Nvidia Analyst Day Notes

These are sim­ply notes from the ana­lyst meet­ing held June 16th at the com­pany HQ.  I lis­tened via web­cast.  [Com­ments in square brack­ets are thoughts.]  For our pub­lished research and mod­els please login to the web­site and visit the library.

Notes:

The stan­dard com­puter today is far from sexy.

They indus­try is at an inflec­tion point that we will remem­ber going through for years to come.

Believes that a new archi­tec­ture is required and the GPU has an impor­tant role to play.  Future sys­tems will include both a sequen­tial proces­sor and a par­al­lel proces­sor co-processing along with it.

In order to meet the pro­cess­ing needs of rich inter­net appli­ca­tions, a pro­cess­ing archi­tec­ture like the above is what is needed.

Growth comes in three slices: mobile/embedded (TEGRA), work­sta­tions (TESLA) and con­sumer PC (GEFORCE / ION) and they have TAMs of $7B, $5B and $26B (con­sumer only) respec­tively. Today they are said to be $0B, $1B and $4B.

It’s not unnat­ural to believe that GPU spend should be equal to or greater than CPU spend.  [Inter­est­ing point, not the con­sen­sus view.]  Given how impor­tant visu­al­iza­tion is the power focused on that part of the prob­lem should be equal or greater to sequen­tial com­pu­ta­tion resources.

Sug­gests that today peo­ple are look­ing at actual per­for­mance and not pow­er­point slides of what chips will do and that Tegra is doing very very well.

The only way there can be main­stream mar­ket shifts in such an estab­lished indus­try area you need giant vol­ume and major changes in over­all per­for­mance.  [Says that CPU “hit the wall” but I don’t think that’s true at all.]

Mov­ing into CUDA dis­cus­sion which enabled them to put GPGPU func­tion­al­ity “on the back of” the stan­dard GPU “work­horse.“  120M units out there that can run CUDA today by down­load­ing and installing soft­ware.  But can the speed improve­ments be delivered?

As a new mar­ket and way of doing things it has taken a ton of work to “make this archi­tec­ture stick” and they think they have achieved it.  [So what about alter­na­tive tech­nol­ogy providers for the CUDA wave if there is in fact one coming?]

Tesla:CUDA as Quadra:GeForce ?

Power effi­ciency is an oppor­tu­nity area where GPGPU com­put­ing can be attrac­tive.  Many tasks like image pro­cesing, video transcod­ing are han­dled well by the GPU.  Gen­eral like this func­tion­al­ity is more inte­grated into the OS and acces­si­ble with Snowl­eop­ard and Win­dows 7.

Silly to have a quad­core x86 CPU pro­cess­ing video if you are watch­ing a movie or pro­cess­ing audio if you are lis­ten­ing to music.  The GPU is much bet­ter and more effi­cient at that.

Power man­age­ment and dis­s­ap­pat­ing power is very expen­sive ver­sus elim­i­nat­ing it.

Lots of dis­cus­sion of pro­gram­ma­ble shad­ing and CG and that it would lead to more gen­eral things like CUDA.  HPC and dig­i­tal film cre­ation are some of the few that don’t want to use Win­dows or MacOS.  Describes appli­ca­tion func­tion­al­ity like auto­matic image recog­ni­tion and tag­ging as some of the “magic” that the GPU archi­tec­ture will enable.

For games it’s not about how things look but what you can do with the game in terms of expe­ri­ence and more dimen­sions.  The PhysX engine is key.  Runs on a CPU but runs on a GPU faster.  3D vision is another way to bring more rich­ness to the experience.

The Nvidia approach does result in a higher cost prod­uct ver­sus ATI.  Main­tains that their gross mar­gin prob­lem has more to do with their $1B in excess inven­tory rather than struc­tural cost disadvantages.

They are try­ing to do some­thing dra­matic and are expand­ing the indus­try.  It means that there can be short-term pain and invest­ments to be made.  CEO “really believes” in what they are doing.  [This part must be scary to investors.]  Lead­er­ship of the com­pany is not tak­ing pay.  [Really?]

The mar­ket oppor­tu­nity will roll out as appli­ca­tions become CUDA-enabled.  One should see the emerg­ing com­pa­nies sum­mit for where Nvidia may invest.  Bought a stake in the com­pany that become Google Earth.  Lots of joint devel­op­ment going on with large com­pa­nies.  Mind­ful of neg­a­tive syn­er­gies from acqui­si­tions.  Tegra was built from the ground up around an indus­try stan­dard proces­sor.  [I won­der if the Intel/Wind acqui­si­tion has any impact on Nvidia?]

Tegra will be focused on the mobile/embedded oppor­tu­ni­ties that demand HD video where there are no other accept­able solu­tions.  [Have to see about that.]

Claims that they are focused on things which are too hard for oth­ers to do, will result in major design wins and gen­er­ate major increases in TAM.  These are the con­tin­ual tests that are run against the com­pany busi­ness plans.

Mike Ray­field, GM of Tegra — “First HD Mobile Processor”

80% of the top 100 web­sites have Flash built in. [?]  Hulu vis­i­tors aver­age 30 min­utes a visit.  Aver­age YouTube watcher spends 28 min­utes a day online.  Going HD.  Even FB has lots of images, video and ani­ma­tion.  Video min­utes on FB up 13x YoY.

Look­ing at Tegra com­plete its like a stick of gum.  8 proces­sors.  2 ARM, GPU, 2D engin, HDV encoder, HDV decoder, audio, image sen­sor.  Means that ARM proces­sor is run­ning at 5% when doing 720p video.  [Cisco backs up the notion that video con­tent is going to be 2/3 of what we are pro­cess­ing in the future.]

42 design wins so far.  About 18 are smart phones.  “House­hold names, house­hold car­ri­ers.“  Work­ing with 27 dif­fer­ent man­u­fac­tur­ers right now.  25m unit oppor­tu­nity in 2010.  First launches are end of 2009.  Work­ing with 27 dif­fer­ent wire­less car­ri­ers.  Many love this lit­tle Fox­conn note­book.  Trans­fer price to car­ri­ers will be about $200 which is per­fect for them.  [Are net­works the gat­ing fac­tor here?]

Car­ri­ers are excited about this netbook/MID oppor­tu­nity to drive their busi­ness.  100 of them were at Com­pu­tex where they don’t typ­i­cally go.  Car­ri­ers want to brand the devices.  (Wistron, Inven­tec, Com­pal and others.)

Aver­age TV watcher still spends 4 hours/day ver­sus the 30 min­utes for YouTube or Hulu.  [This could take a bite out of Win­dows use at the low end.]  Users typ­i­cally have lots of tabs open and many will have video running.

Able to show amaz­ing and real video at under 1w. (!)  [Early days but we will have to see what the Qual­comm Snap­dragon looks like.]  Work­ing closely with Adobe to be able to accel­er­ate Flash. (Not just Lite but also 10.)

The Mobile TAM is under $4B now but will reach $12B by 2013 (smart phones, portable play­ers, MID, auto­mo­tive have design wins today and oth­ers will join the mix.)

Tegra 650 is the prod­uct today.  T2 is 4x and due 1H2010 adn T3 is 10x and due out 1H2011.  Same power enve­lope (1/2 watt.)  [Can they do it?]  Main­tains that Intel is 2 years behind and before they catch up the 4x and 10x prod­ucts will be out.

“Dehy­dra­tion of the PC” strat­egy is a great neg­a­tive tagline for Intel.

Sug­gests they can be mid-upper 40’s in GM and “not be a drag” on cor­po­rate margins.

Com­pet­ing ver­sus QCOM which has modem around which they can wrap addi­tional tran­sis­tors to gain share of the smart­phone.  Sug­gests that pace of change in mul­ti­me­dia and modems are dif­fer­ent such that it makes no sense to lock them together.

A slower mar­ket does help incum­bents which can slow tran­si­tion to new tech­nolo­gies like Nvidia.  Basi­cally says that how much rev­enue falls into the cur­rent year is hard to say but by Q1 of next year the momen­tum will be very clear.  Said that the ASP range is $18–30 depend­ing on functionality.

Devices now are dom­i­nated by the dis­play which means there is “plenty of room” for chips.  Sup­ports Win­dows CE, Win­dows Mobile and Android.  Have designs going in all those.  [What about Palm WebOS, iPhone and Symbian?]

Ana­lysts seem to feel that they com­pany is involved with mostly “non-Tier 1″ part­ners so would like to see that change.

Talk­ing about $50-100M in rev­enue for the “back half of the year” and could be into the fol­low­ing year depend­ing on the car­ri­ers.  Divi­sion has 500 peo­ple so you can do the math to fig­ure out what the breakeven is.  [Not­ing a lit­tle testi­ness in talk­ing about numbers.]

Com­pany is start­ing a mag­a­zine?  What does the mix look like in three years?  1/2 Tegra and the other 1/4 (?) servers and work­sta­tions.  GeForce will be the rest.  In 20 years com­put­ers will be every­where, they will be very small and they will all be con­nected to the web.

Ana­lysts are very ner­vous over costs at this com­pany.  That’s clear.  IPTV will be a huge dri­ver.  It’s all going 1080p HD.  Not too many com­put­ers today can do that, let alone a $300 MID.

On the x86 they want to avoid it like the plague because of power con­sump­tion. It’s too much more than ARM.  When you start think­ing in terms of bil­lions of devices, 10 watts is too much.

Where are all the soft­ware devel­op­ers going?  It’s not desk­top.  It’s mobile.  iPhone, RIM, FLASH, etc.

Went through the gorey details of the 10 point col­lapse of mar­gins due to inven­tory and price cor­rec­tions.  The pro­fes­sional indus­try remains weak.  Auto indus­try is suf­fer­ing and they tend to buy lots of workstations.

Sug­gest­ing that they will get into the dig­i­tal video stream­ing mar­ket?  Needs a 30x price/performance improve­ment. [I won­der if Aka­mai is look­ing at this?]

Refers to an “enor­mous amount of mid­dle­ware soft­ware” to sup­port these proces­sors.  Invest­ing heav­ily in tools to help build soft­ware for the Nvidia plat­form and edu­ca­tion pro­grams at universities.

Enabling a new archi­tec­ture is a big deal, lots of cost.  Out­come is uncer­tain.  Look at Sony with Cell.  That cost a for­tune and basi­cally is a fail­ure.  Com­pany has 1000 hard­ware engi­neers out of 5600 but lots of soft­ware.  The devel­op­ment tech­nol­ogy orga­ni­za­tion is over 200 peo­ple strong and around the world.

“ION suc­cess would be 2-3x what it is today if not for the Intel FUD tac­tics.  The law­suit is about a future prod­uct and only con­fuses the mar­ket into think­ing it is related to cur­rent products.”

The goal is to put a GPU into the com­puter.  The chipset busi­ness is about the glue that makes that possible/easy to do.  It’s all just I/O.

Doesn’t a com­bined CPU/GPU vision favor the com­pe­ti­tion?  CEO ques­tions old views of mar­ket seg­men­ta­tion around user expe­ri­ence.  [Very inter­est­ing point, worth more thought.]

[Are we really talk­ing about old data-intensive ver­sus compute-intensive here?]

Nahalem is a great CPU. Atom is a great CPU.  Very unique ver­sus what Nvidia does.  The com­ple­ment is what mat­ters.  Mak­ing it too sim­i­lar defeats the pur­pose.  Larrabee x86 cores don’t make sense in this context.

Claims that there is a 10–15% pre­mium in the mar­ket­place for Nvidia.  Price wars hurt play­ers in the chan­nel if they are violent.

Inte­grated graph­ics have been inte­grated for a long time. It will set it’s own limit on what is pos­si­ble.  Would never make sense to do gam­ing or any­thing inten­sive on that inte­grated graph­ics platform.

Dis­cus­sion of USB and SATA in chipsets and why any­one would care about where it sits.  Seems to con­fuse most ana­lysts.   The goal is to have a CPU and GPU in every computer.

Is Nvidia behind on DX11?  Response is nobody has shipped it yet.  Aspires to be Apple in terms of talk­ing about prod­ucts that are released and avail­able now, not in the future.

Don’t con­sider cor­po­rate an oppor­tu­nity.  (!)  [They are not get­ting it on RealVR in the enterprise.]

The real fear finally voiced by ana­lysts is that Intel will kill them.  Because you can see the value that the Nvidia chip brings it makes it impos­si­ble for Intel to defeat them by inte­grat­ing a com­mod­ity func­tion like I/O.

Sug­gests that Intel is basi­cally say­ing that they will give $30 to you if you use Atom with­out Nvidia.  But for many mar­kets that still won’t do.  Those are the Nvidia cus­tomers.  The cost of an MP3 player is basi­cally zero but peo­ple still flock to Apple because for $49 or more it’s great.

When you turn CUDA on in a video edit­ing envi­ron­ment you will be amazed at what is possible.

Drew Henry — GM Desk­top GPUs

Dis­cusses the fact that the two main OS ven­dors are mov­ing to embed­ded sup­port for the GPU. You can write to OpenCL, DirectX, CUDA or FORTRAN.

ION is a new brand.  It’s aimed at the small, fully-functional part of the mar­ket. Specif­i­cally designed for this mar­ket and $299 price points.  GeForce is aimed at the higher end of the mar­ket.  Gam­ing, cre­ative, work­sta­tions, etc.

Run­ning 1080p Blue Ray full speed on an Atom PC as a demo.  To copy to a sony video walk­man it must be copied but also con­verted.  Both the Atom/Ion and Atom only sys­tems are run­ning Win­dows 7. The video transcod­ing is built into the OS which is very nice.  The GPU sys­tem can do a 2 minute clip in just over a minute the non-GPU sys­tem takes 15 min­utes.  The real dif­fer­ence is about 5x if you let them run.  So a one hour video con­verts in 38 mintues on an ION and 3.3 hours on an Atom.

And there are over 100 mil­lion GeForce GPU’s out there now that can all use this abil­ity.  Moves into dis­cus­sion about PhysX.  Inte­grated into major game engines, is cross plat­form and licensed by most of the major devel­op­ers.  [No dis­cus­sion on Havok.]  See Dark­est of Days for exam­ples of the effects.  See also Dark Void, U-Wars, Ter­mi­na­tor Sal­va­tion as well.   Giv­ing a demo with the GTX 285 (high end stuff.)

758 titles said to be sup­port­ing SLI today which will show up when peo­ple move to Win­dows 7.  3D Vision being demoed. Sev­eral hun­dred games can work out of the box in 3D.  Still expen­sive so it’s not mass mar­ket.  Work­ing on a low end 3D glasses approach to help more con­sumers get the experience.

The GPU is becom­ing the de facto stan­dard for pro­cess­ing video.  Parn­ers are tak­ing Nvidia tech­nol­ogy into this mar­ket for edit­ing, enhance­ment, con­ver­sion, upscal­ing, HD and Blu-ray.

Demon­stra­tion of VRe­veal which was devel­oped by the CIA to clean up video. Now this can be used for that huge inven­tory of VHS, cell­phone video and all that con­tent to be uploaded and used.

Show­cas­ing some devices like the Lenovo S12 note­book.  Turns an Atom into a fully-functional PC.  Sims3 sold 1.4m units in 7 days.  Over 100m units sold to date.  ION makes it playable on a net­book.  [Lots of the same points being made here for 20–30 min­utes or so.]

[Should do a blog post on Intel “crimes against humanity.”]

Intel and AMD are sup­port­ing Havok.  What does that mean?

Claims that PhysX and Havok are totally dif­fer­ent animals.  (?)

Basi­cally the fact that Win­dows 7 and Mac OS XI makes the GPU auto­mat­i­cally a full player in the com­put­ing envi­ron­ment is the key for the mar­ket shift.

Good ques­tion on what level to write to in terms of being able to have the OS opti­mize performance.

Andy Keane — GM Tesla Computing

Cray, Con­nec­tion Machines, BBN, Mas­par, etc.

Inno­va­tion in the HPC mar­ket is gone.  Mostly x86 now, some GlueGene.

Where is that next fac­tor of 10 or 100?

Want to do super­com­put­ing? Go down to Fry’s, get the $49 card and down­load the software.

Key indus­tries are Seis­mic, super­com­put­ing, uni­ver­si­ties, defense and finance (300m, 200m, 150m, 250m and 230m TAM for each one.)

“World’s High­est Per­for­mance 1U server.“  New category.

Two areas of Tesla.  Per­sonal Super­com­puter and Tesla co-processing clus­ter.  300 instal­la­tions of Tesla clus­ters.  Can democ­ra­tize the top end of the super­com­put­ing pyra­mid.  100’s will spend 10M, 100,000’s will spend $50k-$1M but mil­lions can spend less than $5K for a PSC.

David White, CFO

Reviewed the charts regard­ing the impact of the inven­tory cor­rec­tion and the impacts on mar­gin.  Com­pany put out a tar­get model get­ting back to 13–15% oper­at­ing mar­gins over time.

End­ing Q & A

Work­ing down inven­tory and of course there is a mix shift but why are wafer starts going up?  Answer is that Q1 was almost none so Q2 has to be up sharply.

How con­strained is 40nm?  Believes they have the terms with TSMC to get what they need.

GPU attach rates are going to go up sharply with the new OS releases later this year.  ION is a good entry level sys­tem but the dis­crete chips add much more pro­cess­ing power.

The busi­ness works bet­ter when chan­nel inven­to­ries are lean.  Win­dows 7 should help drive demand in the sec­ond half.  Dell Quick­sil­ver has two dis­crete GPU untits with SLI and a 9400 on the moth­er­board. Win7 and Nahalem should be a good com­bi­na­tion and a viable new enter­prise platform.

Tagged as: Nvidia