You’re truly a bold person. I could never even think of shorting Intel.
What if, by any chance, 18A succeeds? What if someone like Liang Mong Song suddenly appears at Intel and revives their process? That kind of uncertainty keeps me from going long or short on Intel.
But one thing is certain: Intel’s long-term outlook doesn’t look good. They’re selling off the companies that Krzanich acquired, and the x86 moat is undeniably crumbling, as you pointed out.
Agree with MG. Epic brain drain. No moral. No leadership. No remaining trust from major customers and partners. There was hope with Gelsinger but the idiot board threw it away. Also I have puts, not short of shares. My downside risk is caped.
Intel has been shedding talent like a dog sheds hair for years. No one believes what they say anymore inside or out. I'm not who you're asking, but just imo, I don't think you can recover from that. They don't have the brain power to think themselves out of their problem, and the problem is titantic-sized, and it gets worse every day.
I am very excited about DIGITS. Maybe I could get a second-hand unit in few years for personal play :-)
This could move copilot-level LLMs out of datacenters and into private server rooms of businesses. Improved security and response time, opportunity for new hw/sw providers of customer-adapted solutions.
In my (rather pessimistic) opinion, the various copilots are the only AI tech which might be profitable & makes sense.
One key area that Intel needs (as in vitally needs) to execute on at least as urgently as in the mobile sector are their server CPUs. Now, Consumer Electronics is not the best place to discuss server hardware, but that is an area Intel is losing a lot of customers right now. Nvidia is still somewhat reluctant to use EPYCs to feed their accelerators, as AMD is also trying to compete with them in that area (Instincts), and the ARM-based CPUs Nvidia has just aren't as good. Have you heard anything about the state of Intel's server CPUs?
Intel Sierra Forrest and Granit Rapids are not competitive against AMD Turin and Turin Dense. That is CPU side. As for accelerator/GPU, Intel has nothing. Gaudi 3 is a failure. Falcon Shores is going to be incredibly late and is on a completely different software stack than the Gaudi line it replaces. As for networking/DPU/IPU, Mount Evans is rumored to be a dumpster fire. Heard that Google is not happy with that program. Rumored C5 stepping.
Why do you say Granite Rapids (P-core) is not competitive against Turin (Zen5)? From a technical perspective, they both have same number of cores (128-cores) and mem. channels (12 ch). On top of that, Intel Granite Rapids has support for higher 8800 MT/s MRDIMMs delivering higher memory bw than Turin..
Turn goes up to 192 cores per socket. AMD uses the same CPU microarchitecture for Turin and Turin dense. All cores have same ISA and overall performance. AVX-512 included. Only the cache configuration and physical design have been altered.
Intel Sierra Forrest has a different E-core CPU uarch. No AVX-512 or AMX.
I agree with most of your analysis, but I laughed when I read:
"Almost every programmer I know uses an ARM-based MacBook."
Clearly you don't know a lot of programmers...
99% of the time, programmers develop on the platform they deploy on. DC is mostly based on x86 (Intel/AMD), so developers for applications that run on DC develop them on x86. And that's including all the deep-learning AI software stack which is run on x86.
A programmer here. I develop on a Mac and deploy on... whatever the hell they are running in the cloud, I don't really care. A bunch of programmers I know do the same. Unless you're doing some fancy low level stuff, why would you care? And most of development is not fancy low level stuff, I bet many programmers don't even know about the CPU architecture differences.
I do boring old client side development and Apple Silicon runs compilers like a bat out of hell. C/C++, Swift, Go, Typescript, whatever. Nobody benchmarks that in their CPU reviews so it’s almost like a secret advantage.
All hyperscalers have ARM based CPUs for internal workloads and cloud customers. Nvidia is rapidly replacing x86 with their ARM CPUs. The trend of datacenter is rapidly towards ARM.
It needs to tank one more time. The gross margin and DC GPU estimates inside sell-side models is crazy. IDK where these Invesment bankers are conjuring these hardcoded numbers from. Going to write about this soon.
I get your point and agree with your analysis, going concern in accounting jargon meaning a company can continue to operate despite it sounding like the opposite.
I did not understand the Intel 7 (10nm) comment. Newer chips are Intel 4/3 or TSMC for the compute dies and N6 for SoC.
Anecdoctally, I am old enough to remember that the "Dell adopts AMD" thing is what market the stock price peak of the mid 2000s. It is not the first time.
Intel back-ported the Sunny Cove architecture from 10 nm to 14 nm for the 11th-gen CPUs. Apparently, they are going to back-port again from Intel 4/3 or TSMC N3 to Intel 7.
Only news that I know of seems to be Bartlett Lake. It does away with multithreading and 16 e-Cores in favor of 12 total P cores of the 12-14th gen variety.
This will hardly be a winner but should make use of the same Intel 7 capacity and maybe be a decent chip for mainstream desktops. Much like what AMD has been doing with Zen 3 and Apple with previous year chips.
Do you think that Intel's fabs, and generally the IFS, might be worth something, especially in the unlikely event that uncle Xi makes a move for Taiwan?
Maybe we underestimate wintel moat, and overestimate nvidia's.
The x86 moat is given by the complexity of the legacy x86 cisc platform, similarly the windows moat is in its obscurity, which is hard to replicate. If you need to run legacy sw which is not under your control (you don't have sources), you really need the full legacy platform for best performance. This does not apply for datacenters of course where they run custom code anyway, and the moat is eroding.
nVidia might not have a big moat. They produce at TSMC as everyone else after all. Their moat is in CUDA. Competitors (amd) must make products compatible with cuda, eg. by software translation in compiler/driver. The products might not even be higher performance, just lower cost / better efficiency, different markets (embedded, cars).
Similarly, I don't see a moat in ARM. In 10+ years AMD might be making successful GPU chips for cars and robots with own AMD-GPU core programmed in CUDA and controlled by their own RISC-V cores inside.
Intel has engaged some big-4 consulting firms who instead of helping with a good strategy amplified age old stuff coming out of Pat and his glorified team. Effectively, causing disservice to Intel and all it’s stakeholders.
A lesson for all to not blindly buy into pretty PPT slides when the Consulting firm folks have never managed anything at the scale of Intel!
Now it started to make sense, x86 is declining platform and it will be obsolete in the near future and AMD is realising that, that's why its developing Soundwave arm chip with Microsoft and may soon jumped to ARM too.
5-10 years. The problem is Intel is not in a position to lose more share. They are already in critical condition. AMD will be fine, especially in x86 HPC servers.
Whenever someone mentions HPC, it reminds me of Cerebras, too bad , they dont want to improve this amazing product instead want to dump the IPO on misinformed investor.
You’re truly a bold person. I could never even think of shorting Intel.
What if, by any chance, 18A succeeds? What if someone like Liang Mong Song suddenly appears at Intel and revives their process? That kind of uncertainty keeps me from going long or short on Intel.
But one thing is certain: Intel’s long-term outlook doesn’t look good. They’re selling off the companies that Krzanich acquired, and the x86 moat is undeniably crumbling, as you pointed out.
Do you think Intel has any way to recover?
Agree with MG. Epic brain drain. No moral. No leadership. No remaining trust from major customers and partners. There was hope with Gelsinger but the idiot board threw it away. Also I have puts, not short of shares. My downside risk is caped.
Intel has been shedding talent like a dog sheds hair for years. No one believes what they say anymore inside or out. I'm not who you're asking, but just imo, I don't think you can recover from that. They don't have the brain power to think themselves out of their problem, and the problem is titantic-sized, and it gets worse every day.
I am very excited about DIGITS. Maybe I could get a second-hand unit in few years for personal play :-)
This could move copilot-level LLMs out of datacenters and into private server rooms of businesses. Improved security and response time, opportunity for new hw/sw providers of customer-adapted solutions.
In my (rather pessimistic) opinion, the various copilots are the only AI tech which might be profitable & makes sense.
One key area that Intel needs (as in vitally needs) to execute on at least as urgently as in the mobile sector are their server CPUs. Now, Consumer Electronics is not the best place to discuss server hardware, but that is an area Intel is losing a lot of customers right now. Nvidia is still somewhat reluctant to use EPYCs to feed their accelerators, as AMD is also trying to compete with them in that area (Instincts), and the ARM-based CPUs Nvidia has just aren't as good. Have you heard anything about the state of Intel's server CPUs?
Intel Sierra Forrest and Granit Rapids are not competitive against AMD Turin and Turin Dense. That is CPU side. As for accelerator/GPU, Intel has nothing. Gaudi 3 is a failure. Falcon Shores is going to be incredibly late and is on a completely different software stack than the Gaudi line it replaces. As for networking/DPU/IPU, Mount Evans is rumored to be a dumpster fire. Heard that Google is not happy with that program. Rumored C5 stepping.
Why do you say Granite Rapids (P-core) is not competitive against Turin (Zen5)? From a technical perspective, they both have same number of cores (128-cores) and mem. channels (12 ch). On top of that, Intel Granite Rapids has support for higher 8800 MT/s MRDIMMs delivering higher memory bw than Turin..
MRDIMM is indeed an advantage to Intel. But overall, AMD wins this generation.
https://www.servethehome.com/amd-epyc-9005-turin-turns-transcendent-performance-solidigm-broadcom/4/
Turn goes up to 192 cores per socket. AMD uses the same CPU microarchitecture for Turin and Turin dense. All cores have same ISA and overall performance. AVX-512 included. Only the cache configuration and physical design have been altered.
Intel Sierra Forrest has a different E-core CPU uarch. No AVX-512 or AMX.
Intel has better DRAM controller than AMD their product is not dead
I agree with most of your analysis, but I laughed when I read:
"Almost every programmer I know uses an ARM-based MacBook."
Clearly you don't know a lot of programmers...
99% of the time, programmers develop on the platform they deploy on. DC is mostly based on x86 (Intel/AMD), so developers for applications that run on DC develop them on x86. And that's including all the deep-learning AI software stack which is run on x86.
A programmer here. I develop on a Mac and deploy on... whatever the hell they are running in the cloud, I don't really care. A bunch of programmers I know do the same. Unless you're doing some fancy low level stuff, why would you care? And most of development is not fancy low level stuff, I bet many programmers don't even know about the CPU architecture differences.
I do boring old client side development and Apple Silicon runs compilers like a bat out of hell. C/C++, Swift, Go, Typescript, whatever. Nobody benchmarks that in their CPU reviews so it’s almost like a secret advantage.
All hyperscalers have ARM based CPUs for internal workloads and cloud customers. Nvidia is rapidly replacing x86 with their ARM CPUs. The trend of datacenter is rapidly towards ARM.
Do you think AMD will be a buy anytime soon? It's pretty beaten up for their DC share, but gaining CPU share from Intel.
It needs to tank one more time. The gross margin and DC GPU estimates inside sell-side models is crazy. IDK where these Invesment bankers are conjuring these hardcoded numbers from. Going to write about this soon.
Perhaps in the last sentence you mean intel is not a going concern?
I phrased this poorly. Think Intel might run out of cash within the next 12 months. Their assets are mostly Fabs and these are extremely illiquid.
I get your point and agree with your analysis, going concern in accounting jargon meaning a company can continue to operate despite it sounding like the opposite.
I did not understand the Intel 7 (10nm) comment. Newer chips are Intel 4/3 or TSMC for the compute dies and N6 for SoC.
Anecdoctally, I am old enough to remember that the "Dell adopts AMD" thing is what market the stock price peak of the mid 2000s. It is not the first time.
Intel back-ported the Sunny Cove architecture from 10 nm to 14 nm for the 11th-gen CPUs. Apparently, they are going to back-port again from Intel 4/3 or TSMC N3 to Intel 7.
Only news that I know of seems to be Bartlett Lake. It does away with multithreading and 16 e-Cores in favor of 12 total P cores of the 12-14th gen variety.
This will hardly be a winner but should make use of the same Intel 7 capacity and maybe be a decent chip for mainstream desktops. Much like what AMD has been doing with Zen 3 and Apple with previous year chips.
Do you think that Intel's fabs, and generally the IFS, might be worth something, especially in the unlikely event that uncle Xi makes a move for Taiwan?
Yes. The problem is who would be the buyer. GloFo (AMD old fabs) had backing from Abhu Dhabi sovereign wealth fund. It did not go well.
Maybe we underestimate wintel moat, and overestimate nvidia's.
The x86 moat is given by the complexity of the legacy x86 cisc platform, similarly the windows moat is in its obscurity, which is hard to replicate. If you need to run legacy sw which is not under your control (you don't have sources), you really need the full legacy platform for best performance. This does not apply for datacenters of course where they run custom code anyway, and the moat is eroding.
nVidia might not have a big moat. They produce at TSMC as everyone else after all. Their moat is in CUDA. Competitors (amd) must make products compatible with cuda, eg. by software translation in compiler/driver. The products might not even be higher performance, just lower cost / better efficiency, different markets (embedded, cars).
Similarly, I don't see a moat in ARM. In 10+ years AMD might be making successful GPU chips for cars and robots with own AMD-GPU core programmed in CUDA and controlled by their own RISC-V cores inside.
You know why this is happening?
Intel has engaged some big-4 consulting firms who instead of helping with a good strategy amplified age old stuff coming out of Pat and his glorified team. Effectively, causing disservice to Intel and all it’s stakeholders.
A lesson for all to not blindly buy into pretty PPT slides when the Consulting firm folks have never managed anything at the scale of Intel!
Now it started to make sense, x86 is declining platform and it will be obsolete in the near future and AMD is realising that, that's why its developing Soundwave arm chip with Microsoft and may soon jumped to ARM too.
If by near future you mean 30 years, then maybe.
5-10 years. The problem is Intel is not in a position to lose more share. They are already in critical condition. AMD will be fine, especially in x86 HPC servers.
Whenever someone mentions HPC, it reminds me of Cerebras, too bad , they dont want to improve this amazing product instead want to dump the IPO on misinformed investor.
I was addressing x86 getting obsolete in the near future.
So you just take screenshots and type in all caps and call it analysis?
It's an irrational analysis
This is a lazy post on a weeknight. The 18A yield note has more details on rumored 10% overall yield of Panther Lake CPU tile. https://irrationalanalysis.substack.com/p/brief-intel-18a-yield-note