Intel Launches 3rd Gen Xeon Scalable "Ice Lake" CPUs With Up To 40 Cores

[wizard69]

Instead, the message they'll likely hear is that you're just another Mac addict who will buy Apple at any price (and any CPU technology).

The problem here is that the price on the M1 Air is extremely competitive for what you get. It is the first time in a long time where Apple has one of the best values in computing for sale. The problem here is that somebody needs to lead the way and show the market that ARM is viable in this sort of application. What M1 demonstraits is that even a cell phone oriented chip can out perform some of the best that Intel has and do it coolly. So if you look at what Apple is doing from the bottom up and what the server manufactures could do from the top down, you have to wonder why people even bother with x86.

So if they don't get the message then there is a good likelihood that Intel will be on very hard times soon. Maybe they still have some magic left and can get power usage down while increasing the performance of their chips. The question is why does anyone as a consumer even want to deal with the likes of Intel anymore?

[coder]

I would definitely be in the market for a mobile workstation class ARM Linux laptop. I've never seen such a thing though, unfortunately. Maybe it will happen after Apple proves it works?

The issue isn't "proving it works", technically. I think we know it's technically possible. The real blocker is proving there's a big enough market for it to be financially viable. And Apple isn't really going to prove that, because they're not going to give their customers the choice. People will keep buying their products because they're Macs (aside from the rare exception, like @wizard69) and nobody can really say otherwise.

The risk is particularly high, when you consider the only fast (non-Apple) ARM chips are currently either for phones & light-weight laptops or big cloud servers. So, you're basically proposing someone like Qualcomm sink hundreds of $M into a high-performance SoC suitable for mobile use, without a clear market and when their existing Snapdragon 8cx and 7cx laptop chips have been seeing lackluster uptake (or so that's the impression I've gotten from the lack of any successors being announced).

IMO, the most likely to do this (after Apple) is probably the Chinese. Huawei already introduced an ARM-based desktop PC about a year ago.

[coder]

They sure did, but the downside was very little support for any PCIe AIC's. The few that were supported, appeared to have been qualified by Raptor themselves, rather than by the card manufacturer, so it seems there was little to no 3rd party validation going on there, at least at the time.

But presumably all POWER-based servers at that time were also supporting PCIe 4.0. I only picked on Raptor, because I don't follow the market for POWER-based servers, so I can't definitely cite anyone else.

The point is that PCIe 4.0-based systems have been available for AIC-makers to test & debug their wares since like 2018. So, I don't get why you're painting it as some kind of risky proposition.

IMO, what's actually risky is PCIe 5.0 and CXL, at least the first of which Intel is promising on Sapphire Rapids, shipping late this year or early next.

Yup, with Big Cloud as the primary market, the only PCIe4 validation occurring was for the short list of cards specific to those cloud implementations. As much as I like Ampere and the idea of an ARM server, it's still very much a niche market that gets deployed in specific use cases, or for specific cloud implementations, and does not attract broad platform validation from the industry.

Anyone building a PCIe 4.0 card can buy an off-the-shelf Ampere Altra server from Gigabyte (and IIRC, at least one other manufacturer). Or a POWER 9 system. Or Ryzen. So, that's 3 ISA, and a bunch of different motherboard brands. Plus, Intel and its partners should've learned some valuable lessons from the issues with Comet Lake.

Intel has been absent. It's hard to validate a platform that exists only on paper. Intel is in catch-up mode. How much ground they still have to cover in terms of PCIe4 validation remains to be seen.

Intel has been sampling Ice Lake server CPUs to partners for at least 2 years!

[coder]

I look at this release and wonder what if it had come out on time. I think the storylines would be much, much different.

Depends on what you mean by "on time" (i.e. which Intel roadmap revision you're going by). If it had gone up against the first-gen Epyc, AMD would've gotten massacred. Even at the time, I recall that its main selling point against Intel was price.

But the 7002-series holds its own. Indeed, 7003 isn't much faster than 7002, in a lot of cases. So, even launching it 1.5 years ago wouldn't have left Intel in a much better position than today. The biggest difference is probably the amount of market traction AMD has gotten. In the first generation of EPYC, they had only a few design wins. During Epyc's second gen, AMD (re-)established itself as a real player in the server market.

And in the interim of getting Ice lake out the door, I am sure they were working on their own internal enhancements, enhancements we will see (hopefuily) over the next few years.

Not sure what you mean by that, but there hasn't been any indication of large-scale design changes to Ice Lake SP, since its planned release. Remember that they did ship Ice Lake laptop CPUs at the end of 2019, so this isn't the first time the world has seen the Sunny Cove cores.

Now they are getting their fab situation under control, makes me wonder how backed up the product pipeline is and how performant they will be.

Tiger Lake laptop chips are performing much better than Ice Lake did (both in IPC and clockspeed), and they're made on a newer iteration of Intel's 10 nm node than either Ice Lake uses. Moreover, its Willow Cove cores and process node are slated to be used in Sapphire Rapids, coming in... let's say about a year (allegedly). Therefore, it does indeed look like some more competitive products are in the pipeline.

As for the threat to AMD, Tiger Lake's Willow Cove cores can't match Ryzen 5000's Zen3 cores, in IPC. So, if AMD can somehow do a refresh of EPYC 7003 with a 7 mn I/O die and some power optimizations, maybe they still hold a comfortable lead. If not, we might have to wait until Genoa (Epyc 7004) for them to regain their leadership of the x86 server performance race.

I am not for any vendor, I am for performance value.

Well, then I hope you're happy to see Intel cut prices in its server chip lineup, for the first time in recent memory! And that, in the midst of a global CPU supply crunch!

[coder]

I love how you legacy x86 guys are still trying to defend this increasingly moribund and aged architecture.

Few are more ready to see the back of x86 than I am. I bought a 1st gen Raspberry Pi just a few months after launch, plus a v3 (and an ODROID-N2) since then. Within the past decade, I even thought ARM would already be dominant, by now.

Ampere will release this year a single socket 128 core Altra. So Milan and Ice Lake are already leapfrogged by ARM.

Don't count your chickens... Let's wait and see how well it scales and what its pricing & power consumption look like.

But right now if I wanted an 800 core Supercomputer I would drop $40,500 USD retail on ten 80 core Ampere Altras. To get a 800 core Intel Ice Lake Super I'd have to drop
$162,000 USD for TWENTY 40 core Ice Lakes.

I can't fathom why you're comparing raw core-counts. Altra has to use more cores specifically to compensate for the fact that they're each slower.

But, if you're going to go by raw core counts, then Anandtech actually looked at the per-core performance under load[1], and found that:

• an Ice Lake core is 19% faster at integer workloads
• an Ice Lake core is 73% faster at floating-point workloads

Now, most people aren't going to run them that way. Most people will run them both as 80-thread CPUs, where Ice Lake's advantage disappears because it has 2 threads sharing each core (albeit with higher net utilization of the core's resources to partially compensate for that). So, Anadtech's data[1] show that with 80 threads on each:

• an Altra thread is 55% faster at integer workloads
• an Altra thread is 18% faster at floating point workloads

So, that means you'd need 8.5 Altra CPUs to achieve the same floating point throughput as 10 Ice Lake SP. Not 5, as your facile analysis would suggest. And that also doesn't take into account whether the workload is optimized for AVX2 or AVX-512. In such cases, Altra is at a disadvantage, having only 128-bit vector support.

And when you look at it like that, and also consider that servers have costs above and beyond just the CPUs, it's not the completely lopsided picture you painted.

The other silly thing about your strawman is to put up Ice Lake as the standard-bearer of x86, when it's clearly AMD's Milan.

But. I think you're just trying to find ways to paint ARM in the best light, for whatever reason. I have to wonder if you have some vested interest in ARM, like certain others seem to, or maybe you're just an ARM fanboy?

References:

1. https://www.anandtech.com/show/16594...lable-review/8

[coder]

I just hope that one of the big manufactures, probably Lenovo, goes all in on a true ARM based workstation class laptop.

They make computers, not chips. There exists no suitable SoC for them to use, even if they wanted to build such a machine!

[coder]

The problem here is that the price on the M1 Air is extremely competitive for what you get. It is the first time in a long time where Apple has one of the best values in computing for sale.

IMO, that's just something Apple is doing to build the market for ARM-native software, in order for software developers have enough incentive to port their code. I 100% guarantee that Apple will be back to their old pricing structure (if not worse), once the market has transitioned.

What M1 demonstraits is that even a cell phone oriented chip can out perform some of the best that Intel has and do it coolly.

But they already did that, without you having to buy into it!

So if you look at what Apple is doing from the bottom up and what the server manufactures could do from the top down, you have to wonder why people even bother with x86.

These transitions take time, and ARM is still just on the cusp of breaking through. Nvidia's acquisition will surely have created additional headwind, for some.

I hope everyone here is aware that AMD already sold an ARM-based server chip[1], but apparently got ahead of the market. And the last thing Jim Keller did at AMD was to design an in-house ARM core[2], at the same time they were working on Zen. Apparently, they had only enough resources to bring one to market and they wisely went with the one that had the (much) bigger market, at the time.

References:

1. https://www.amd.com/en/amd-opteron-a1100
2. https://en.wikipedia.org/wiki/AMD_K12

[Jumbotron]

Few are more ready to see the back of x86 than I am. I bought a 1st gen Raspberry Pi just a few months after launch, plus a v3 (and an ODROID-N2) since then. Within the past decade, I even thought ARM would already be dominant, by now.


Don't count your chickens... Let's wait and see how well it scales and what its pricing & power consumption look like.


I can't fathom why you're comparing raw core-counts. Altra has to use more cores specifically to compensate for the fact that they're each slower.

But, if you're going to go by raw core counts, then Anandtech actually looked at the per-core performance under load[1], and found that:

• an Ice Lake core is 19% faster at integer workloads
• an Ice Lake core is 73% faster at floating-point workloads

Now, most people aren't going to run them that way. Most people will run them both as 80-thread CPUs, where Ice Lake's advantage disappears because it has 2 threads sharing each core (albeit with higher net utilization of the core's resources to partially compensate for that). So, Anadtech's data[1] show that with 80 threads on each:

• an Altra thread is 55% faster at integer workloads
• an Altra thread is 18% faster at floating point workloads

So, that means you'd need 8.5 Altra CPUs to achieve the same floating point throughput as 10 Ice Lake SP. Not 5, as your facile analysis would suggest. And that also doesn't take into account whether the workload is optimized for AVX2 or AVX-512. In such cases, Altra is at a disadvantage, having only 128-bit vector support.

And when you look at it like that, and also consider that servers have costs above and beyond just the CPUs, it's not the completely lopsided picture you painted.

The other silly thing about your strawman is to put up Ice Lake as the standard-bearer of x86, when it's clearly AMD's Milan.

But. I think you're just trying to find ways to paint ARM in the best light, for whatever reason. I have to wonder if you have some vested interest in ARM, like certain others seem to, or maybe you're just an ARM fanboy?

References:

1. https://www.anandtech.com/show/16594...lable-review/8

Ahhh...see...legacy x86 thinking. You're so wrapped up in legacy x86 fanbois-ism that you think ARM server SoCs are only capable of fixed SIMD Neon instructions at 128 bit. But in fact ARM based chips like Ampere's Altra have SVE (Scalable Vector Extensions) that START at 128 bit and scale (hence the use of Scalable in SVE) to 2084 bits in 128 bit chunks. It's up to the chip maker to implement that but you are NOT limited to "just" 512 bits of legacy AVX.

And with SVE2 already here, SVE will now be able to be implemented not just for HPC or ML but EVERY use case from phones, to base stations, CV, in memory processing, in line data processing, etc.

Oh, I almost forgot, you don't need to recompile your binaries or need to even know your vector variable length.

We'll see Apple first take advantage of SVE2 probably by this time next year as they roll out the 3nm (yes...3nm and the very first to do it) version update to M1. Ampere may be rolling this out with new Altras next year as well. But SVE AND SVE2 are BOTH Scalable to 2084 bits.

And ARM chips and SoCs will NOT have to be the size of the palm of your hand in order to provide the die area needed to implement them. Pretty soon, both Intel's "Tile" based architecture and AMD's "Chiplets on SIP" will be the size of a Micro ATX motherboard themselves.

[coder]

Ahhh...see...legacy x86 thinking. You're so wrapped up in legacy x86 fanbois-ism

I already explained that I'm not. I take the fact that the best defense you can mount against having vested interests in ARM is try to tar me with the same label as veritable proof that you have some undisclosed conflict of interest.

you think ARM server SoCs are only capable of fixed SIMD Neon instructions at 128 bit.

Wow. That's the one point you pick up on? I guess it's the only one you believe you have an answer for...

But in fact ARM based chips like Ampere's Altra have SVE (Scalable Vector Extensions) that START at 128 bit and scale (hence the use of Scalable in SVE) to 2084 bits in 128 bit chunks. It's up to the chip maker to implement that but you are NOT limited to "just" 512 bits of legacy AVX.

Except, it doesn't. That's right!! The N1 core (used in the Altra) doesn't have SVE!

So, you're left with zero legs to stand on.

And for the record, I have been watching SVE with interest, since it was first announced (5 years ago). But, since you have no apparent interest in facts, I'm sure you don't care about that, either.

Next time, try putting even half the energy into getting your facts straight as you put in your ARM boosterism, and you might not come off as such an idiot fanboy. Sheesh. You almost make me feel embarrassed to be an ARM supporter.