Intel Corporation's Knight's Landing Is a Huge Threat to NVIDIA

NVIDIA (NASDAQ: NVDA  ) has managed to carve out a market for its GPUs in the enterprise, selling its Tesla line of GPU accelerator cards for use in supercomputers and high-performance computing applications. These GPUs are typically paired with a CPU from Intel (NASDAQ: INTC  ) , but Intel has also attempted to break into the accelerator card market in recent years with its Xeon Phi line of co-processors. While Intel has seen some success, NVIDIA still has an overwhelming share of the HPC accelerator market, which was as high as 85% last July.

But, the next iteration of Intel's Xeon Phi, known as Knight's Landing, promises to give NVIDIA a real run for its money. Set to launch in 2015, Intel could have a real competitor in the HPC market, which threatens NVIDIA's growing enterprise business.

Why Knight's Landing might take a bite out of NVIDIA
Intel's Knight's Landing and NVIDIA's Tesla are very different products. Knight's Landing will be comprised of up to 72 Silvermont cores, the same cores used in Intel's low-power Atom processors, and it promises to deliver nearly 3 TFLOPS of double-precision computing performance. Tesla is comprised of thousands of GPU cores, and while the highest-end product only produces about 1.4 TFLOPS of double-computing performance, NVIDIA will likely refresh the line before Knight's Landing launches next year.

Certain applications will benefit more from having thousands of specialized GPU cores, while some will perform better with fewer, more powerful CPU cores. But, the big advantage for Intel is that Knight's Landing will be based on the company's 14nm process. NVIDIA still uses a 28nm process, and while it plans to eventually move to a 20nm process, NVIDIA is constrained by the abilities of the foundries to which it outsources manufacturing. NVIDIA only designs its chips, while Intel both designs and manufactures them, and with Intel spending $10 billion on R&D and another $10 billion in capital expenditures annually, it's safe to say that the company will be able to maintain a manufacturing edge over NVIDIA.

This move to 14nm should lead to more power-efficient chips, and with the performance-per-watt very important in supercomputers and HPC applications, this could give Intel a significant advantage over NVIDIA.

One last advantage that the new Knight's Landing will have over NVIDIA's Tesla is that Knight's Landing will be available as a stand-alone processor, as opposed to previous iterations that were solely add-on cards. This means that, instead of needing a CPU as well as an accelerator card, customers will be able to replace both with a single Knight's Landing chip.

This does, however, pose a threat to Intel's Xeon CPUs, which are typically paired with an accelerator card for HPC applications. It also means that Knight's Landing is unlikely to be cheap, since it would need to make up for lost Xeon CPU sales. But, it's still likely to be less expensive than a Xeon CPU plus an NVIDIA Tesla card, and that's what matters.

NVIDIA has one important advantage
While NVIDIA is a hardware company, software is proving to be a source of competitive advantages for the company. Much like NVIDIA provides code to game developers that is optimized for its GPUs, the company also offers tools and libraries written in CUDA, its proprietary GPU compute language, for common mathematical computations. This cuts down on development time and creates some lock-in for NVIDIA's enterprise hardware.

Of the top 100 supercomputers in the world, 18 use NVIDIA GPUs, while only nine use Intel's Xeon Phi. Coupled with NVIDIA's massive HPC market share, and it's clear that a lot of developers are using CUDA, this creates some significant switching costs. So, even if Knight's Landing provides better efficiency and performance than NVIDIA's Tesla, that may not be enough to end NVIDIA's dominance.

The bottom line
Intel's upcoming Knight's Landing will have some serious advantages over Tesla accelerator cards from NVIDIA, but NVIDIA has built an ecosystem around its hardware that may prove difficult to beat. Knight's Landing probably won't move the needle much for Intel, but its success has the potential to disrupt NVIDIA's enterprise ambitions. It's hard to say how this will all shake out, especially with Knight's Landing at least a year away, but NVIDIA certainly has something to worry about.

Leaked: Apple's next smart device (warning, it may shock you)
Apple recently recruited a secret-development "dream team" to guarantee its newest smart device was kept hidden from the public for as long as possible. But the secret is out, and some early viewers are claiming its everyday impact could trump the iPod, iPhone, and the iPad. In fact, ABI Research predicts 485 million of this type of device will be sold per year. But one small company makes Apple's gadget possible. And its stock price has nearly unlimited room to run for early-in-the-know investors. To be one of them, and see Apple's newest smart gizmo, just click here!


Read/Post Comments (2) | Recommend This Article (5)

Comments from our Foolish Readers

Help us keep this a respectfully Foolish area! This is a place for our readers to discuss, debate, and learn more about the Foolish investing topic you read about above. Help us keep it clean and safe. If you believe a comment is abusive or otherwise violates our Fool's Rules, please report it via the Report this Comment Report this Comment icon found on every comment.

  • Report this Comment On July 03, 2014, at 12:14 PM, dtor wrote:

    <snip>Tesla is comprised of thousands of GPU cores, and while the highest-end product only produces about 1.4 TFLOPS of double-computing performance, NVIDIA will likely refresh the line before Knight's Landing launches next year.</snip> The article cited below pre-dates this article by 3+ months.

    GPU Doubletake: Nvidia Launches Dual-GPU GeForce Titan Z for $2,999. (Dated 3/25/2014) http://www.maximumpc.com/gpu_doubletake_nvidia_launches_dual...

    <snip>What that translates into is an insane spec sheet. Covering the basics, the Titan Z is overflowing with 5,760 CUDA cores and 12GB of memory for 8 TeraFLOPS of performance. Don't bother sending emails about typos, those numbers are correct.</snip> And the new next-gen Maxwell is right around the corner - before Intel launches its brand new legacy hardware.

    Just to put this in some perspective, my *DESKTOP* motherboard (ASUS) is designed to hold 4 of the Titan Black or two of the Titan Z linked over an SLI bus. Am I a gamer? Nope. I'll give you two hints: 1) Fool One and 2) a copy of Hull's "Options, Futures, and Other Derivatives" next to my keyboard.

    FWIW, one does not need to learn CUDA to program a heterogeneous computer. M$FTs Visual Studio 2013 has AMP (Accelerated Massive Parallelism) http://www.google.com/url?sa=t&rct=j&q=&esrc=s&a....

    BTW, I believe your article should have noted that a significant portion of the NVIDIA customer base is financial institutions.

  • Report this Comment On July 05, 2014, at 7:42 AM, nobodyspecial2u wrote:

    Knights Landing is a H2 2015 product. 20nm NV cards will all be out before June2015. So any 14nm advantage is moot and they are quickly losing the fab war to people either joining forces or just plain people with more cash (samsung makes 3.5x Intel profits and can spend Intel's entire year profits on fabs and laugh). The exact things Intel gains from 22nm to 14nm, NV gets from 28nm to 20nm which will happen first. H2 isn't very specific, and Cuda6 has already laid the software groundwork for Maxwell based hardware (the next tesla's). We can already see maxwell has awesome power properties via 750TI. Tesla will be no different and the die shrink will double the cores and probably easily hit 3TF also. Again, nothing gained on NV here for Intel (I'm not counting the dual chip monster here already out as noted above, comparing chip to chip instead), it's just moving the goal posts for both sides.

    Pascal on the horizon with 3d memory/TSV, Nvlink etc also.

    It's almost comic you act as though NV is alone and without a fab they're crushed. They don't need fabs, with TSMC/Samsung/GF all being capable and sharing tech now (samsung/GF just moved on with IBM tech, now both can pump out the same chips easily, no need to retool/redesign). I'm fairly sure NV can feed the tesla's with Tegra's with the next gen as that is the point of the cuda 6 unified memory tech etc. There will be no need for Intel/amd to feed them soon if not already in testing. Imagine the next 18 NV supercomputers using NV ONLY and more surely coming as NV has never lost a gpu war to Intel. i740, Larrabee ring a bell? Phi is the leftovers of that last failure ;) Their drivers are abysmal also for gpu.

    NV has nothing to worry about, and it's easy for me to predict the winner here. There is no previous evidence to say otherwise and NV is gaining in top500, not losing it. Phi did nothing so far and maxwell shows NV has doubled down on the power side.

    Pascal, NVlink (no need for infiniband/hypertransport etc), 3d memory (TSV too), unified etc all mean they'll have the same stuff Intel does. There's an empty fab here in AZ to prove Intel's fab story isn't working so well either ;) They're failing to stop ARM also, as they have to give away Baytrail to compete. This will only get worse with 64bit Denver and it's enemies. WINTEL has problems coming massively from all angles. For 90% of the world WINTEL isn't needed. Most just check mail, play games, and browse the web. If you're not hooked on MS apps or other pro apps on windows you can leave as 21% of ALL notebooks leaving for chromebooks shows.

    Desktops are next. Servers too. Wintel will look completely different in 5-10yrs :) Fabs will catch Intel, and Windows will lose a lot of casual users to ARM64/Android L (slowing intel profits slows R&D, they also only have a few years of dividends left at current rate, if they cut dividends at least some will leave the stock, again hurting Intel). Usually Intel beats others to the next process by 6-12 months, this time competition's 20nm comes first vs. Intel's 14 or at worst a tie. Basically a year gained for their enemies. Intel already has finfet, and the competition gets that soon too, so again every move Intel makes (or already made), will be met by something offsetting their gains.

    Nvidia has to screw up to lose this, and for at least this next gen that appears to be a closed case based on 750ti. They have a ton of power savings to play with even on an old 28nm process as shown. Good luck Intel ;) Pascal appears to be heading in the exact direction needed too, so not sure the next gen after maxwell isn't a closed case already too. The sharing of process tech with GF/Samsung will surely lead to NV moving some production to samsung as you can quickly ramp if needed at GF with no design time needed like now. IF samsung keeps selling phones/tablets at their current rate or more they will be leading fabs in 5yrs. Intel can't afford to keep burning a Billion a quarter on mobile losses and keep R&D for fabs high. As noted if they have to cut dividend rates at some point (due to lower sales lost to ARM, burning 4B a year on mobile etc etc), their stock will be hurt. Look at AMD for the same show (R&D now dropped last 4yrs, lower than NV now, stock dropped from $41 to $4 in 5yrs etc etc).

    Ashraf overstates Intel's lead.

    http://www.eetimes.com/document.asp?doc_id=1321656

    Samsung capex $35B+ Intel 32B+, granted 60% of samsung goes to mem, but that leaves 14B for other stuff. His article last month acts as though Intel doesn't spend on modems etc too (and nand with Micron etc). It also ignores the profits of the two here, and the fact that the competition gets to 20nm at or before Intel's 14nm. That doesn't usually happen, so they are gaining. Samsung doesn't give the breakdown of where that 14B goes, just the 2.86B as he notes to R&D. That may not be spent on anything but direct competition to Intel's 2.5B that he notes. I think that other 11B left over from samsung's capex here is going to some of the stuff he's assuming is from the 2.86B R&D on semi tech.

    http://www.fool.com/investing/general/2014/04/22/3-reasons-i...

    You can argue this guys reasons, but cash down etc shows weakness, and accelerating mobile losses (looks like 4B this year, now 1B/Q) won't help grow R&D spending or dividend funding etc. If ARM64/AndroidL etc steals from desktop like they did already with chromebooks from notebooks (21% WOW) Intel will get even more pressure financially. It isn't Intel vs. Nvidia, it's Intel vs. Samsung/tsmc/GF/NV/Google/Valve etc. Valve hates windows/directX so they push opengl/linux/steamos. Surely NV/Valve will work to port SteamOS to ARM64 (it's the best way to make a gaming system without WINTEL and hurt DirectX/Windows which GABE N. would LOVE) and Denver K1, Erista, Parker etc @3.5-4ghz would surely make great CPU's for a full on gaming PC like box with 500w psu etc and discrete NV gpu card, multiboot (SteamOS, Linux, and Android L probably?) to cover all kinds of apps and games. Sure valve will still be in the Intel market, but the point is how much does this new ARM full PC like box steal from WINTEL? The loss to chromebooks (21%) says a LOT probably.

    http://investing.money.msn.com/investments/financial-stateme...

    10yr sheet on Intel, peaked 2011, down each year since as mobile takes it's bite. Hard to see it slowing as 64bit chips and Android L hit this xmas. Note their long term debt has nerly DOUBLED since 2011 also. So nearly double debt, profits down $3B, cash down 20% to 5B, mobile losses accelerating, delays to 14nm etc. Where is the sign of strength? It's not just the FAB war, it's the entire shebang that shows Intel will get weaker vs. this huge group out to steal WINTEL casual users (web, email, games) and even hardcore gamers if they do a full PC like box as described (16-32GB memory, TB HD's/SSD's etc, all like a PC). With NVlink (with TSV/3d memory), 64bit NV cpus, NV gpus, and 64bit OS's they have no need for Wintel. I'm not saying you win the entire market share, but they will take a sizable amount no doubt as shown by 32bit chromebooks.

    Ashraf seems to be spamming useless articles these days with no real point (thus no comments), so I commented on yours instead ;) He needs to go back to taking a side and defending it, instead of it could go this way or it could go that way. Blah, flip a quarter...LOL He used to make a point and lay out the details of why it is so, now he sounds like a democrat trying to figure out which way to go this week on obamacare. :(

Add your comment.

DocumentId: 3013635, ~/Articles/ArticleHandler.aspx, 7/23/2014 1:57:50 AM

Report This Comment

Use this area to report a comment that you believe is in violation of the community guidelines. Our team will review the entry and take any appropriate action.

Sending report...


Advertisement