Discussion Intel current and future Lakes & Rapids thread

[igor_kavinski]

do you really think it's an advantage that a 1P4E 13th gen ADL CPU with 1.6/1.2 GHz base frequencies is truly ADL versus a 4C/8T Zen2 with a 2.4 GHz base frequency? I would take the Zen2 every time, despite it being, "old". Maybe you would prefer the 1P4E ADL, that would be your preference but that's all it is.

Wanted to say something similar but you just said it better. Thanks!

I would take 4C/8T fat cores over 1 or 2 fat cores plus however many dinky Intel cores any day of the week because the fat cores can be tuned to run slower/cooler if needed but good luck trying to extract more performance out of the E-cores. I might consider them seriously only if a mobile CPU has 16 or 32 E-cores, then maybe I could live with 2 P cores. Seeing that many cores in Task Manager is gonna make me giddy!

TBH, the whole point of weak cores was that you could have a lot of them. So far, that's not what we see in mobile CPUs so booooooooooooo!!!!!

Heck, give me 128 Larrabee Pentium cores on 5nm!

I wonder if any time Pat is being a nuisance to his wife, she just mentions Larrabee and he leaves her alone and goes and cries himself to sleep

 

[Hitman928]

Oh no I agree. But at the least the difference is more clear than a single digit in the middle of the naming scheme from AMD. And again, I didn't like what Intel did with ICL and CML as well. I have repeated that numerous times. Which is why I think Ultra is better.

Nope, it makes no difference when the consumer has no clue what the naming scheme is. Ultra doesn't really help either as it does give the impression of the superior product but superior how? Most likely MTL won't have superior top end performance so then the meaning is still obscure and it also gives no indication that it's a new architecture. Ultra can, and many times has, just meant the same CPU with a higher frequency or more cores. Ultra is just more marketing obfuscation to get you to buy the more expensive product.

You don't even need to say that. No one really knows how much better a 9th gen i7 is better than a 9th gen i5. But when people say I got a 9th gen i5 in my laptop, it's much more clear then when people might say there's a 7000 series ryzen 5 in my system. Most people just ignoring the last couple monikers on the processor label. It's just how it is.

Apparently I did. People don't talk in 9th gen, 10th gen terms. They're lucky if they understand i9 > i7 or Ryzen 5 < Ryzen 7. Even then, they have no clue that some i5 n600s are faster than i7 m700s or that some i3 n300s are no better than i3m300s. It's all a mystery to the average consumer. If you can't clearly decipher what is what in my examples, that is the same thing consumers see when they go to buy a laptop.

Yes, I agree. But one way causes more harm to the non-hardware inclined person than the other- AMD's method. It's only in theory where everyone understands everything hardware related that AMD's method is better. But that's not the world we live in.

What is this harm you speak of? Do you think AMD is selling 7020s CPUs for the same price as 7040 CPUs?

You are comparing this with different clock speeds too. RPL's advantage only comes from much higher Fmax, and a small IPC uplift in games which is like ~5%, or even less IIRC. An i5, even if it used RPL silicon, would have had an artificial frequency limit in order to ensure higher skus sell. Realistically, the only advantage an i5 using RPL silicon would have gotten is the small IPC benefit from the L2 cache uplift, as well as perhaps slightly better perf/watt from binning.

No, the greater cache and improved efficiency (or clock scaling at the upper end) allowed RPL like 15% improvement in games, I literally just posted a chart showing that. It also had higher official memory speed support which is a bigger deal in laptops.

ADL's 1+4's configuration isn't even marketed as part of 12th gen in mobile lol, it's literally marketed as a Pentium CPU. Has a whole different naming scheme too, it's called the 8505. Kinda just blows up your whole argument there.

Say hello to the 13th gen i3-1305u with a 1P4E configuration and 1.6/1.2 GHz base clock. This and the 1315u are what the Ryzen 7020 SKUs with 4C/8T are positioned against. To me, selling that 1305u as an i3 is way worse than AMD selling Zen 2 as part of their 7000 lineup but that is my take on it. If you disagree, that's fine, but there's nothing of an argument there outside of opinion and to critique one or the other (but not both) for how they handle it is silly.

And I have no problem with Zen 2 being sold, just don't call it part of the 8000 series. Or at least, make the generation number the 2nd digit, rather than the tier being the second digit. That would have made infinitely more sense. It still wouldn't have been as good as Intel making the generation number the 1rst digits, but at least it would be less misleading than what AMD is doing now.
Regardless, 12th gen's features and performance variety is just way more narrow than the perf and feature difference you will see in the 8000 series.

You believing it is better to have it as 2nd digit or is more misleading is your unprovable opinion or preference as again, any argument one way can be made in the opposite direction.

 

[TESKATLIPOKA]

I would take 4C/8T fat cores over 1 or 2 fat cores plus however many dinky Intel cores any day of the week because the fat cores can be tuned to run slower/cooler if needed but good luck trying to extract more performance out of the E-cores.

But in this case we are talking about Raptor lake vs Mendocino, which has Zen2 cores. E-cores have similar IPC.

If we compare i3 1305U vs 7520U, then Mendocino will be faster in MT, but slower in ST.

edit: I just checked and this 7520U cost the same as i3 1315U(2P+4E) in the same HP laptop.

 

[Geddagod]

Ultra doesn't really help either as it does give the impression of the superior product but superior how? Most likely MTL won't have superior top end performance so then the meaning is still obscure and it also gives no indication that it's a new architecture.

It likely will have superior top end performance, in the segment it's competing in. MTL isn't likely to be replacing RPL-HX systems.
The only thing it likely won't be superior in is ST max, by like ~5%. Should honestly be close to margin of error stuff. Considering everything else is supposed to be better, saying MTL isn't a superior product vs RPL, again, just sounds disingenuous to try to get some other random point across.

Ultra can, and many times has, just meant the same CPU with a higher frequency or more cores. Ultra is just more marketing obfuscation to get you to buy the more expensive product.

A more expensive product... that is newer and better? What??

Apparently I did. People don't talk in 9th gen, 10th gen terms.

They do. That's literally what people say. I have a 9th gen i5. I have a 12th gen i7.

They're lucky if they understand i9 > i7 or Ryzen 5 < Ryzen 7.

This is such a reach. If you think the general population doesn't know 9 is greater than 7, then you also must believe most people haven't passed like 3rd grade lol.

If you can't clearly decipher what is what in my examples, that is the same thing consumers see when they go to buy a laptop.

No. People who recommend a 12th gen i9 will have a better idea on what they are buying than people buying a 8000 series ryzen 5. All the differences (TDP, binning) that can apply to the 12th gen i9 apply to the Ryzen counterpart, while also having the added caveat of all the different generations (Zen 4-2) also counting for differences.

What is this harm you speak of? Do you think AMD is selling 7020s CPUs for the same price as 7040 CPUs?

The harm is people knows less abt what they are buying. Not that they knew everything before this new naming scheme, but they did know more.

No, the greater cache and improved efficiency (or clock scaling at the upper end) allowed RPL like 15% improvement in games, I literally just posted a chart showing that. It also had higher official memory speed support which is a bigger deal in laptops.

The chart is irrelevant because most of that gain came from the clock scaling- as you said in the upper end. Even if the i5 used RPL silicon, that clock scaling wouldn't have been relevant because Intel would have artificially limited frequency regardless to ensure that the upper tier RPL silicon sells.

Say hello to the 13th gen i3-1305u with a 1P4E configuration and 1.6/1.2 GHz base clock.

Oh dang I was wrong. That's ma fault.

To me, selling that 1305u as an i3 is way worse than AMD selling Zen 2 as part of their 7000 lineup but that is my take on it.

Except people expecting an 1305u will get what they should be expecting (at least compared to buyers of AMD). Should there be more cores in that segment? Sure. But it is a low end GLC (well marketed as RPL prob lmao) product. No one is going to be expecting ARL i3 performance in this product.
AMD selling Zen 2 as their 7000 series is worse, because people could also associate the 7000 series as having Zen 3 or Zen 4 cores/features. When they buy a 7320u, they might expect a 7340 level of performance because that's what their friends got when they said, wow my brand new Ryzen 3 has amazing performance.
This isn't opinion based.

If you disagree, that's fine, but there's nothing of an argument there outside of opinion and to critique one or the other (but not both) for how they handle it is silly.

It's not an opinion that consumers know products by the generation and tier. Intel still has generation and tier first. AMD hides the generation, instead putting the year in front. That is more misleading for general consumers than what Intel is doing.

You believing it is better to have it as 2nd digit or is more misleading is your unprovable opinion or preference as again, any argument one way can be made in the opposite direction.

Unprovable opinion or preference? Are you seriously telling me that consumers are more likely to look at the 3rd digit than the 1rst digit in AMD's new naming scheme? It might be unprovable, since I doubt there are any studies on this lmao, but I think it should be obvious consumers are drastically more likely to just pay attention to the "generation" and "tier" digits, which have been the 1rst two digits for a while now. The problem is that AMD is replacing the "generation" digit with a new "generation" meaning- the year- rather than what "architecture generation" the product is from. Misleading.

 

[Abwx]

Except people expecting an 1305u will get what they should be expecting (at least compared to buyers of AMD).

3500 pts in GB5 for the 1305U and about 3800 pts for a 7520U, that s what you get from both sides, notice that the 1305U turbo up to 55W, so dunno what is the precision of thoses numbers, the AMD part seems to be within 20-25W in the NBC test.

 

[Hitman928]

It likely will have superior top end performance, in the segment it's competing in. MTL isn't likely to be replacing RPL-HX systems.
The only thing it likely won't be superior in is ST max, by like ~5%. Should honestly be close to margin of error stuff. Considering everything else is supposed to be better, saying MTL isn't a superior product vs RPL, again, just sounds disingenuous to try to get some other random point across.

A more expensive product... that is newer and better? What??

That's my point, Intel's Ultra CPUs will be slower than RPL-HX. So their Ultra line isn't "ultra" performance, there will be a tier above it without an ultra or something moniker for top performance. That's not confusing, right? So the consumer is paying for something Ultra but what that is isn't clear.

They do. That's literally what people say. I have a 9th gen i5. I have a 12th gen i7.

I have never talked to an average consumer who knows nothing about computers who knows gen generation core they have in their laptop. Just like I've never met the average person who knows what size engine they have, they just know it's a, e.g., v6. I have friends who do, but they are a little more tech savvy.

This is such a reach. If you think the general population doesn't know 9 is greater than 7, then you also must believe most people haven't passed like 3rd grade lol.

It was a tongue in cheek comment.

No. People who recommend a 12th gen i9 will have a better idea on what they are buying than people buying a 8000 series ryzen 5. All the differences (TDP, binning) that can apply to the 12th gen i9 apply to the Ryzen counterpart, while also having the added caveat of all the different generations (Zen 4-2) also counting for differences.

But the average user doesn't care about any of that, you said so yourself. . .

The harm is people knows less abt what they are buying. Not that they knew everything before this new naming scheme, but they did know more.

You keep arguing both sides, that the average consumer doesn't care about the architecture or any technical details (or they can google it if they care) but Intel's scheme is better because these details are (supposedly) clearer. Which one is it? And there is no harm in AMD's marketing tactic because they are charging less for the lower performing parts. Maybe they thought they were getting a deal when really they were getting what they paid for is not harmful in anyway.

The chart is irrelevant because most of that gain came from the clock scaling- as you said in the upper end. Even if the i5 used RPL silicon, that clock scaling wouldn't have been relevant because Intel would have artificially limited frequency regardless to ensure that the upper tier RPL silicon sells.

And yet, that's not what happened, look at the chart again, there are i5s on that chart too.

Except people expecting an 1305u will get what they should be expecting (at least compared to buyers of AMD). Should there be more cores in that segment? Sure. But it is a low end GLC (well marketed as RPL prob lmao) product. No one is going to be expecting ARL i3 performance in this product.

So they are getting what they should expect (which is the "correct" architecture, something you previously said they don't know or care about), they just shouldn't expect i3 performance from an i3 part?

AMD selling Zen 2 as their 7000 series is worse, because people could also associate the 7000 series as having Zen 3 or Zen 4 cores/features. When they buy a 7320u, they might expect a 7340 level of performance because that's what their friends got when they said, wow my brand new Ryzen 3 has amazing performance.

First off, no one is going to say that. Second, no one is buying individual mobile CPUs, this isn't DIY desktop. No one is cross shopping Zen2 and Zen 4 SKUs, even at the Ryzen 3 level as they will be in different tier laptops. Last, it's not any difference than the span between i5s or i7s. There is a huge disparity between the list of 13th gen i7 SKUs so once again, Intel is just as guilty here as AMD, just in a different way which again gets back to opinion as to which one is preferred or the "better" way.

It's not an opinion that consumers know products by the generation and tier. Intel still has generation and tier first.

Except when they don't, but those don't count because reasons.

AMD hides the generation

By clearly marking it in the model name, the jerks.

instead putting the year in front. That is more misleading for general consumers than what Intel is doing.

Unprovable opinion or preference? Are you seriously telling me that consumers are more likely to look at the 3rd digit than the 1rst digit in AMD's new naming scheme? It might be unprovable, since I doubt there are any studies on this lmao, but I think it should be obvious consumers are drastically more likely to just pay attention to the "generation" and "tier" digits, which have been the 1rst two digits for a while now. The problem is that AMD is replacing the "generation" digit with a new "generation" meaning- the year- rather than what "architecture generation" the product is from. Misleading.

If you don't believe it is opinion based, what is then your theory on why Intel's naming methodology is better and what evidence can you provide to support that theory? I'm not talking evidence about the SKU lineup, but evidence that proves why one is better than the other.

 

[Geddagod]

3500 pts in GB5 for the 1305U and about 3800 pts for a 7520U, that s what you get from both sides, notice that the 1305U turbo up to 55W, so dunno what is the precision of thoses numbers, the AMD part seems to be within 20-25W in the NBC test.

I mean I literally just said this:

And I have no problem with Zen 2 being sold,

The problem is calling it an 8000 series CPU. People can easily be confused between this and and a 8040 or 8030 CPU, because people don't pay attention to like anything but the generation number and tier.

 

[Geddagod]

That's my point, Intel's Ultra CPUs will be slower than RPL-HX. So their Ultra line isn't "ultra" performance, there will be a tier above it without an ultra or something moniker for top performance. That's not confusing, right? So the consumer is paying for something Ultra but what that is isn't clear.

People are paying for top performance with Ultra. Because in the laptops these CPUs are going too, those are going to be the top performers in most categories. You aren't stuffing 13900HXs into thin and lights.

I have never talked to an average consumer who knows nothing about computers who knows gen generation core they have in their laptop.

They could just as easily say "the newest i5" and mess up due to AMD's new marketing segmentation lmao

But the average user doesn't care about any of that, you said so yourself. . .

Not the specific core architecture. But what it does have a larger impact is the features/performance the end product will have than just Intel's regular generational product differentiation, and creates more confusion for the end user. A 13th gen i5 has only so much differentiation, unlike a 8000 series Ryzen 5.

And yet, that's not what happened, look at the chart again, there are i5s on that chart too.

What? That's exactly what happened. The i5's on that chart isn't the 13400f, the part that mixed ADL and RPL silicon. Your chart proves or shows nothing.

So they are getting what they should expect (which is the "correct" architecture, something you previously said they don't know or care about), they just shouldn't expect i3 performance from an i3 part?

What I've been saying for multiple messages now. With a 8000 series, Ryzen 5 part, people can expect anything from a 8540 to 8520 series performance, because people don't look at the 3rd digit in the naming scheme, just generation and tier. With 13th gen i3, the product differentiation is way less.

First off, no one is going to say that.

Everyone says that. Tier and generation are the two biggest things normies know.

Second, no one is buying individual mobile CPUs, this isn't DIY desktop. No one is cross shopping Zen2 and Zen 4 SKUs, even at the Ryzen 3 level as they will be in different tier laptops. Last, it's not any difference than the span between i5s or i7s. There is a huge disparity between the list of 13th gen i7 SKUs so once again, Intel is just as guilty here as AMD, just in a different way which again gets back to opinion as to which one is preferred or the "better" way.

Oh but one company is way more misleading than the other. AMD. Idk why you keep on bringing up the "both sides" do this shtick. I have acknowledged that numerous time. One side has it worse though- AMD.

Except when they don't, but those don't count because reasons.

For ICL and CML? Yawn. As said like 3 times before, they made it more obvious than AMD does, and I still claimed I didn't like it when they did that either. Which is why Intel moving to Ultra is better.

By clearly marking it in the model name, the jerks.

Something which most general consumers aren't going to pay attention too.

what is then your theory on why Intel's naming methodology is better and what evidence can you provide to support that theory?

Literally what I have been saying in the past couple replies to you.

I'm not talking evidence about the SKU lineup, but evidence that proves why one is better than the other.

I literally have explained this numerous times to you. Intel's naming scheme is less misleading, because the generation and tier models have less differentiation than they do with AMD.

Idk, it really sounds like we keep on going in circles here, where you just ignore reality. Like this entire past message I'm replying too, hasn't had any actual rebuttels smh.

It doesn't look like either of us are getting anywhere with this anymore, so idk, lets just let this drop.

 

[Abwx]

I mean I literally just said this:

The problem is calling it an 8000 series CPU. People can easily be confused between this and and a 8040 or 8030 CPU, because people don't pay attention to like anything but the generation number and tier.

If you can find a 7040/8040 laptop at about the same price, even remotely, as a 7020/8020 one please do tell us as fast as you could, i mean, the 7020/8020 are for the 400-500$ segment, there cant be much confusion when buying such an item....


That being said the time of single core CPUs is long gone, there s currently a lot of differentiation even within a same line, there s 6C and 8C Zen 4 and 3 as well as 4C Zen 4, 3 and 2, that s a lot of products, at least with AMD scheme you know what is what for any of theses APU, that s better than obscure numbering schemes that tell nothing within the number itself.

 

[Hitman928]

People are paying for top performance with Ultra. Because in the laptops these CPUs are going too, those are going to be the top performers in most categories. You aren't stuffing 13900HXs into thin and lights.

They could just as easily say "the newest i5" and mess up due to AMD's new marketing segmentation lmao

Not the specific core architecture. But what it does have a larger impact is the features/performance the end product will have than just Intel's regular generational product differentiation, and creates more confusion for the end user. A 13th gen i5 has only so much differentiation, unlike a 8000 series Ryzen 5.

What? That's exactly what happened. The i5's on that chart isn't the 13400f, the part that mixed ADL and RPL silicon. Your chart proves or shows nothing.

What I've been saying for multiple messages now. With a 8000 series, Ryzen 5 part, people can expect anything from a 8540 to 8520 series performance, because people don't look at the 3rd digit in the naming scheme, just generation and tier. With 13th gen i3, the product differentiation is way less.

Everyone says that. Tier and generation are the two biggest things normies know.

Oh but one company is way more misleading than the other. AMD. Idk why you keep on bringing up the "both sides" do this shtick. I have acknowledged that numerous time. One side has it worse though- AMD.

For ICL and CML? Yawn. As said like 3 times before, they made it more obvious than AMD does, and I still claimed I didn't like it when they did that either. Which is why Intel moving to Ultra is better.

Something which most general consumers aren't going to pay attention too.

Literally what I have been saying in the past couple replies to you.

I literally have explained this numerous times to you. Intel's naming scheme is less misleading, because the generation and tier models have less differentiation than they do with AMD.

Idk, it really sounds like we keep on going in circles here, where you just ignore reality. Like this entire past message I'm replying too, hasn't had any actual rebuttels smh.

It doesn't look like either of us are getting anywhere with this anymore, so idk, lets just let this drop.

You provided zero evidence. That is your perception of the modeling system and opinion of it. I am fine if you prefer Intel's scheme but you are posting like it is verifiably superior which is silly. Intel violates your own criteria multiple times and without any transparency but it just gets brushed aside (CFL, CML/ICL. RPL/ADL, RPL/MTL) because Intel doesn't codify it into their system. You say I am ignoring reality when all you've done is give an opinion that many people in this thread obviously disagree with and you can not give me any verifiable proof or evidence that your opinion is more than just that. If you have anything like that, I'm happy to change my mind and declare Intel's system the superior branding methodology. Otherwise, have a great day.

Edit: Just as a note, we don't even know for sure yet if Zen2 will be a part of the 8000 series. AMD has only announced Zen 4 in the 8000 series. I do expect additinoal SKUs so it may end up being an 8320 or something, but AMD hasn't announced anything like that yet.

 

[Geddagod]

If you can find a 7040/8040 laptop at about the same price, even remotely, as a 7020/8020 one please do tell us as fast as you could, i mean, the 7020/8020 are for the 400-500$ segment, there cant be much confusion when buying such an item....


Thatg being said the time of single core CPUs is long gone, there s currently a lot of differentiation even within a same line, there s 6C and 8C Zen 4 and 3 as well as 4C Zen 4, 3 and 2, that s a lot of products, at least with AMD scheme you know what is what for any of theses APU, that s better than obscure numbering schemes that tell nothing within the number itself.

Very classy of you to get in the last word when I legit just said, Imma drop this topic, cuz you know I have stuff to do... like study for finals... or you know, not be online 24/7 lol. Very classy.

 

[Abwx]

Very classy of you to get in the last word when I legit just said, Imma drop this topic, cuz you know I have stuff to do... like study for finals... or you know, not be online 24/7 lol. Very classy.

That s just discussion, dunno why you re taking it like this, as to drop the subject that was actually brought by Intel themselves, why not, unfortunately we dont have much communication these times other than this recent smear campaign, i would hope that they released something about MTL instead...

 

[Geddagod]

That s just discussion, dunno why you re taking it like this, as to drop the subject that was actually brought by Intel themselves,

Nah, I'm saying I'm dropping it, and then you reply to me... right after I said I was dropping it? Like cmon bruh : |

i would hope that they released something about MTL instead...

I mean considering MTL isn't out yet... but whatever, I gotta go study.

 

[Timorous]

Nah, I'm saying I'm dropping it, and then you reply to me... right after I said I was dropping it? Like cmon bruh : |

I mean considering MTL isn't out yet... but whatever, I gotta go study.

You could just not respond...

 

[nicalandia]

Just leaving this out here...




I've been saying this for a while now. These E cores would be pack in clusters(Quad) each of them with only 4MiB of L2 and a smaller ring bus for inter comunication. L3 Being less than 1MiB per core will really hurt the.

 

[gdansk]

Will it hurt GB? Yes but that doesn't mean much. These aren't consumer parts, these are basically micro vm/container/serverless cores. Shouldn't be communicating between clusters much.

 

[Saylick]

Just leaving this out here...

View attachment 90016


I've been saying this for a while now. These E cores would be pack in clusters(Quad) each of them with only 4MiB of L2 and a smaller ring bus for inter comunication. L3 Being less than 1MiB per core will really hurt the.

Ehh, I'm not sure how much intercommunication is needed for a product like this and what it is intended to do, i.e. offer strong INT performance per dollar for mostly independent tasks. As some would say on these forums, it's a bunch of "poverty cores" for cheap.

Edit: Dang, @gdansk beat me to the punch lol

 

[dullard]

Is that limited to the Linux maximum number of cores of 256? Or does the fact that it is two 144 core CPUs get around that limit? Not that 288 is much above 256, but I was just curious.

 

[gdansk]

Is that limited to the Linux maximum number of cores of 256? Or does the fact that it is two 144 core CPUs get around that limit? Not that 288 is much above 256, but I was just curious.

For x86/64 Linux: when CPUMASK_OFFSTACK is enabled, the maximum supported CPU count is 8192, otherwise the maximum value is 512.

Shouldn't be a problem yet even if it was a single CPU using common server kernels.

 

[TESKATLIPOKA]

3500 pts in GB5 for the 1305U and about 3800 pts for a 7520U, that s what you get from both sides, notice that the 1305U turbo up to 55W, so dunno what is the precision of thoses numbers, the AMD part seems to be within 20-25W in the NBC test.

From that single review on NBC, 7520U has 30W PL2 and after that It drops to 25W in Prime95 Plus Furmark.
There is nothing about 1305U, but that 55W is max PL1 that can be set, doesn't mean It has to be set that high.

 

[lightisgood]

As far as Gaudi3, Intel is claiming 1.5x the memory and 2x the compute.

You are simply wrong.
Gaudi3 will provide 4x AI power (BF16).

P.S.
MI300X has 1.3x higher peak BF16 of H100.
This is a lack, isn't it?

I'll leave it at that, if you feel the need to continue trying to spin the narrative against MI300, I suggest you go do it in the appropriate thread in the graphics section where much of this has already been discussed.

I would like to see Gaudi3(2)'s perf rather than MI300's .
Is this graphics?

 

[Hitman928]

You are simply wrong.
Gaudi3 will provide 4x AI power (BF16).

Intel is the one who claims 2x compute with Gaudi3, not me. But you are correct, if only using BF16, they are claiming a 4x speedup. I believe modern models are pushing towards 8 bit or use mixed precision so I'm not sure how much of an effect that will have on overall training performance.

Spoiler

P.S.
MI300X has 1.3x higher peak BF16 of H100.
This is a lack, isn't it?

So AMD will have a higher peak BF16 than H100 and H200, why is that lacking? BTW, has Intel published any of Gaudi's peak numbers anywhere? I have looked twice and can't seem to find them whereas AMD and NV make theirs very easy to find.

I would like to see Gaudi3(2)'s perf rather than MI300's .
Is this graphics?

What? If you are asking for a chart of Gaudi3/2's performance, I posted Gaudi2 vs H100 already. There's no performance charts for Gaudi3 yet.

One interesting note is that it looks like Gaudi will go from the highest memory capacity with Gaudi2 out of all the players to the lowest with Gaudi 3.

There was a mistake on the original slide that said 1.5x HBM Capacity. That is now “BW” or Bandwidth. 144GB (96GB x 1.5x) seemed strange in the eight HBM packages shown. Our best guess is that Gaudi3 might be targeting 8x 16GB stacks for 128GB

Intel still needs to confirm the memory capacity, but if correct, this will put it behind H200 (141 GB) and well behind MI300X (192 GB).

 

[jpiniero]

Hadn't seen this before... Emerald Rapids also has an monolithic LCC die. It has up to 20 cores.

 

[JoeRambo]

5th Gen Intel Xeon Processors Emerald Rapids Resets Servers by Intel

We delve into the 5th Gen Intel Xeon Scalable Emerald Rapids generation and talk both the performance and platform gains around the new chips

www.servethehome.com

Some really nice gains there. Seems like proper memory subsystem can go a long way




Some real great engineering. Average to access 160MB in SNC2 mode ( half of chip's 320MB ) is 25ns, when previuosly 26MB took 22ns in SNC4 and full chip L3 cache was 34ns previuosly, when now crossing to different SNC2 domain is just 35.
Competitive results in Phoronix test workloads where this matters, Intel finally can reap benefits of having "almost" monolithic server CPU and it surely will generate them extra sales in those lovely "per core licensing" realms.

Kudos to Intel's engineering team, that's a genuine achievement on 10nm process.

 

[jpiniero]

Intel Xeon Platinum 8592+ "Emerald Rapids" Linux Benchmarks Review - Phoronix

www.phoronix.com

Phoronix has some benchmarks too. Decent improvment... The top model is close to the 64 core Genoa but still chugs way more power.