I'll just stay on AM4 and hopefully get clearance priced 5950x. My board with 5950x will be plenty enough for me until AM5 platform costs come down in price and I can find AM5 motherboard that fits my needs.

That said, I'm sure plenty people will be getting Raptor Lake/Zen4 when they come out. There are always plenty of people who have to/want to buy new stuff.

 

[shady28]

Other than academic debates, does IPC potential have any use whatsoever? All that matters in real life is how a particular system does on a user's applications.

• Performance = (Max IPC from technical docs) * (Efficiency of system for given benchmark) * (Clock)

Is that really any different for the user than:

• Performance = (Effective IPC) * (Clock)

The system matters and that will impact the effective IPC.

Performance of the system is not IPC * clock, because we don't know what IPC is in the first place, and if we did it would be very task specific.

My point is and has been, IPC is irrelevant and no one is measuring it anyway, they're just abusing the concept - mostly to arrive at a pre-conceived conclusion.

For example, this is my 10850K vs a 5800X single core. Can I derive IPC by dividing clock rate by this?


Or is this?


The correct answer is, neither. They are just testing an algorithm that is commonly used for a specific task. They may, or may not, be limited by the speed the CPU can execute that particular mix of instructions for that task.

The only thing that matters is how the system performs on applications you care about. Anything else is noise.

 

[coercitiv]

Can you please take this IPC discussion elsewhere? At this point it's completely off-topic.

 

[Vattila]

Can you please take this IPC discussion elsewhere? At this point it's completely off-topic.

Probably a good idea, as some very arrogant posters have their own idea and definitions that are far from the textbook definition. Which would be somewhat acceptable, if they were clear about that. But refuting textbook definitions is annoying, as it probably confuses readers that do not know these definitions.

I like it when these forums are educational as well as informative (and entertaining), so let me round up the IPC discussion by sharing the actual textbook definition, so that any readers not familiar with these terms are not mislead.

Mopetar had it largely right in his well-written post:

The program has some number and type of instructions, the IPC (as a general number, but it could be more specific for a given program) will indicate the number of cycles it will take that processor to execute the program, and the clock speed will let you determine the amount of time it will take for a CPU to complete those cycles.

To be more specific, the number of instructions executed by the program has to be multiplied by the CPI (i.e. clock cycles per instruction) to give you the total number of clock cycles a particular program takes to run. Then multiplying by the cycle time (which is the same as dividing by the frequency) will give you the total time it takes to run the program. This is the first formula on the inside of the front cover of the computer architecture bible ("Computer Architecture — A Quantitative Approach" by Hennessy and Patterson):

CPU time = Instruction count × Clock cycles per instruction × Clock cycle time

As regards IPC, page 42 of the same textbook states: "Designers sometimes also use instructions per clock (IPC), which is the inverse of CPI".

To make it clear, the inverse is calculated as follows: IPC = 1 / CPI

So, with the definitions settled, it should now be clear that memory hierarchy, cache design and size — as well as all other architectural features of a CPU that has a bearing on the total number of cycles it takes a particular program to run — all influence the effective IPC for that particular program. It should also be clear that IPC is highly workload dependent.

 

[FangBLade]

Talking about IPC is funny because 5800x-3d have lower performance in cinebench than regular version, while much higer performance in gaming, in couple of games like MSFS significantly better even than Intel best model.