Pat Gelsinger at NextSilicon Headquarters Full Interview

Pat Gelsinger
1 follower
Jan 31, 2026
6.8.2025 | Pat Gelsinger at NextSilicon Headquarters
Pat Gelsinger, distinguished General Partner at Playground Global and renowned semiconductor industry luminary, visited NextSilicon's offices.
During his visit, Pat joined NextSilicon founder & CEO, Elad Raz, and the entire NextSilicon team for an insightful talk about the current state of acceleration and the groundbreaking vision of Intelligent Compute Architecture.
Pat, I'm excited of having you here at MEC Silicon offices in Israel.
Is it your first time in Israel?
Well, not quite my first time.
This is probably trip 100, 110 to Israel.
As you know, many know, many, many visits to Israel due to Intel.
And, you know, visited much of Intel Israel work for me.
So I was here, you know, twice, three times a year for 30 years,
right, with Intel.
And, of course, this trip, right, the pleasure of seeing MEC Silicon.
But maybe the main reason is to be at Technion and getting an honorary doctorate
tomorrow at Technion.
So congratulations.
Well, thank you. Thank you.
Super exciting.
And we all know that Technion and Intel have had just, you know,
an incredible relationship over many, many decades to really make Israel
the tech nation that it is today.
You know, we have spent the past four hours here in the offices.
We walk around.
The amount of excitement that people have towards you and everything that
you have done in Intel and specifically in Intel in Israel is tremendous.
I don't know if many know, but Intel was one of the cornerstone of the
Israeli high tech in general.
A lot.
I mean, I think that most of the employees at MEC Silicon in the
hardware section were hardware employees that used to work for Intel
and started companies like Mellanox, later to be acquired by NVIDIA.
My company got acquired by Mellanox, later acquired by NVIDIA.
So it's all connected.
And I think that many people or you as one of the early sites here in
Israel, a big congratulation.
I spent my youth, you know, programming in x86.
And then when the 386 came, you know, and you had the 66 and 67 prefix
in the opcode in order to use the large 32-bit register, I was excited.
And, you know, meeting you in person, especially after all my history,
was a great excitement.
So thank you for doing that.
And maybe I would like to hear a little bit of your voice over about
how you see CPUs.
What's the role inside the compute ecosystem?
Yeah.
And, you know, such a wonderful history that I've had in computing and,
you know, being in the 286, engineer number four on the 386, engineer
number one, an architect of the 486, and then the design manager of the
486, you know, that x86 instruction set is near and dear to my heart and
soul.
And, you know, but it never really was about the x86 or the instruction
set.
It was always about the workload.
It was always about what you ran on it, you know, Windows and then Linux
being those operating systems that enabled all the application stacks.
And, you know, whether that was the, you know, the first DOS and Windows,
whether that was the first multimedia, whether that was the database,
whether that was the cloud software, you know, it was always enabling the
software stack.
And, you know, for that, the x86 was a seminal, right, enabler because so
much software was run on it.
It wasn't whether it was good.
It was whether it ran.
And then being compatible, right, was this magic of being able to have
generation by generation compatibility because it takes years to develop a
mature software stack.
And how can you take that software stack forward?
If you're compatible and you're faster, it was like magic, right?
And that was really the heydays, if you could, you know, the magic period of
Moore's Law where we just kept making the CPUs better and better, both
microarchitecture and process technology.
So, you know, this period of the history of the CPU was sort of magic, right,
in the sense that Moore's Law, the process technology kept getting better,
but we also were making the microarchitecture better, you know, run
more software, make the software run faster, advancements and, you know,
clocks per instruction.
All of that just sort of came the magic period of Moore's Law, the 386, 486.
You know, what a joy it was to be in the middle of that.
I think it was a tremendous run of the processor in x86.
As you said, we enjoy so much from a process innovation going down into the
nanometer, better frequency, and then the scaling came and frequency halted
and power consumption increased and GPUs suddenly came.
What do you think was the trigger that caused people to design massively
parallel applications?
Because CPU is always good for serial execution, for the operating system.
We will always need a CPU.
Most of my use cases when I'm, I mean, with AI, not anymore, but most of the
use cases is based on CPU.
So we are going to have them for life.
But what do you think was the trigger that causes people to look for new
architecture, GPU specifically, when the CPU performance got into a barrier?
You know, it was always about the workload, right?
And if the workloads were on the CPU, the CPU just kept winning more workloads,
right, and driving more of that ecosystem.
And I remember many a discussion between Jensen Wong and NVIDIA and myself on
scalar versus throughput and, you know, the work that they were doing.
And the only volume deployment for the GPU was graphics, you know, ergo graphics
processing unit, right?
No magic, right?
But, you know, and I'll say Jensen and I applaud him for this and NVIDIA, they
stayed true to that mission.
We're just going to increase compute density.
You know, we're not going to worry about, you know, the speed of execution or how,
you know, branches were being done and more and more of the silicon area was
being spent on non-execution, you know, branch buffers and target buffers and
buffers.
Because essentially there wasn't more parallelism to harvest in the CPU.
And that's where the GPU really started to emerge.
And, you know, it was, you know, this throughput, increasing density of compute,
creating more general purposeness to the software stack through CUDA and the
other steps that they were taking.
And then, of course, the magic moment was AI, right, where it was just not a
little bit better, it was 10x better to run on the GPU.
And then as AI became a mainstream workload, as opposed to this little niche
thing off of the side, the GPU broke out of graphics and really became a GP, GPU.
And today the history of computing is different because of that.
I think that, I mean, if we are looking at a CPU core, then the CPU processor,
2% of the CPU core is allocated for the ALUs and everything else is for out of
order execution, reorder buffer, all the queues and the queues in order to
optimize, as you said, latency.
How fast we can run instruction per clock.
And with GPUs, you have smaller cores, more of them, more throughput oriented.
Lots of threads, right?
Management of state is minimized.
It is all about compute density.
But one of the downsides of a GPU is, as you mentioned, or what now is known as
the software mode, is the CUDA, or domain specific language.
You couldn't take a CPU code, a general purpose code, a C, C++, Python code, and
simply run it on a GPU.
It will not run in an efficient way because of cache coherency, the way that
memory laid out.
In order to squeeze the performance, you need to know exactly what you're doing.
CUDA was invented back in 2006.
And it took around a decade until users were adopting the GPU itself as the
native compiler to run their application.
And I think that NVIDIA was super smart, super thoughtful, how they gave away
GPUs for free for university education program.
Think about other CPU alternatives like ARM, and a decade or two decades until
they got an adoption.
This is a GPU which was way, way more painful.
Now, Jensen said that he saw that a small percentage of the application runs
the majority of the time.
And a trillion of dollars were spent for general purpose acceleration, and it
just failed because if you could do that, that will be called a CPU.
So he decided to go after GPGPU and CUDA.
And it all started with CUDA.
So my ask to you is, what do you think about CUDA, software mode, domain
specific languages?
Well, you know, obviously, we've seen it being successful now.
But as you say, it took a decade plus of maturing that software stack, because it
was ugly.
It was hard, right?
Because all of the software energy was over here and general purpose compilers
and operating systems and lightly threaded workloads.
You know, that's where all the energy was.
Databases, user interfaces, we're very accustomed to that.
And figuring out how to program in highly parallel ways was hard.
And it just took a long time for that to develop.
But then important workloads, obviously, high performance computing was much more
parallelizable.
And then, of course, AI machine learning was the breakthrough workload, because it
isn't just, can I get the application there?
But do I care about that application?
And all of a sudden, it was this idea that I had a domain specific language that was
dramatically better and an important domain.
And to me, that was the thing that moved CUDA from this fairly niche, ugly
language to increasingly becoming a mainstream computing paradigm for the
industry.
But of course, now we have these two worlds, CPU and GPU.
What do you think?
Can we fix that and bring those back together?
Well, first of all, I believe that there is going to be a new era of compute.
The main reason is that GPU performance plateaued.
And why do I say that?
Because we are seeing GPUs with increased performance, but also increasing power
consumption.
Yeah, right.
And if you don't improve power performance efficiency, you're not improving, you're
just sucking more power right into higher density footprints and increasing power.
Exactly.
Power density, power per rack, power per data center.
And you need to normalize everything for dollar.
So it's power, performance, and dollar amount that need to be assigned.
I think that, and we are seeing GPUs getting to 3600 watts.
Yeah, which is just crazy.
I mean, it wasn't that long ago where a 400 watt CPU was like, man, right?
You know, can you push that back to 300 or 350 watts?
But now we're talking about thousands of watts for a GPU.
What happened, right?
So in order to show more performance, you are making the die bigger and bigger.
You move to chiplets.
So now you have four full retica chiplets all together aggregated to 3600 watts.
And like 4x more performance with 4x power consumption, that's a slight increment.
I do believe that the big next generation compute company will be the one that show
more performance, like 10x more performance, quarter of the power consumption.
But I think that one of the most important aspects on doing all of that is not by inventing
a new CPU or a new GPU.
Because if someone can develop a new GPU, a new tiny cores that can do all of that magic,
probably NVIDIA with the amount of resources or AMD or Intel that they are spending will
figure it out.
New novel architecture, like Nexilicon, a smart algorithm that understand what matters,
reconfigure the chip in a 20 nanosecond in order to give the actual kernel that run the
majority of the time, the bigger space of the die.
That's the way that the future goes.
But we are not stopping there.
It's not about the 10x more performance with quarter of the power consumption.
CUDA is a fact of life.
You have CPU application that run in C, C++ and for supercomputing industry, even Fortran.
The ability to have a software solution that is drop-in replacement, meaning it can run
GPU code, it can run CPU code out of the box, that's mandatory in order to ease the transition
into this new architecture.
And one of the key aspects that we have placed here at Nexilicon is the ability to be agnostic
to the software stack.
You can run CPU application written in Fortran or you can run AI workload VLLM with custom
CUDA kernel.
We can run them all.
Different architecture, smart runtime solution that understand and reconfigure the chip.
This is what we are all about.
And I'm excited about it and excited about the progress the team has made.
And I see that there's two dimensions of innovation.
To me, they're exciting.
This idea that we're moving past CPU, GPU and creating highly performant underlying
architecture that's driven by data flow, which is ultimately the limitation of how
dense we can make computing.
So bringing that together in a much more flexible underlying architecture.
Of course, we have RISC-V cores to run that serial code.
But to me, the software innovation then that that enables on top of that, because now we're
redefining because even in CUDA and GPU or the instruction set between the CPU and the
operating system, you had this hard separation between the physical silicon and the software
instruction set that was coming into that hardware.
Smashing that together as you're doing the next silicon to create the software hardware
boundary, maybe it's software or something that we're pushing those together to enable
this dynamic optimization in real time that maximizes the hardware utilization of the
running workload dynamically.
That's a pretty exciting set of innovations.
So I see it both GPU and CPU coming together, but then the hardware level coming together
with the low level software.
To me, that's sort of the magic of what we're doing here at NEC Silicon.
And it's really exciting to see the update today.
One of the things that you can see here when you're on the NEC Silicon offices is how
tightly integrated the hardware engineers and the software engineers, like same flow
hardware design, play some routes, software, low level, because I think that one of the
challenges that we had, and we have taken the long road.
We spent the last seven years building this general purpose accelerator, the one that
can take CPU workload, GPU workload, smash them together to have these workload processor
units.
And it's needed tight integration between all aspects of software, hardware together
in order to invent what we are seeing here at NEC Silicon.
And I do think the idea that you're not changing CUDA, you're not changing Python,
you're not changing Linux.
To me, that's all the way back to the beginning of the x86, right?
The magic was compatibility.
You weren't changing the software stack.
It takes years to build a mature software stack, to get tools deployed, to get applications
optimized, and being able to say, we're not changing the software, we're just making
it great with this new hardware.
That's the magic of what we're talking about.
But then combining it all so that it's future-proofed, because I can keep upgrading that runtime
software to make it more performant on the hardware that I've already installed.
So to me, that's sort of the magic of not just backward compatibility, but forward
compatibility as well.
Hundred percent.
I mean, right now we are living in 2025, middle of 2025.
Transformers is the architecture to go, talking about LLM and AI machine learning.
But as we have seen over the past five or six years, neural network architecture changed
so rapidly.
From CNN, Convolutional Neural Network for image classification, to Transformer today,
and who knows what the future holds.
Having the ability to be a future-proof accelerator is important.
Because neural network in 27 and 28 will look in a different way.
Looking at the future, you have recently joined Playground as a full-time partner.
How does it feel to be a venture capitalist?
How do you bet on those novel startups that will change the future of compute and silicon
fabrication and more?
You know, it's a thrilling time.
To be thinking about the future of computing.
Because we sort of got to a period that was sort of boring.
You know, the CPU was 400 watts.
And now we're talking about thousands of watts.
And, you know, we're sort of slowly incrementing the network and the memory.
Wow.
You know, the doors have blown open.
We're power envelopes for, you know, network designs, for size of configurations, data
center spend.
So everything's on the table.
So we're going to create a new microarchitecture with the next silicon.
We're really moving from copper to optical, right?
You know, power delivery networks are going to be rebuilt and optimized.
But it's also new forms of computing.
You know, we're about to launch a new superconducting company.
We have two of the premier quantum companies as part of the portfolio.
I call it the trinity of computing, right?
You know, quantum, classical, and AI all coming together.
We're going to be, new materials are going to be discovered.
And we're seeing that occur.
And, you know, the ability then to, I'll say, get comfortable operating in cryogenic
environments near absolute zero, single photon operation.
And, you know, it really is everything's on the table.
And being able to be part of a deep tech investor like Playground.
OK, this is a lot of fun.
And since I've been a deep tech guy my entire career, you know, being able to do that across
a broader number of companies, not just seven by 24 focused on one company, you know, but
now having a portfolio of companies.
And I'm thrilled that Next Silicon is part of that, you know, is a lot of fun.
And I expect that this next phase of my career journey is going to be one of the most fruitful
of my career.
I want to thank you for coming here.
I know that everyone, me included in Next Silicon, got really excited and thrilled of
having you around us.
Thank you for all your insight and everything that you have done.
I'm excited to be part of the Next Silicon team.
And with my history in computing, it's a real thrill that I get to express some of that
through you and the team here.
We have a great thing going on and I look forward to the bright future that we have
together.
