News Intel Launches Sapphire Rapids Fourth-Gen Xeon CPUs and Ponte Vecchio Max GPU Series

[Admin]

Intel launched its Fourth-Gen Xeon Scalable Sapphire Rapids processors along with its Xeon MAX CPUs and Max GPU Series.

Intel Launches Sapphire Rapids Fourth-Gen Xeon CPUs and Ponte Vecchio Max GPU Series : Read more

 

[Giroro]

So Intel is making lower performing, less efficient, higher priced processors with confusing product segmentation and a supremely unethical upgrade model.

Let's see how that works out works out for them.

 

[Tech0000]

My take: Intel is betting on an investment in very specific accelerator silicon (as opposed to using that silicon budget to add more general cores etc) will be competitive against AMD EPYC many more core products.
The bet is that the largest investments that their customers do will be in server capacity for AI (obviously), data traffic/ encryption (also obviously) etc. etc. (read their marketing above) and not in general server performance improvements (across all work loads - even though there is some of that as well with gen 4).

No doubt AI etc. is important but it is a relatively big and risky bet, trading potential general server volume (sales that may slip to AMD with their many more core general high performing EPYC servers) vs. accelerated workload server volume.
They said they listened to their customers and this is what they said so...
We'll see it intel got the market prediction right...

It also felt like they got some of their customers to even say (in the intel's video stream today) that with the new gen4 Xeons they do not need special AI accelerators (from NVidia etc) since the new Xeons are good enough. Intel quickly jumped on that nice set up (in the video) and (re)introduced the "Democratization of AI" slogan. So in some ways it's a NVidia compete as well...

Again, We'll see...

Strategically, it feels like Intel is getting squeezed and is fighting a two front war AMD on one side and NVidia & co on the other side... not a comfortable position to be in...

 

[jkflipflop98]

So Intel is making lower performing, less efficient, higher priced processors with confusing product segmentation and a supremely unethical upgrade model.

Let's see how that works out works out for them.

Must be opposite day here on Tom's.

 

[Tech0000]

Forgot to say that intel mentioned, they are going to launch the workstation W790 series W2400 and W3400 chips on February 15th.
hopefully they will be priced at a level that we can build reasonable HEDTs out of them. W3400 looks great on paper, but priced wrong, it will be out of reach and Intel consumer HEDT will stay dead. sigh...

 

[jkflipflop98]

A 24 core 13900k is $599 at newegg. I'm not sure what you're going on about "consumer HEDT will stay dead". Looks like it's alive and kicking to me.

 

[Samlebon2306]

These chip are going to get Threadripped really hard.

 

[rluker5]

So Intel is making lower performing, less efficient, higher priced processors with confusing product segmentation and a supremely unethical upgrade model.

Let's see how that works out works out for them.

Depends on the workload.
Gaming GPUs have a hard time competing with Asics when they can mine the same crypto.

 

[truerock]

This is question that is not specific to these Intel CPUs.

Why do workstation and server CPUs run slower than desktop PC CPUs?

I googled it and all I'm getting is: more cores means more heat which means server CPUs need to run slower. But that is obviously not true.

Xeon SKU 6434 has only 8 cores and max turbo is only 4.1GHz.

The 24 (8p/16e) core Intel ‘Raptor Lake’ Core i9-13900K will run 5.8GHz.

Is there any hope of a Xeon 8-core CPU running boost 5.8GHz someday?

Thank you for any insight.

 

[Amdlova]

And Iam trying to get an 7259cl 24 cores of raw power and some optanes to play. Epyc has some Market but have amd problems. Cannot guarantees if will work right.

 

[Diogene7]

I am curious to see what will be the next gen Non Volatile Memory (NVM) options that will connect through CXL will appear on the market in a couple of years ?

And also what will be the latency because it will be usung the PCIe5 or PCIe6 bus, but Intel Optane DC PMEM were pluggable on the DRAM bus ?

Any guess ? Nantero carbon nanotube NRAM ? Everspin / Avalanche magnetic MRAM (STT-MRAM, SOT-MRAM, VG-SOT-MRAM,…) ?

 

[davebyrd]

we all know proof-reading isn't a thing at TH, but this article is shockingly poor.

 

[thestryker]

Forgot to say that intel mentioned, they are going to launch the workstation W790 series W2400 and W3400 chips on February 15th.
hopefully they will be priced at a level that we can get build reasonable HEDTs out of them. W3400 looks great on paper, but priced wrong, it will be out of reach and Intel consumer HEDT will stay dead. sigh...

The W3400 is pretty much guaranteed to be their Threadripper Pro competitor so it will very likely be equally expensive. The W2400 line is the one that has a chance to be more reasonably priced.

A 24 core 13900k is $599 at newegg. I'm not sure what you're going on about "consumer HEDT will stay dead". Looks like it's alive and kicking to me.

HEDT is not just about cores, it's about the entire platform. Consumer platforms are still PCIe lane starved and are limited in memory capacity/bandwidth.

This is question that is not specific to these Intel CPUs.

Why do workstation and server CPUs run slower than desktop PC CPUs?

I googled it and all I'm getting is: more cores means more heat which means server CPUs need to run slower. But that is obviously not true. In fact - with these Intel Xeon CPUs it seems the fewer cores the slower they run.

Thank you for any insight.

It's two things: core count and market segmentation.

core count: They are limited by TDP so there is a maximum per core worth of power they can consume. A 13900K can pull over 250W in stock configuration with 8 p-cores and 16 e-cores whereas a Xeon 6430 has 32 p-cores that have to fit into 270W (there's more to the power consumption than just this, but I figured it was an easy to understand comparison which gets the point across).

market segmentation: Just like how consumer chips clock lower the further down the stack you go the same can be said for server/workstation chips. The other part is lower power consumption parts to fit a specific segment of the market.

 

[gg83]

Can someone explain the accelerator on demand thing? Are the chips cheaper to produce with unactivated accelerator cores? How does this work? The cores have to function ahead of time right? So it's a software thing? It's all very weird to me.

 

[truerock]

The W3400 is pretty much guaranteed to be their Threadripper Pro competitor so it will very likely be equally expensive. The W2400 line is the one that has a chance to be more reasonably priced.


HEDT is not just about cores, it's about the entire platform. Consumer platforms are still PCIe lane starved and are limited in memory capacity/bandwidth.


It's two things: core count and market segmentation.

core count: They are limited by TDP so there is a maximum per core worth of power they can consume. A 13900K can pull over 250W in stock configuration with 8 p-cores and 16 e-cores whereas a Xeon 6430 has 32 p-cores that have to fit into 270W (there's more to the power consumption than just this, but I figured it was an easy to understand comparison which gets the point across).

market segmentation: Just like how consumer chips clock lower the further down the stack you go the same can be said for server/workstation chips. The other part is lower power consumption parts to fit a specific segment of the market.

Thank you for the feedback

Xeon SKU 6434 has only 8 cores and max turbo is only 4.1GHz.

The 24 (8p/16e) core Intel ‘Raptor Lake’ Core i9-13900K will run 5.8GHz.

Your "market segmentation" comment. Are you saying Intel somehow physically slows down (crippels?) CPUs and sells them for less as a marketing thing? I assummed it was a matter of "binning" various CPUs according to testing or something - but, come to think of it... that doesn't seem possible. What... is there like a micro-code instruction where they dial down the CPU speed to make it run slower?

 

[JamesJones44]

Can someone explain the accelerator on demand thing? Are the chips cheaper to produce with unactivated accelerator cores? How does this work? The cores have to function ahead of time right? So it's a software thing? It's all very weird to me.

Accelerators are usually created for specific uses cases that reduce the overhead related to general purpose compute units (aka what most people call a CPU or GPU core). The easiest comparison is Bitcoin mining, while general GPU cores can do it, they don't do it well due to all of the instruction, pipeline, etc. overhead. ASIC on the other hand can mine Bitcoin extremely fast compared to a general GPU because it's optimized for a specific software path (aka no overhead). Intel and AMD due this to a degree with things like SIMD instructions (not 1 to 1, but a similar idea).

In most cases these accelerators sit idle unless used and to my knowledge aren't really cheaper to produce. Software in most cases needs to be aware of these accelerators, but this is typically handled by compilers (aka my not required software overhead). Depending on the application the performance boost can be more than anything you will likely ever get out of a general compute silicon (bitcoin example). If a service comes along that trains a ML model for detecting types of birds, then these application specific accelerator units would be used largely over their general compute counterparts and do so in half the time with half of the power draw. Another example of this is what Apple has done with their M series processors and in some cases gives them an edge over much more expensive and more power hungry desktop level components while staying in a 60 watt envelope. Where they suffer is if the workloads don't use those accelerators. This is why you get some very different benchmarks results on Apple M series processors (in some can get close to a rtx 3080 in performance while others are miles away).

The best way to think about it as targeted application performance improvements. It makes sense when you know what your typical customer bases uses your components for. I would say for most server side deployments these days what Intel has done seems to make sense, but only time will tell if it hits the mark enough to make a difference overall once you trade away those general compute units or ends up being a failed experiment.

 

[thestryker]

Thank you for the feedback

Xeon SKU 6434 has only 8 cores and max turbo is only 4.1GHz.

The 24 (8p/16e) core Intel ‘Raptor Lake’ Core i9-13900K will run 5.8GHz.

The TDP of that part is 195W and they have capped turbo at the all-core turbo speed so it doesn't have single core scaling like consumer parts (this is fairly typical for server chips to ensure equal performance on every core at all times). All-core for the 13900K is around 4.9 when running 32 threads (taken from TPU's review) while pulling about 280W. On the 6434 the memory controller is much wider and there are a lot more PCIe lanes so the IO part of that server chip consumes significantly more power than the consumer version which limits the amount of power the CPU cores get. I wouldn't be surprised if 195W server was at best equivalent of 150W desktop because of this.

Your "market segmentation" comment. Are you saying Intel somehow physically slows down (crippels?) CPUs and sells them for less as a marketing thing?

Every manufacturer does some version of exactly that. Look at the clock speeds of the 13600K/13700K/13900k the maximum core clocks move up the higher you go. As far as I've seen all 3 of these are using different CPU die so it isn't a matter of binning (finding the best chip to use for higher end parts). I'm sure you could overclock any 13600K/13700K to match the stock boosts on the 13900K. Intel purposely sets the clocks lower to ensure someone looking at boost numbers sees everything higher as you go up (AMD does the same thing). Now this isn't universally true at the server/workstation level, but it certainly can happen for the exact same reason.

 

[Tech0000]

A 24 core 13900k is $599 at newegg. I'm not sure what you're going on about "consumer HEDT will stay dead". Looks like it's alive and kicking to me.

We have very different understanding of what HEDT is.

• 13900 has 8 performance core and and 16 so called efficiency cores (aka crap cores)
• 13900 has 20 PCIe

That is not HEDT. Just the 20 PCIe alone disqualifies it as a HEDT.
Great game machine, but a serious workstation class or enthusiast class cpu.

Last HEDT intel did, was the core i9 10980XE (2019) with fully enabled cores.

• 18 "performance" cores (real once)
• 48 PCIe lanes

I'm hoping W3400 will be priced in the same way they priced the core i9 10,9,7 X series, continuing that tradition.

 

[-Fran-]

I read about their Quick Accelerator thing and they require to modify OpenSSL (for example on crypto) libraries so it works, so it makes me raise an eyebrow already. I hope they don't create special libraries tied to the hardware and it can be made open/general, but this gives me a bad feeling already. This is IBM and Mainframes all over again =/

Regards.

 

[PlaneInTheSky]

Why do workstation and server CPUs run slower than desktop PC CPUs?

For the same reason you want to drive a car at 80km/h if you want to maximize efficiency. About 65%-90% of the gasoline you put in your car is wasted and transferred to heat.

Any appliance using energy humans have come up with, has some kind of optimal range where loss to heat is minimized. Anything outside of that tight range and almost all the energy is wasted to heat. It's the case with engines, computers, lights, etc.

Same reason many crypto farmers figured out that if they slightly undervolt their hardware, they can get way better efficiency.

Same reason why my I turned off the CPU turbo feature in the BIOS of my old notebook. An extra 40% battery life and no spinning fans. Now I have an ARM notebook and I don't have to worry about that stuff anymore.

 

[cyrusfox]

I am curious to see what will be the next gen Non Volatile Memory (NVM) options that will connect through CXL will appear on the market in a couple of years ?

And also what will be the latency because it will be usung the PCIe5 or PCIe6 bus, but Intel Optane DC PMEM were pluggable on the DRAM bus ?

Any guess ? Nantero carbon nanotube NRAM ? Everspin / Avalanche magnetic MRAM (STT-MRAM, SOT-MRAM, VG-SOT-MRAM,…) ?

The problem with all those techs[NRAM/MRAM] is density, largest dies too small and not cost competitive option. With Intel's exit on emerging memory[Optane], we will see the same old solutions in the new dressing of CXL (Battery back up DRAM, DRAM w/ SLC NAND hybrid, SLC NAND with aggressive controller, etc...).

Crystal ball/my prediction, we will not see a new memory emerge for another 3-5 years if ever. Only 4 companies are still working in this space: Micron, Hynix, WD (Purchased Sandisk), and Samsung[last I heard from Kioxia is them calling other companies foolish for wasting money in this space, they called it unworkable economically).
As only WD in that line up doesn't have a DRAM business, I would bet on them being the most aggressive as they don't have to worry about cannibalizing a current business segment. All of these efforts may not ever get out of the lab/fab and to the market, NAND is still scaling quite aggressively and still moving to more bits per cell (PLC!!!).

Emerging memory needs to scale 3D like NAND in order to be competitive and survive ( i.e. it needs to be able to lay down vertical stacks of memory cells). Otherwise it will follow the same fate as Optane(Better endurance and speed not enough to justify $/bit difference).

 

[kjfatl]

So Intel is making lower performing, less efficient, higher priced processors with confusing product segmentation and a supremely unethical upgrade model.

Let's see how that works out works out for them.

What is unethical about it? This is no different than buying an entry level software package and complaining that you don't get all of the features in the high end version of the software developed of the same code base. For the high end accelerators, most users will claim they don't need it if it if free, but enough will pay for it for Intel to justify developing the next accelerator aimed at a small customer base.

 

[Diogene7]

The problem with all those techs[NRAM/MRAM] is density, largest dies too small and not cost competitive option. With Intel's exit on emerging memory[Optane], we will see the same old solutions in the new dressing of CXL (Battery back up DRAM, DRAM w/ SLC NAND hybrid, SLC NAND with aggressive controller, etc...).

Thanks for the feedback.

I am not a specialist but I would think there might be, at start, a premium niche (= low volume) market for different type of Non Volatile Memory (NVM) with different optimization point in terms of speed, endurance, power consumption,… for different target applications where Non Volatility has enough value to justify a premium.

For example, as of January 2023, there seems to be quite a lot of ongoing government (IMEC,…) and private companies R&D in MRAM, and also already some sales by start-ups (Everspin / Avalanche technology) and manufacturing by foundries (Globalfoundries, TSMC, Samsung foundry,…) mainly for embedded / SRAM applications.

So MRAM seems the NVM the furthest along on the commercialisation/scaling volume path which could help reduce costs, and then help address other markets.

That said, it is a very slow process, and some government (ex: DARPA investing several 100M$ in Kioxia or TSMC to set-up a US Fab for HVM of 64Gbits VG-SOT-MRAM die) and/or regulations (ex: a law that would mandate that by 2035 all IT systems must use bistable Non Volatile Memory (NVM) instead of DRAM…) would tremendously help to accelerate the transition from volatile to Non Volatile Memory (NVM).

I believe that NVM is the holy grail of any IT system as it would be really disruptive in the way IT systems are architectured (no more boot-up time) and really, really want it to happen as soon as possible.

But again unfortunately, I understand that as the economics of NVM currently stands versus what is already on the market, it tremendously slows down the process of transitionning from volatile to bistable NVM memory .

 

[ottonis]

Depends on the workload.
Gaming GPUs have a hard time competing with Asics when they can mine the same crypto.

Is crypto-mining still a thing nowadays?

 

[ottonis]

I believe that NVM is the holy grail of any IT system as it would be really disruptive in the way IT systems are architectured (no more boot-up time) and really, really want it to happen as soon as possible.

To my knowledge, servers are usually running 24/7, so boot up time is not really relevant.

Also, boot up times could be massively shortened by software optimizations, e.g. when the OS does not need anymore to load myriads of processes up front and to initialize dozens of hardware elements at start, but only those it absolutely needs.

That being said, I would of course appreciate some non-volatile memory that would be instantaneously available and maybe hold the most critical hardware drivers or even some user data, e.g.a,static, fixed database of sorts.