Charlie call 19th July.

Charlie call July 19th




Q: high level take?
A: things are slowing, question is how and how long,  hard/obvious question.  Not trying to answer.
Everything softening.  Technical look is who wins/loses and why based on tech.

Overview is; server side AMD benefits, unquestionable leader + production bound,  can’t make and sell enough.
Also have significant higher margins and pref per mm and $.

If market softens, AMD provide larger % of market because less consumer sales, turned into server SoCs.  Because of chiplets.  Chiplets are REALLY GOOD IDEA.

When have to pivot supply it pays off.  And they have to pivot and will pay off.  AMD takes market share and grow numbers in server relative to Intel.

Don’t think will hurt AMD pricing because they control and demand still over supply.  Little/no price reduction pressure.  Dunno hyperscaler negotiatons go.

Everything else enterprise/consumer Called “consumer”.
Intel will do better because competitive products, mid term Intel/AMD trading blows/equal in segment.
Intel in fall, AMD in spring launching,  leapfrogging who wins.

Trends showing Intel raising performance faster than AMD.  While trading blows, Intel slowly gaining lead.  Intel wins mid/long term consumer, slow back and forth between than.

Will take marketshare from AMD.  More like slowing AMD (potential) rate of growth.  Intel product comparable every way, if not better.  Thinks Intel will consistently gain ground on performance until in lead.

INTEL Can supply, AMD can’t.  Burned bridges last few years.  They get along emphasize server, but doesn’t help AMD case with OEM viewpoint.

Intel better support/engineering “and their drivers work, which I still can’t say for AMD”.

Non-tier 1 OEMs shifting towards Intel shifting next 18 months.  Starting next year.

/////////

Q: sapphire rapids, latest update on timing and reason it might be delayed.

A: two bits of timing.  PR launch and ship date.  Not correlated.  In may/June Charlie said SPR delayed.  Hard September 27/27 launch, found a bug.  Did right thing delaying it.  Charlie rarely says this.  Intel telling OEMs chips in February now.  Dunno if end of Q3 release is on, but chips are at February.  Usually Quarterly updates, hasn’t seen one yet.   Hold change.

Say did right thing because potentially (have to be careful about wording) security implications to what is being done and why.  Until know there isn’t anything that effects in wild or a fix, Charlie not getting into details
It could be serious, Intel did right delaying.


/////////

Q: large MCM, cost? Gross margin drag?  Technicals of it, works well?  Suck to much power?  Competiveness?

A: it’s a fine high performance chip that will more than make up for high cost and high power with extremely high performance when it releases in mid 2020!
Oh wait, 2023 chip—oops!.

Would have been fine 2.5-3 years.  Doesn’t move needle compared to Genoa due to delay.

Cost and power through the roof.  North of 400W, 450W last one seen for highest end.  More for HBM versions, no hard numbers.
Genoa is in same ballpark, hard to call out of line.
Extreme power hungry, extreme expensive to make, yields not an issue.  Even with renewables cores needed.  Die are quarter of size as before, and lots of products under 60 cores.

Packaging has yields, but sounds like very good yields.

Expensive, needle doesn’t move on performance.  If AMD not horrendously delayed, Intel in big trouble.  AMD delayed, so Intel dodged bullet barely.

//////.
Q: any DDR qualification issues?  If not, emerald/granite rapids?

DDR issues, yes.  Lots of issues and lots of features turned off that wil never be on.
Skylake to cascade “bug fix” was turning on lots of a buggy stuff.

Emerald bug fix and cost down for sapphire, new features are what was meant to be in sapphire, but couldn’t debug.

Performance same, cost lower, new features nothing to get excited about.  Think security stuff still broken, not out until granite.  Could change due to delays.  Non-issue unless u look at Intel margins

///////
Q: timing on granite overall?

A: short decision, don’t know.  Last time official guides be nothing had changed on Emerald and granite.  Than SPR delayed, and now launching where emerald was going to be.

Emerald delayed? Simultaneous launch?  Who knows.  No official word.

Educated guess is emerald delayed, but not by full year than sapphire has.  Granite will follow again in under a year.  No immodest pull in.

Problem not development or debug time.  It’s large customers hate platform changes.  Very expensive.  Less than a year on a platform is badly received by hyperscaler/tier1.

Less of can Intel pull in, more of if customers will stand for it.  Not a technical question.   Pull go either way. Depends who takes fall and how much piss off customers.

//////
Q: sierra forest, any updates?

A: yes, one after it is named publicly.  “Clearwater forest”.  More of the same.  A lot of little cores that certain markets demand.  For certain workloads very efficient.  Other workloads, 25% less big cores is better trade.

Is what it is, throughout/INT workloads, not heavy math. Can be signicantly more efficient than traditional way.  Some are hyperscaler internal.
Will probably sell really well.

Another server offering that meets certain workloads

///////
Q: mov to AMD now.  Thoughts on Milan?  If not go to Genoa and timing.  Said 2022 publicly, what is “horrendous delay”.

A: is the case, that Genoa launche in 2022 maybe.  PR launches versus chips delivered.  Being told Genoa is also February for chips.  Could be Q4.
Reason for delay is memory subsystem was broken.
Plans he told was would ship with 1DPC, couldn’t get more reliably working.

If familiar with history Zen1, AMD has had miserably time with memory working at speed.  Same issues with Zen1.  Said would fix in BIOS, they did.

We’re planning of shipping broken part and patching, for some markets okay, others less so.  Hyperscaler, can do 100K boxes push of button, but sending intern out of add DIMMs for half a million nodes isn’t palatable situation.

Sub-optimal to launch broken.  They delayed when heard Intel delayed.  To get chip in order and fix issues.  Ship it working.

Could launch 2022? Sure.  Fully working, doubtful.  Actual volume be February? Most likely.
PR LauncH is marketing decision.  PM decisions, not technical.

////////
Q: so respin? DDR5 issues?  Platform?

A: was DDR5, early this year, late last year.  All technical documents said DDR5-5200.  Few months ago DDR5-4800, like sapphire.
Told OEM would only ship 1 DPC working.

Delayed wheb Intel does so could fix all.

100% DDR5 issue,  in bad shape, they’ve been in worse shape in the past and fixed it.

Because of way they were gonna ship than patch in firmware, means didn’t need respin.  Could have used, maybe.  But more software/firmware.  Traditional AMD weak point.

//////
Q: AMD talked about Sienna telco part, interesting there?

A: Wrote an article it’s maybe  small Genoa, was wrong,  just a fused off Genoa, probably higher reliability parts/boards/packaging.

Nothing more than trying to start conversion than Telco, with a present in anticipation of upcoming Xylinx and real telco products.  No major sales.

More of “we have this products let’s talk”.

//////
Q: Turin?

A: no to early.

/////
Q: anything on C cores?

A: when it comes to big.LITTLE, AMD doing right.  Same chip as C, is what is in bergamo. Same chip, same code base, optimized different.  Does everting big chip can bit slower and way more efficient.

If so big.LITTLE with it, of version after, will just work.

Intel on other hand did every few wrong, than screws up more.
Chip requires tremendous software support and OS support.  Very little gain. Fused of AVX-512. Which they had said was killer feature.  Botched in every way possible.

AMD just said same chip, laid out and organized and tweaked different, so works on all code in all way.  Right way to do it

Same thing Apple did for their first big.LITTLE.

Technically right every way, Intel way wrong in every way.

Bergamo is 128 of those core.

//////
Q: hyperscaler have ARm options, bergamo/SF limit arm?

A: Arm has place, not mainstream.  Savings not worth risk for mainstream users.  ARM (SF?) and bergamo aimed at more specific internal company workloads than generic “throw it out there and have people buy it” could instances.

ARM Mainly been thrown out and can buy because we can.  Let people try and use.  If people will paying to pay, why not

Not to say ARM inferior, just not fully there in some way.  Edge cases, “many 9s worth of reliability questions”.

Same problem AMD/ARM have getting into data center.
Big difference in “software compile and run” versus “hardened and tested” and integrated into data enter, button press on console and things happen across AMD/Intel/ARM”

ARM last first, for most use cases fine.
You wanna take risk software errors out and crash, versus $20/month.

Hard call if your job.  No one got fired for buying IBM type scenario

Is there for some workload? Unquestionable.  Others coming up.  Lot of mindset issues and things that take time to prove.


///////
Q: how much she’s hyperscaler five to AMD, realistic stars? Over 50% any companies.

A: don’t do marketsharea.  AMD has massive TCO lead Intel can’t match.  Unless hyperscaler is stupid will buy as much AMD as they can, thAn buy Intel.  More what AMD supply than what Intel does.

Back to initial statements, now that demand slacking, AMD can gain share.


///////
Q: AM5 platform?

A: looks interesting, people think I’m insane.  But believe pluton untenable.  Says could end company if goes badly.   See his articles for rant.

In technical side, modern platform, first update in far to long.  Nothing Terrible, nothing incredible.  Intel and AMD trade performance blows as Intel inches ahead.


/////
Q: how heavily Intel use TSM for granite/graphics? Potential on CPU? Anything on PC?

A: granite is Intel.  Moved to better process node .  Right move.  First things he saw that out technical merit versus manager bonuses.  Sea change and Pat’s changes having effect and in right way.

Meteor some TSMC, don’t think what’s made where matters in significant way.


////
Q: Raptor lake thought?  Meteor lake?

A: Raptor next one, it’s fine.  Not much to say.

Meteor interesting.  Issue is high cost product.  Intel appears.  Trying to limit amount make and time on market.  Fine chip but expensive, potential to crater margins because packaging expensive way out of line in markets it is in.  Short lived and low volume.


/////
Q: Intel splitting consumer roadmap?  Meteor lake timing? Some say 2024

A:  last heard hard news it was quote “late 23”.  If it slips a little 24 is not a stretch or major delay.  If did slip, haven’t heard hard reasons.  No official work recently.

Latest wording to OEMs is 3 lines of consumer, low/middle/high.  Think i3/5/7.  Always been artificially crippling chips and selling back to consumers.
Now appears those lines being translated into significant silicon changes versus blowing fuses.
Changes like, meteor high cost.  Not in low end.
Low end “essentiall (-N on roadmap) doesn’t have meteor, stretch alder and pull in Cyprus.

$50 less tier 1 OEM pricing, two EMIB bridges, say $10, is a 5th of the ASP.  Doesn’t fly.  To expensive.

No info on Cyprus.  Being done to avoid packaging cost in markets not sustibained.

Meteor appears not to be on high end, Raptor -> arrow -> panther (not public if I know).
Unsure what arrow is, haven’t Hunted down details.

Changes doing in silicon for first time aware of tend to be fitting it to market segments for real instead of artificially.

Here’s the thing, for technical reason, don’t wanna get into, for next long while, everyone be monolithic than chiplets and tiles.  Technical reason forced meteor to be multi-die.  Reason will not occur again.

Don’t thin multi-die in consume space for a while.  Point products, but not a trend.

Fact they’re doing Silicon changes petty profound.


/////
Q: what caused meteor dekay?


 A: didn’t hear.


////
Q: outlook for 3D vCache? When mainstream


A: great tech if you need, if don’t need, is stupid.  Was stupid on consumer side.  To physically fragile to overclock.  It’s fine to limit, engineering done right.  No issue.
But no margin to overclock and so many handcuffs to get product out, it’s crap on consumer product.  Expensive shouldn’t have been done.  No reason can’t do again for every product on every line.

Big win on server, but unless they find way to not screw up no future on consumer


////
Q: AMD share gains as world softens and supply weakens?  If Intel raising pricing?  Thoughts on Intel price hike?

A: digitimee not news outlets, outlet for people who can’t speak because of culture,  when heard asked around.  Everything he looked to about it show as not true.

Contracts for long term CPUs are locked in quarter of two ahead of time.  Don’t think real.  Someone stirring the pot.  First 3-4 sources said heard nothing/getting nothing,  sources that should know.

Intel could, but Dell/HP/Lenovo, pricing long locked in.  Raising it on tier 1.5 or lower, ASUS/similar, you just destroy their competivenes.  Price hikes on paper? Maybe.  20% no way.  If they had performance advantage could do a thing.  Haven’t had for years


////
Q: AMD efforts in gaming GPU?


A:  AMD SW shit, object garbage.  20 years saying next version good.  Still sucks.  NOT GSMING.  Everyone in field GPGPU/compute.  Doesn’t work real world.  Works in demo.

Zero faith of AND fixing software issues.  Done with their bullshit.

For years, decade, higher run in performance in GPUs, not translate into sales.  Rarely into better performance.  Drivers don’t work.  Won’t sample because afraid of what Charlie will find.  ThT’s their issue.

Moving on, really think Intel will be #2 GPU player, 18-24 months, maybe longer.

Intel has volume, working drivers.  First gen and alchemist not real ones.  DG3 first one raja control.  Put early next year.  Real make or break.  Think Intel will take 2nd place with DG3 or DG4.

DG3 fist one designed by someone who knows GPUs and not 2 years late.  Much closer to right,  have working drivers unlike AMD.  Will have volume unlike AMD.

Don’t think AMD can fix problems in time to make difference.


////
Q: does SPR/Geno ramp delay hyperscaler to pause CAPEX/server purchases?


A: don’t think so.  If need cycles for customers, you need them.  If cost higher, cost higher.  Delay and turn away customers or have delays that piss them off.  Better to pay 10% to keep customers happy.  They can’t delay.

If anything they’ll go to AMD and say messing up TCO and ask to make up difference somehow with pricing games.


But AMD in drivers seat overwhelming TCO and performance.  Intel not in same game.  Any gestures AND makes is goodwill/face saving.


////
Q: with meteor with extra cost, how much ASP? What percent of corporate market address?


A: address a lot, solid chip,  haven’t seen numbers.  Never performance issue, just cost issue.  Can address anything, question is cost.  Can’t answer.
Last few years Intel long tail on products, old products go to lower tier systems.

Don’t think cost is an issue, will come out, win some benchmarks, how many hit market not relevant.  Could be halo product

////
Q: when does low end/mainstream refresh?  With meteor as silicon sub-straight how much cost?

A: packaging adds double-digit percent to BoM of meteor.  Depends on cost u give to EMIB.  $5 each =$10 on a <$100 chip to OEMs.

Cyprus is 2024, ALD-N low end could come out any time,  no rates for high end, RPT this year, arrow 24(?) Raptor 25.