cloudy skies: Intel's future in the datacenter

Jump to latest Follow Reply
E

ev9_tarantula

Ars Praefectus
5,221
  • #41
    Megalodon said:
    Qualcomm acquiring Nuvia https://arstechnica.com/gadgets/2021/01 ... 4-billion/

    This is weird. I understand the impetus to hedge against Nvidia cutting off access to sophisticated ARM designs, but Qualcomm already had one and they got rid of it.
    Click to expand...

    That was because of all the drama around the Broadcom deal and wanting to cut costs, and Qualcomm's Centriq products wouldn't have been any better than using ARM IP like Amazon and Ampere are doing. Nuvia on the other hand are claiming absolute performance leadership against anything available at the time of release. If they can pull that off then it's a very good deal, and given Nuvia's belief that mobile and server processors now need to operate within the same power constraints (on a per core basis) you can see how Qualcomm would want to have that option for smartphones, laptops and the like.

    What will be interesting is to see if Qualcomm takes another run at servers, but my hunch there is that many of the potential customers are more interested in developing their own chips. Of course Qualcomm could license their CPU IP if they wanted to.

    I suspect it won't be long now until the fastest server processor is an ARM chip.
     
    Megalodon

    Megalodon

    Ars Legatus Legionis
    34,201
    Subscriptor++
  • #42
    ev9_tarantula said:
    That was because of all the drama around the Broadcom deal and wanting to cut costs
    Click to expand...
    That's a good point.

    ev9_tarantula said:
    If they can pull that off then it's a very good deal, and given Nuvia's belief that mobile and server processors now need to operate within the same power constraints (on a per core basis) you can see how Qualcomm would want to have that option for smartphones, laptops and the like.
    Click to expand...
    Yes, I think this suggests an expanded mobile+PC+server lineup.

    Nvidia seems interested in pursuing that strategy as well so there's every reason to expect competitive ARM PC chips from at least those two, and probably others like mediatek depending on how far Nvidia thinks they can abuse their ARM acquisition.

    ev9_tarantula said:
    What will be interesting is to see if Qualcomm takes another run at servers, but my hunch there is that many of the potential customers are more interested in developing their own chips. Of course Qualcomm could license their CPU IP if they wanted to.
    Click to expand...
    I'm not super convinced any of the companies like Amazon are getting anything exceptional in terms of the value proposition from their own chips, they just have to do it themselves to get specs and schedules that work for them, rather than waiting around for someone to make a commodity chip they can use off the shelf.

    ev9_tarantula said:
    I suspect it won't be long now until the fastest server processor is an ARM chip.
    Click to expand...
    I think that result was already essentially locked in. Looking at the current x86 server chips, they generally sacrifice clock speed to get massively multi-core scalability. Which means ARM cores that are not quite competitive with x86 when running 20W/core single-core turbo, easily win at the 5-10W/core power budgets more commonly seen in server chips. This is already true in some cases, but I think once the Neoverse N2 and V1 cores start becoming available it'll be true nearly across the board. That's not even considering a new Nuvia core with even better performance characteristics.
     
    W

    wrylachlan

    Ars Legatus Legionis
    12,768
    Subscriptor
  • #43
    Megalodon said:
    I'm not super convinced any of the companies like Amazon are getting anything exceptional in terms of the value proposition from their own chips, they just have to do it themselves to get specs and schedules that work for them, rather than waiting around for someone to make a commodity chip they can use off the shelf.
    Click to expand...
    A lot depends on volume and pricing. At the volume that Amazon is producing Gravitons now, a commercial ARM server chip would be better for them. If ARM takes over the bulk of AWS over the next few years that’s a different story. It also depends on how competitive the ARM server market ends up being. If Qualcomm gobbles up startups they could own the high end ARM server market and command pricing.

    All of which is to say that having your own ARM design seems like smart business for Amazon - if nothing else then as a hedge against the ARM server market consolidating and leading to high prices.
     
    E

    ev9_tarantula

    Ars Praefectus
    5,221
  • #44
    It looks like Google are going to build SoCs for servers.

    The future of cloud infrastructure is bright, and it’s changing fast. As we continue to work to meet computing demands from around the world, today we are thrilled to welcome Uri Frank as our VP of Engineering for server chip design. Uri brings nearly 25 years of custom CPU design and delivery experience, and will help us build a world-class team in Israel.
    Click to expand...

    https://cloud.google.com/blog/topics/systems/the-past-present-and-future-of-custom-compute-at-google
     
    Y

    YoHo

    Ars Praefectus
    4,319
  • #45
    I rushed here expecting to see posts about Intel's announcement today...Instead we're talking about Google...

    https://www.theverge.com/2021/3/23/2234 ... ufacturing

    Intel to outsource certain things to other foundrys, build 2 new fabs in Arizona (where were the last new fabs built? Oregon?) and create an arm that is a foundry for 3rd party chips.

    Also some sort of partnership with IBM.

    Microsoft is a partner on the foundry business, so I assume intel will be Fabing their next ARM design.
     
    E

    ev9_tarantula

    Ars Praefectus
    5,221
  • #47
    YoHo said:
    I rushed here expecting to see posts about Intel's announcement today...Instead we're talking about Google...

    https://www.theverge.com/2021/3/23/2234 ... ufacturing

    Intel to outsource certain things to other foundrys, build 2 new fabs in Arizona (where were the last new fabs built? Oregon?) and create an arm that is a foundry for 3rd party chips.

    Also some sort of partnership with IBM.

    Microsoft is a partner on the foundry business, so I assume intel will be Fabing their next ARM design.
    Click to expand...

    I posted about Google because they are one of Intel's largest customers, and now they are more or less saying "we'll build our own damn chips". Google is also a partner with Intel on the foundary side of things, as well as with Samsung (Whitechapel?) if rumours are correct, and TSMC for TPUs.

    With Amazon building Graviton chips, Microsoft apparently building an Arm server chip for Azure, and now Google on the same path you have to wonder how much of the web-scale market will be using the x86 architecture in 3-5 years time. We could see a situation where Intel remains the dominant manufacturer supplying that market, but most of chips are using IP from elsewhere.
     
    Megalodon

    Megalodon

    Ars Legatus Legionis
    34,201
    Subscriptor++
  • #49
    The Cortex X3 is reportedly up to a 6-wide decode, meaning it will be equal to the latest gen Intel chips for the first time. https://fuse.wikichip.org/news/6855/arm ... -cortex-x3

    It's also 64-bit only. The X2 was also 64-bit only, but the linked article suggests they didn't fully remove and optimize for 64-bit until the X3. Aarch64 removes a bunch of special cases present in 32-bit, such as predicates and thumb instructions. Since those are gone there's a kind of pseudo-branch case that no longer happens, and instructions are now always fixed width at 32-bits.

    The predicate style logic has returned in SVE2, which has predicate registers that allow each lane to have branch-style logic without requiring the processor to deal with flow of control. This is very similar to how GPUs deal with this. Since the X3 is ARMv9, SVE2 is mandatory.
     
    Megalodon

    Megalodon

    Ars Legatus Legionis
    34,201
    Subscriptor++
  • #50
    Reading the linked article the part on the "decoupled front end" caught my attention, I don't think I was aware of it even though the article notes it's appeared in predecessor cores to the X3.

    What that appears to mean is that the branch predictor is able to run ahead of the decoder's instruction fetch, sometimes quite significantly. This seems to improve branch prediction but also has the side effect of creating a very good instruction prefetcher. I believe Itanium had something similar and some of the later SPARC designs had something I think may be similar, what they called a "hardware scout".

    Anyway, this seems to be something they can get away with because of the fixed instruction width for Aarch64, as well as IA-64 and SPARC. It is always possible to look at a word anywhere in memory and know whether it contains a branch instruction. That's not possible in x86 because instruction alignment cannot be known without knowing the entry point into a piece of code. They work around this in various ways like a trace cache, but it seems to be quite simple to do for fixed instruction width architectures. My impression is that the branch predictor prefetcher has enough decode logic to identify branches from scratch, which would be prohibitively difficult on x86.

    My impression is that the Cortex X3 basically breaks even with x86 on a per-thread basis so there's no longer going to be much advantage there over ARM for Intel/AMD.
     
    E

    ev9_tarantula

    Ars Praefectus
    5,221
  • #51
    Megalodon said:
    Reading the linked article the part on the "decoupled front end" caught my attention, I don't think I was aware of it even though the article notes it's appeared in predecessor cores to the X3.

    What that appears to mean is that the branch predictor is able to run ahead of the decoder's instruction fetch, sometimes quite significantly. This seems to improve branch prediction but also has the side effect of creating a very good instruction prefetcher. I believe Itanium had something similar and some of the later SPARC designs had something I think may be similar, what they called a "hardware scout".

    Anyway, this seems to be something they can get away with because of the fixed instruction width for Aarch64, as well as IA-64 and SPARC. It is always possible to look at a word anywhere in memory and know whether it contains a branch instruction. That's not possible in x86 because instruction alignment cannot be known without knowing the entry point into a piece of code. They work around this in various ways like a trace cache, but it seems to be quite simple to do for fixed instruction width architectures. My impression is that the branch predictor prefetcher has enough decode logic to identify branches from scratch, which would be prohibitively difficult on x86.

    My impression is that the Cortex X3 basically breaks even with x86 on a per-thread basis so there's no longer going to be much advantage there over ARM for Intel/AMD.
    Click to expand...

    The Rock SPARC chip had hardware scout, as well as hardware transactional memory, but that never shipped. Nvidia's Denver has run-ahead execution after a cache miss. Itanium could use software scout threads, I don't know to what extent they had hardware support. Nvidia are still in the CPU game, so we will likely see some new Arm cores from them in the not too distant future, but I'd expect them to be more normal than their predecessors.
     
    C

    Claropus

    Wise, Aged Ars Veteran
    112
  • #52
    targetnovember said:
    TRX302 said:
    Whoever comes up with a CPU that isn't backdoored in hardware and can make their claim believable, will get our future business, even if we pay more money for less performance.
    Click to expand...
    Raptor Computing and their Power9 systems might be working towards what you want.
    Click to expand...

    For big iron servers, POWER10 systems look attractive. I saw a thing where a data center replaced over a hundred x86 server units with half a dozen POWER9 units, looking to go to 2 POWER10 units.

    That does sound like major energy/space savings (POWER10 is on 7nm), and who would not want that? Add in the fact that POWER architecture is not in several billion homes, making it slightly more challenging to binary-hack. The downside, of course, is that when you have a hundred units and one fails, you still have 99% of capacity – if POWER10 units are so expensive that you cannot afford enough to support your redundancy imperative, that could be a problem.
     
    theevilsharpie

    theevilsharpie

    Ars Scholae Palatinae
    1,199
    Subscriptor++
  • #53
    Intel release their Q2 2022 financial results, and it's an unmitigated disaster.

    • Top-line revenue is down 22%
    • Gross margins are down 20%
    • Datacenter and AI revenue is down 16%
    • Intel Foundry Service is down 54%

    On top of all that, the continued delays for Sapphire Rapids pretty much ensures that it'll be going up against Genoa rather than Milan, at a time when Intel's server processors still lose to Rome more often than not.

    Intel has also apparently thrown in the towel on Optane, and is completely winding down that business line.

    It'll be interesting to see what AMD's report looks like next week, but I doubt it will be anywhere near this bad. Intel has some dark days ahead.

    Edited to fix broken press release link.
     
    theevilsharpie

    theevilsharpie

    Ars Scholae Palatinae
    1,199
    Subscriptor++
  • #57
    According to Igor's Lab, Sapphire Rapids continues to be a train wreck.

    Igor's Lab":2y2vly75 said:
    Intel’s new Xeons, i.e. the scalable server processors of the 4th generation, are delayed again. Once again, one might say, and the reasons are very diverse. The problems and errors are collected under NDA in internal documents (“so called NDA sights”) and currently amount to almost 500 (!) entries, with the trend continuing to rise. With a total of 12 (!) steppings, they haven’t exactly covered themselves with glory either, and I can’t remember any project so far that needed so many steppings before you could even use it to some extent. This started with A0 and A1, then proceeded via B0, C0, C1, C2, D0, E0, E2, E3 and E4 until currently stepping E5! The market launch was planned about 1.5 years ago and the plan was recently updated again.

    ...

    Until now, Intel’s customers and partners assumed that the chips would be available in the second half of 2022, but internal information is now becoming a bit more concrete. Intel has now announced the “launch window” for Sapphire Rapids (SPR) for calendar week 6 to 9 (Feb. 6, 2023 to March 3, 2023), while the first shipment to selected recipients is still scheduled for 2022 in two waves. Calendar week 42 for the smallest models (2S) and calendar week 45 for the larger models (4 and 8S) are being speculated.
    Click to expand...
     
    theevilsharpie

    theevilsharpie

    Ars Scholae Palatinae
    1,199
    Subscriptor++
  • #58
    theevilsharpie said:
    Intel release their Q2 2022 financial results, and it's an unmitigated disaster.

    • Top-line revenue is down 22%
    • Gross margins are down 20%
    • Datacenter and AI revenue is down 16%
    • Intel Foundry Service is down 54%

    On top of all that, the continued delays for Sapphire Rapids pretty much ensures that it'll be going up against Genoa rather than Milan, at a time when Intel's server processors still lose to Rome more often than not.

    Intel has also apparently thrown in the towel on Optane, and is completely winding down that business line.

    It'll be interesting to see what AMD's report looks like next week, but I doubt it will be anywhere near this bad. Intel has some dark days ahead.
    Click to expand...

    AMD released their Q2 2022 financial results, and they've had a solid quarter.

    • Top-line revenue is up 70%
    • Gross margins are down 2%
    • DC revenue is up 83%

    Although not the focus of this thread, AMD is up in all other business segments, notably in Client. Intel reported dreadful results in their Client group (down 25% YoY), which people theorized was the demand for client PCs evaporating as a result of pandemic restrictions easing. AMD? Up 25% in client.

    It might be time to start an "Intel death watch" thread, because the so-called "rapid decline in economic activity" is increasingly sounding like an Intel-specific problem. :D

    EDIT:
    Dr. Ian Cuttress was live-tweeting a short-hand transcript from a post-earnings Q&A call, and while a lot of the questions were around AMD's closed acquisition of Xilinx, there was an interesting question and answer relating to overall market share.

    https://twitter.com/IanCutress/status/1 ... zL_5IrAAAA
    Dr. Ian Cuttress on Twitter":2jlnbnuk said:
    Jeffries Q: DC biz aligns to Intels. AMD gained 6.6% share from Intel, pro-forma from Xilinx is about 6%. That would be highest share gain ever reported, even back to 2005. AMD market share in the mid 20% now? Any outside contribution from Xilinx or Pensando that might mask that?

    A: This is why we went to new reporting segments. Your math is in our zip code. We're pleased we're gaining share, our expectation was as the portfolio expands, as we ramp, as we increase supply across the customer set, we would see share gains.

    To your question, there was no outside contribution - the other pieces in the segment are small. Mostly driven by EPYC.
    Click to expand...
     
    M

    md5crypto

    Ars Centurion
    258
  • #59
    Pretty shocking to see Intel post their first loss since 1999. I know they're on the right track but they might have 2 years of pain ahead.

    Here's the juxtaposition of what just happened:

    AMD Data Center segment revenue was $1.5 billion, up 83% year-over-year driven by strong sales of EPYC™ server processors.
    Intel Datacenter and AI Group (DCAI) revenue $4.6 billion down 16%
    Click to expand...

    That's literally AMD Epyc taking data center market share away from Intel as we speak. EPYC is killing it.

    The most interesting part is how Microsoft has untethered their existence from Intel. They are absolutely crushing it in the same period Intel is crashing it. Nobody would have ever thought that possible 10 years ago. I bet Azure DC are switching out to AMD EPYC as we speak.
     
    E

    ev9_tarantula

    Ars Praefectus
    5,221
  • #61
    Megalodon said:
    Unsurprising but Qualcomm sounds like they're going to take another swing at the datacenter with the Nuvia design. https://arstechnica.com/gadgets/2022/08 ... quisition/
    Click to expand...

    It doesn't surprise me. I follow this stuff quite closely and everything I've read about Qualcomm/Nuvia has been positive. Qualcomm seem to be doing a lot of hiring, and reputedly they are beating offers from any competitor. Right now it seems to be the place to go for ambitious CPU engineers. This points to Qualcomm being in it for the long-term, not just some acquihire to take out a potential competitor.

    Qualcomm have also been more specific about performance targets, and at the least they aim to beat Intel in the same product category (laptops presumably) if not across the board, if they pull that off they'll likely beat Apple as well, as Qualcomm will need to beat Meteor Lake come 2023. The Lenovo ThinkPad X13s using the Snapdragon 8cx Gen 3 is a good chip using Arm IP, it wouldn't take a big jump in performance for Qualcomm to have a nice part for laptops.

    I think we are going to see a lot of interesting new Arm CPUs over the next couple of years, and there will be some serious RISC-V silicon as well.
     
    Megalodon

    Megalodon

    Ars Legatus Legionis
    34,201
    Subscriptor++
  • #62
    ev9_tarantula said:
    I think we are going to see a lot of interesting new Arm CPUs over the next couple of years, and there will be some serious RISC-V silicon as well.
    Click to expand...
    Yeah, I think China especially really wants to de-risk dependencies that might be subject to sanctions. Unsurprising we're seeing a lot of action there.
     
    theevilsharpie

    theevilsharpie

    Ars Scholae Palatinae
    1,199
    Subscriptor++
  • #64
    AMD issued a warning ahead of the 2022 Q3 earnings report that earnings would likely come in well below estimates, primarily driven by abysmal performance in their Client business segment (down 53% QoQ and 40% YoY). In fact, the drop has been so significant that Client is now AMD's smallest business segment by revenue.

    More relevant to this thread, AMD's Datacenter and Embedded business are still growing, although by single-digit QoQ percentages.

    It's unclear how Intel is affected by this, however due to Intel's underperformance in Datacenter over the past few years, a similar cratering of their Client revenue could leave them in a really tough spot.
     
    theevilsharpie

    theevilsharpie

    Ars Scholae Palatinae
    1,199
    Subscriptor++
  • #66
    From the article:

    Anandtech":8itzz3gl said:
    At a high level, AMD’s situation mirrors the rest of the tech industry, particularly other chip rivals like Intel and NVIDIA. For a multitude of reasons from reduced consumer demand to preparations for a likely recession, PC OEMs are significantly reducing their stockpiles of completed systems and parts. OEM inventories had become relatively bloated following the pandemic, as OEMs surged to meet the spike in demand for new client systems for a newly-remote workforce. But as the pandemic has subsided, so has the demand for new systems.

    As a result of these inventory drawdowns, OEMs are currently ordering relatively few new client processors. Which in turn is hitting chip suppliers hard as orders from their biggest customers are in a heavy decline.
    Click to expand...

    Whether this is an "industry-wide" problem or an "AMD" problem is unclear. Intel's Q3 report later this month should shed some light on it.
     
    theevilsharpie

    theevilsharpie

    Ars Scholae Palatinae
    1,199
    Subscriptor++
  • #67
    “Thousands” of Intel layoffs planned as PC demand slows and revenues fall

    Ars Technica":3orwqfu1 said:
    Intel is planning its first major layoffs in almost six years, according to a new report from Bloomberg. The report says that layoffs will "likely" affect thousands of its 113,700 employees, particularly in its sales and marketing departments, and that they could happen as soon as this month. Bloomberg says that Intel's last major layoffs happened in 2016.

    The alleged layoffs are the latest sign of trouble for the PC market and for the companies that make and sell PC components.
    Click to expand...

    Not looking good for Intel.
     
    W

    wrylachlan

    Ars Legatus Legionis
    12,768
    Subscriptor
  • #68
    theevilsharpie said:
    “Thousands” of Intel layoffs planned as PC demand slows and revenues fall

    Ars Technica":1xs57scj said:
    Intel is planning its first major layoffs in almost six years, according to a new report from Bloomberg. The report says that layoffs will "likely" affect thousands of its 113,700 employees, particularly in its sales and marketing departments, and that they could happen as soon as this month. Bloomberg says that Intel's last major layoffs happened in 2016.

    The alleged layoffs are the latest sign of trouble for the PC market and for the companies that make and sell PC components.
    Click to expand...

    Not looking good for Intel.
    Click to expand...
    If they lay off all the marketing people whose going to name the chips?
     
    theevilsharpie

    theevilsharpie

    Ars Scholae Palatinae
    1,199
    Subscriptor++
  • #71
    One more day until Intel's Q3 2022 report, and even more storm clouds are brewing:
    • Microsoft's Q3 report showed Windows license sales to OEMs have declined by 15%, which is yet another sign of softness in the client market
    • Seagate is laying off 3K, citing a lowered demand from client and data center customers
    • Microsoft claimed that higher energy costs materially impacted Azure gross margins this quarter

    That last bullet point is ominous for Intel, because benchmarks of Raptor Lake show that Intel is still well behind AMD (and presumably ARM server chip vendors) in perf/watt, and if that showing is at all reflective of Sapphire Rapids' performance, it's going to continue to struggle competing with Rome (nevermind Genoa) at a time when energy costs are making up an increasing percentage of a data center's TCO.

    The broader tech slowdown also frees up capacity at TSMC that AMD and ARM vendors can potentially use, which can help alleviate the situation that AMD has reported previously where their growth was limited by the manufacturing capacity. If that's no longer a bottleneck, Intel server processors are likely going to be a hard sell to anyone that hasn't tuned their workload specifically for Intel's proprietary accelerators.

    Speaking of accelerators, it remains to be seen how Intel will perform in datacenter GPU workloads, but, well... given that "it remains to be seen" is still the current status, Intel's execution on GPUs probably isn't keeping anyone at AMD or Nvidia up at night.

    I'll guess we'll see tomorrow, but if things are as dire for Intel as it looks right now, you really have to wonder where they go from here. They are clearly underperforming in the data center, they have much greater exposure to shrinking Client markets than their data centers peers (with the possible exception of Nvidia), and the breathing room afforded by profits from other areas of Intel is rapidly shrinking.
     
    Megalodon

    Megalodon

    Ars Legatus Legionis
    34,201
    Subscriptor++
  • #72
    theevilsharpie said:
    Speaking of accelerators, it remains to be seen how Intel will perform in datacenter GPU workloads, but, well... given that "it remains to be seen" is still the current status, Intel's execution on GPUs probably isn't keeping anyone at AMD or Nvidia up at night.
    Click to expand...
    My understanding is that Nvidia's first mover advantage is still going strong, and there's still not really any true alternatives to them if you want really good developer tools etc. CUDA is still ahead of the alternatives.
     
    theevilsharpie

    theevilsharpie

    Ars Scholae Palatinae
    1,199
    Subscriptor++
  • #73
    Intel released their Q3 2022 financial results. Company-wide, it's a bit of a mixed bag. The results still aren't great, but they slightly beat revenue expectations, Client was better than expected (although still down 17% YoY), and the company announced plans to cost costs by $10 billion through 2025.

    However, on the data center side of things, it's still a slaughter:
    • Datacenter and AI revenue is down 27% YoY, and missed estimates
    • Network and Edge is up 14% YoY, which missed estimates
    • Accelerating Computing and Graphics was up 8%. I'm not sure how that faired compared to estimates, but it looks like the operating loss for that division is getting smaller, at least

    More telling is the operating income for Data Center and AI. On quarterly revenue of $4.2 billion, Intel recorded an operating income of $17 million. :eek: In comparison, last year, DC and AI recorded an operating income of $2.3 billion, on a quarterly revenue of $5.8 billion. Looking at AMD for Q2 2022, their data center segment recorded on operating income of $0.5 billion on $1.5 billion in revenue.

    In other words, in order to achieve even the current data center revenue, Intel has had to sell their products basically at cost. Intel's "Financial Horsepower" is doing some pretty damn heavy lifting right now. :D

    I guess we'll see what AMD's results look like next week as a comparison, but Intel's data center folks are in a very deep and dark hole, and it's hard to see how they exit with their current roadmap. Their focus on accelerators seems reasonable, but with the slowdown in adtech, social media, and the broader tech world, they may have missed their window of opportunity.
     
    E

    ev9_tarantula

    Ars Praefectus
    5,221
  • #74
    If this article about Qualcomm's counter claims in their dispute with Arm is even close to the truth then the entire Arm business model is about to be upended. Even if these claims are merely threats, rather than planned changes to licensing, such behaviour seriously undermines the rationale for using Arm processors.

    https://www.semianalysis.com/p/arm-changes-business-model-oem-partners

    It might as well be summarised as "why you'll be using RISC-V in 5 years time".
     
    theevilsharpie

    theevilsharpie

    Ars Scholae Palatinae
    1,199
    Subscriptor++
  • #76
    AMD released their Q3 2022 financial results, and as they warned, their client revenue has cratered by over 50%. Gaming is also down slightly.

    However, Data Center revenue continues to grow (if only slightly), up 8% from Q2 and 45% YoY.

    Overall data center revenue for Intel and x86 (combined) is down about $1 billion YoY. On the earnings conference call, Lisa Su explained this as enterprise pulling back spending, leaving more of the revenue share to cloud and hyperscalers, where ASPs are lower. Given that these customers have likely had early access to Genoa and Sapphire Rapids already, the financial results paint a pretty clear preference among cloud buyers for Genoa.

    Genoa is launching on November 10, whereas Sapphire Rapids is launching on January 10. So not only will Sapphire Rapids be competing with Genoa, it looks like Genoa will actually beat it to market by a few months.

    Speaking of revenue share, this quarter, AMD is closing in on 30% of the x86 server market.

    Turning back to client, it looks like there's downward pressure on price for that market. In their conference call, Lisa Su explained that they aren't chasing after low-profit markets, and it looks like they are fine ceding some of the lower-margin Client business to Intel while they focus on growth in Data Center and Embedded. We'll see how that works out for AMD, but if Genoa lives up to expectations (and Intel continues to underperform), AMD could actually be positioned to become the data center market leader within the next 2-3 years.
     
    theevilsharpie

    theevilsharpie

    Ars Scholae Palatinae
    1,199
    Subscriptor++
  • #77
    AMD officially launched EPYC Genoa today, and embargos on reviews have also lifted.

    Compared to Milan, Genoa is a mind-boggling performance uplift.

    From Phoronix's review: https://www.phoronix.com/review/amd-epyc-9654-9554-benchmarks

    Phoronix said:
    When taking the geometric mean of all the benchmarks that successfully ran on all processors, here is how things shake out. Even a single EPYC 9554 comes out ahead of the 2P EPYC 7773X configuration overall... AMD 4th Gen EPYC is great with its AVX-512 implementation, DDR5 system memory, twelve memory channels, and other Zen 4 architectural improvements. The 64-core EPYC 9554 2P was 64% faster than the 64-core EPYC 7763 2P configuration overall, or 67% if running the EPYC 9554 2Ps in the power determinism mode. Meanwhile the flagship EPYC 9654 2P was 74% faster than the EPYC 7763 2P or that went up to 85% when the Genoa flagship CPUs were running in the power determinism mode. The AMD EPYC 9654 2P was running at over 2x the speed of Intel's current flagship, the Xeon Scalable 8380 2P "Ice Lake" processors.

    ...

    From my testing of the EPYC 9554 and EPYC 9654 (and my preliminary testing with the EPYC 9374F -- stay tuned for my full review there!), the performance is incredible generationally and against the current Xeon Ice Lake competition while waiting for Sapphire Rapids. Particularly for the HPC workloads, the generational gains from Milan(X) to Genoa were some of the most captivating generational improvements and against the competition that I've seen out of the past 18+ years of running Phoronix for Linux hardware reviews. The huge raw performance gains are matched by competitive performance-per-Watt and competitive performance-per-dollar for driving great value in the data center. For many of the AVX-512 workloads, the performance-per-dollar of 4th Gen is extremely compelling.
    Click to expand...

    From ServeTheHome: https://www.servethehome.com/amd-epyc-genoa-gaps-intel-xeon-in-stunning-fashion/

    ServeTheHome said:
    The AMD EPYC 9004 series, codenamed “Genoa” is nothing short of a game-changer. We use that often in the industry, but this is not a 15-25% generational improvement. The new AMD EPYC Genoa changes the very foundation of what it means to be a server. This is a 50-60% (or more) per-socket improvement, meaning we get a 3:2 or 2:1 consolidation just from a generation ago. If you are coming from 3-5 year-old Xeon Scalable (1st and 2nd Gen) servers to EPYC, the consolidation potential is even more immense, more like 4:1.
    Click to expand...

    I suppose we'll have to see how Sapphire Rapids performs, but it's extremely unlikely to be able to close that kind of gap unless the workload can utilize one of Sapphire Rapids' accelerators.
     
    Nevarre

    Nevarre

    Ars Legatus Legionis
    24,110
  • #79
    There's... a lot to say about Sapphire Rapids since the lineup is so highly optimized for various workloads and various accelerators with a dizzying array of models but they did finally launch to try and counter AMD's Genoa. Keep in mind Sapphire Rapids was made public in 2019 and supposed to launch in 2021, not Jan. 2023. It replaces Ice Lake/Sunny Cove Xeons and that's not a particularly recent uArch (relatively early 10nm, same as Ice Lake on mobile, contemporary with 10th gen on desktop, but 10th gen was still 14nm for the non-mobile, non-Xeon parts.) There's been a long time between generations.

    www.servethehome.com

    4th Gen Intel Xeon Scalable Sapphire Rapids Leaps Forward

    We deep-dive into the 4th Generation Intel Xeon Scalable processors, codenamed Sapphire Rapids, to see a huge leap in Xeon capabilities
    www.servethehome.com www.servethehome.com

    Genoa is a game changer and great for consolidation with 96-core units, but the one gotcha is that it only supports 1S and 2S configurations. While it's much more dense per socket compared to a maximum 40 cores per processor on Intel Ice Lake and 60 on Sapphire Rapids, Sapphire Rapids has SKUs that support 4S and even 8S configurations. So you can technically have more cores on a single board for Sapphire Rapids because you're scaling up the max number of sockets. Those servers will not be anything most mainstream buyers will ever see, though.

    There are a lot of caveats to performance -- the Sapphire Rapids cores come with a variety of accelerator modules, optionally. If you buy a model with the right accelerator and it lines up with your workload and the developers support the accelerator for the apps you're using (a lot of gotchas) it can be a good product despite lower clocks, but it's not the case that those all are guaranteed to line up.

    The risk here is that Intel is having problems with its next die shrink. Meteor Lake (desktop) is probably delayed to 2024 and the Xeon variants are likely to lag beyond that. Meanwhile, Genoa X with a 3DVcache is coming this quarter and Bergamo -- the 5nm die shrink of Genoa-- is due shortly after for a density of 128 cores/socket. Even greater density is planned, so AMD is poised to leapfrog them again and isn't dependent on accelerators. The Xeon Max (Platinum) models designed for HPC are still coming and those may be more desirable and may start to be sold in more types of servers as Intel tries to come up with a solution for every customer... As an outsider looking in now, though, I have to wonder how many generations ahead AMD will be by the time Intel can replace Sapphire Rapids. 192-core future Epyc processors are already planned and they're scaling higher in clocks.
     
    theevilsharpie

    theevilsharpie

    Ars Scholae Palatinae
    1,199
    Subscriptor++
  • #80
    Nevarre said:
    There are a lot of caveats to performance -- the Sapphire Rapids cores come with a variety of accelerator modules, optionally. If you buy a model with the right accelerator and it lines up with your workload and the developers support the accelerator for the apps you're using (a lot of gotchas) it can be a good product despite lower clocks, but it's not the case that those all are guaranteed to line up.
    Click to expand...

    Regarding those accelerators...
    ServeTheHome said:
    In the launch materials Intel shows, and others discuss, you will hear the benefits of acceleration. That is Intel’s key benefit with this generation. At the same time, looking at three of the big four new accelerators, QAT, DLB, and IAA, less than 45% of SKUs actually have these turned on. DSA is the only accelerator on every SKU, but only 27% of the SKUs have the full DSA configuration, and most SKUs have only one-fourth as much acceleration capacity. Perhaps the strangest part of the launch is that Intel has been discussing acceleration in Sapphire Rapids for months, it is the company’s key competitive edge, yet more than half of the SKUs are either light on accelerators or do not have them enabled.
    Click to expand...

    So either Intel is having yield issues with these accelerators, or they're still clinging to their approach of exceedingly-fine market segmentation when they are in absolutely no position to be doing so.
     
    You must log in or register to reply here.
    Jump to latest Follow Reply
    Jump to latest Follow Reply