News Intel and AMD forge x86 ecosystem advisory group that aims to ensure a unified ISA moving forward

[bit_user]

I got this from the review you linked - what im saying all along

AmpereOne is a bad example.

It's sort of like if you compared a Zhaoxin-designed x86 CPU to Snapdragon X and concluded x86 was garbage on that basis, alone.

 

[TheHerald]

What are you talking about? 72 cores vs. 96 or 128 cores!

Uhm ,transistors? It''s literally over twice as big as the 128c epyc. The 9754 is 80 something B, the nvidia chip is 200b, lol.

 

[TheHerald]

AmpereOne is a bad example.

It's sort of like if you compared a Zhaoxin-designed x86 CPU to Snapdragon X and concluded x86 was garbage on that basis, alone.

I get that, im not in disagreement, but you can argue the same about the nvidia chip. If nvidia had an x86 license Intel and AMD would go out of business, so...

 

[Kamen Rider Blade]

Bro, it's not about how many Transistors you have, it's about how you use it.

Let's take a simpler comparison in the ICE (Internal Combusion Chamber) Engine world. Maybe that'll be easier for you to understand.

Engine vs Engine, it's not about the # of Cylinders, but about the amount of HP you can output/extract for your design.

That's why everybody is going for lower cylinder counts, more energy efficiency / Thermal Efficiency per combusion cycle, and more HP while having ever lighter weight engines.

Most folks aren't just making big engines for the sake of it anymore.

Everybody is going for Maximizing Efficiency while minimizing weight within the optimal band of operation that they plan on targeting.

 

[bit_user]

Uhm ,transistors? It''s literally over twice as big as the 128c epyc. The 9754 is 80 something B, the nvidia chip is 200b, lol.

No, this is the same issue you had with the transistor count of Alder Lake. You can't just type something into Google and take the first thing that pops up, without checking it!

Nvidia is partly to blame for this, because they have this concept of a "superchip", which is two processors (each can be CPU or GPU) on a single SXM board. A lot of the info you see is on a Grace-Blackwell pairing. I think Nvidia never gave a figure on the number of transistors in a single Grace CPU.

Fortunately, The Next Platform did some analysis on Amazon's Graviton 4, which uses the same Neoverse V2 core. They estimated it uses between 95B and 100B transistors for 96 cores. That would put the expected number for 72 cores at just 71B to 75B.

AWS Adopts Arm V2 Cores For Expansive Graviton4 Server CPU

For more than a year, we have been expecting for Amazon Web Services to launch its Graviton4 processor for its homegrown servers at this year’s re:Invent,

www.nextplatform.com

 

[TheHerald]

Bro, it's not about how many Transistors you have, it's about how you use it

Exactly. That's why we are comparing 1 to 1 transistors between m3 and rpl. To see how they use them.

 

[Kamen Rider Blade]

Here's the reality, nobody cares about Performance per Transistors when doing product reviews.

At the end of the day, it's all about how much performance can I get, at this price bracket.

Am I willing to pay that much for it, what can I get.

 

[TheHerald]

No, this is the same issue you had with the transistor count of Alder Lake. You can't just type something into Google and take the first thing that pops up, without checking it!

Nvidia is partly to blame for this, because they have this concept of a "superchip", which is two processors (each can be CPU or GPU) on a single SXM board. A lot of the info you see is on a Grace-Blackwell pairing. I think Nvidia never gave a figure on the number of transistors in a single Grace CPU.

Fortunately, The Next Platform did some analysis on Amazon's Graviton 4, which uses the same Neoverse V2 core. They estimated it uses between 95B and 100B transistors for 96 cores. That would put the expected number for 72 cores at just 71B to 75B.

​

AWS Adopts Arm V2 Cores For Expansive Graviton4 Server CPU

For more than a year, we have been expecting for Amazon Web Services to launch its Graviton4 processor for its homegrown servers at this year’s re:Invent,

www.nextplatform.com

​

See, now we are getting somewhere. You know you could have posted that in page 1 instead of arguing about this useless disgustingly bad m3 chip.

So sure, if that chip is ~75B it seems to be competitive with the 128 epyc that uses 80b transistors. But the epyc was faster, wasn't it?

 

[TheHerald]

Here's the reality, nobody cares about Performance per Transistors when doing product reviews.

At the end of the day, it's all about how much performance can I get, at this price bracket.

Am I willing to pay that much for it, what can I get.

Exactly, performance at the price bracket. A 500B chip is going to be more expensive than a 10B chip. Pricing comes down to transistors. If ARM (hypothetically) needs 10 times the transistor to get the same performance as the x86 then simply it's going to be a lot more expensive.

 

[Kamen Rider Blade]

See, now we are getting somewhere. You know you could have posted that in page 1 instead of arguing about this useless disgustingly bad m3 chip.

So sure, if that chip is ~75B it seems to be competitive with the 128 epyc that uses 80b transistors. But the epyc was faster, wasn't it?

Nobody compares directly "Transistor Count vs Transitsor Count".

Everybody compares Core Count to Core Count or CPU to CPU for the price bracket.

The Amount of Transistors in the CPU or Core is largely irrelevant to most people & reviewers.

It just happens to be a impressive stat at the end of the day.

 

[Kamen Rider Blade]

Exactly, performance at the price bracket. A 500B chip is going to be more expensive than a 10B chip. Pricing comes down to transistors. If ARM (hypothetically) needs 10 times the transistor to get the same performance as the x86 then simply it's going to be a lot more expensive.

Pricing per Transistor isn't that simple, you're OVER-Simplifying the pricing structure into a linear count of Transistors when there are ALOT more factors that goes into the cost of a chip.

And Transistor count is a VERY small part of the cost.

 

[TheHerald]

Nobody compares directly "Transistor Count vs Transitsor Count".

Everybody compares Core Count to Core Count or CPU to CPU for the price bracket.

The Amount of Transistors in the CPU or Core is largely irrelevant to most people & reviewers.

It just happens to be a impressive stat at the end of the day.

Uhm no, You do realize that the reason the 7950x is 599$ while the 7700x is 299$ is because of the transistor count, right? And the reason it's faster is also because of the transistor count. You literally cannot say "I don't care about the transistor count, I care about performance / price" cause that's exactly what the transistor count comparison gets you.

 

[bit_user]

So sure, if that chip is ~75B it seems to be competitive with the 128 epyc that uses 80b transistors. But the epyc was faster, wasn't it?

First, I haven't verified the transistor count of the EPYCs, so I'm neither agreeing nor disagreeing with you on that point.

As I said before, the link I posted had the Geomeans of Grace + all three of the top-end EPYCs within a couple percent of each other, on the benchmarks performed in the article.

 

[bit_user]

Uhm no, You do realize that the reason the 7950x is 599$ while the 7700x is 299$ is because of the transistor count, right?

A lot else goes into their pricing, besides that.

But yes, obviously a bigger die (or more of them of a given size) on the same node is going to cost more. Assuming they bin the same, etc.

 

[Kamen Rider Blade]

Uhm no, You do realize that the reason the 7950x is 599$ while the 7700x is 299$ is because of the transistor count, right? And the reason it's faster is also because of the transistor count. You literally cannot say "I don't care about the transistor count, I care about performance / price" cause that's exactly what the transistor count comparison gets you.

You're also buying twice as many cores, so of course you're paying more for it.

You're trying to compare Transistor Counts between CPU Architectures that are vastly different, sometimes made on vastly different Process Nodes between different Fabs within this long argument with bit_user.

It's like trying to compare a 4-Cylinder vs 8-Cylinder engine.

One is going to be obviously more expensive than the other for obvious reasons.

But that's not why people buy them.

You just don't buy the bigger Cylinder engine for the sake of it.

Most people buy what they need, for the applications they plan on using things.

 

[TheHerald]

A lot else goes into their pricing, besides that.

But yes, obviously a bigger die (or more of them of a given size) on the same node is going to cost more. Assuming they bin the same, etc.

Well obviously a lot more goes into it but all else being equal - bigger = more expensive. A lot of people in this thread seem to disagree, they almost called me crazy for saying that. You are the first person that agrees actually, lol.

 

[stuff and nonesense]

Uh... do me a favor and remember you said that. It'll be worth noting, when AMD comes along with a version of Zen 5 with its front end swapped out for an AArch64 decoder, next year. Or maybe it'll be Zen 6. Anyway, I expect we'll have an even better apples-to-apples comparison than this Lunar Lake vs. M3 match up. Prepare to be surprised.

I suppose I should allow for the possibility there might be a handful of cases where x86 does better, but I think the vast majority are going to be either equal or swing in ARM's favor.

If you disagree, please tell us which use cases will favor which ISA, and why.

Yes, the apple chip is more efficient. It’s also tied to the apple ecosystem. It has good single core and good multi core.

Ignoring emulation they are code incompatible.

If you have an apple use case.. buy apple. If you have an Amd64 use case buy intel or Amd. Your use cases are whichever gives you best value.

Arm in general, not Apple, has made great strides since the first windows on ARM pc. It needs more native programmes. It also needs to capture the mindset of the corporate buyers, when this happens people might realise the potential benefits.

What you buy depends on your use case.

 

[Kamen Rider Blade]

Arm in general, not Apple, has made great strides since the first windows on ARM pc. It needs more native programmes. It also needs to capture the mindset of the corporate buyers, when this happens people might realise the potential benefits.

That also depends on the programs that corporations run smoothly, especially proprietary stuff.

If they have 'Zero Desire' to port the code to ARM & validate it, they might be out of luck.

That all depends on if they have some apps that won't work with the new ARM chips, regardless of emulation / translation layers. Also, if their dev teams want to spend the time to make sure all the software stack works for a few ARM LapTops vs just buying a x86 based one.

Which one is easier, spending more dev hours making things work, or just buying the x86 LapTop and moving on with your life?

Also, most Employees don't have a say in what kind of hardware Corporate IT gives them.

 

[Kamen Rider Blade]

Anyways, I'll be back later to check on this heated argument.

 

[TheHerald]

You're also buying twice as many cores, so of course you're paying more for it.

No, you are not paying more because it has more cores. You are paying more cause it's more expensive to make. And the reason it's more expensive to make is because it's a bigger die because it has twice as many transistors.

Transistors are the building blocks, if one ISA needs 10 times the materials to produce the same result as another ISA then yeah, it's going to be more expensive.

 

[Kamen Rider Blade]

No, you are not paying more because it has more cores. You are paying more cause it's more expensive to make. And the reason it's more expensive to make is because it's a bigger die because it has twice as many transistors.

It's more expensive to make because it literally takes 2x CCD's to make that chip.

So you're paying the price of 2x CCD's in the CPU + AMD Profit Margins + AMD Cost Margins + TSMC Fab Margins.

Transistors per CCD is just a stat.

Especially when you want to compare against Intel or ARM.

Everybody is designing for their own target goals.

Transistor Count per Core or per CPU Complex just happens to be what is needed to get things done for the performance capabilities of the core that company designed.

Some need More, some need Less. It's all about your engineering skills available at a company, and how they choose to employ it.

I don't get why you're so hung up on Transistors when we're telling you it's not a metric most people test against, and for good reason.

You're obseessed by it, but NOBODY else in the industry really cares about it that much.

Transistors are the building blocks, if one ISA needs 10 times the materials to produce the same result as another ISA then yeah, it's going to be more expensive.

The ISA plays a VERY small chunk of the over-all die space and performance equation.

The computer cores inside and how it handles the processing is what matters.

Jim Keller himself has stated it multiple times. ISA isn't nearly that relevant in this day & age.

The argument of CISC vs RISC is old, outdated, and largely irrelevant.

The rest of the computing engine matters ALOT more. And that is largely unique to how each company builds their CPU cores, regardless of ISA.

If you're going to be talking about ISA (Instruction Set Architecture), you should really go learn what it is, and how it plays into the Die Are of any CPU, it's usually a VERY small amount of total Die Area and very small amount of performance difference in the modern age.

 

[stuff and nonesense]

That also depends on the programs that corporations run smoothly, especially proprietary stuff.

If they have 'Zero Desire' to port the code to ARM & validate it, they might be out of luck.

That all depends on if they have some apps that won't work with the new ARM chips, regardless of emulation / translation layers. Also, if their dev teams want to spend the time to make sure all the software stack works for a few ARM LapTops vs just buying a x86 based one.

Which one is easier, spending more dev hours making things work, or just buying the x86 LapTop and moving on with your life?

Also, most Employees don't have a say in what kind of hardware Corporate IT gives them.

Fully agree

wrt employees, if they feel windows on arm is ok there is more chance of them buying for themselves.

 

[TheHerald]

It's more expensive to make because it literally takes 2x CCD's to make that chip.

So you're paying the price of 2x CCD's in the CPU + AMD Profit Margins + AMD Cost Margins + TSMC Fab Margins.

Transistors per CCD is just a stat.

Yes, to get the performance of the 7950x you need twice the transistors of the 7700x and therefore it's going to cost twice as much. Saying it's 2xCCD's is irrelevant. At some point theyll end up packing 16core in one ccd, the 16 core part is still going to be more expensive than the 8core part.

You're obseessed by it, but NOBODY else in the industry really cares about it that much.

That's just you saying it. Everyone is obsessed by it. That's why their chasing node shrinks etc. I mean literally both Intel and AMD trying to find ways to extract more performance per transistor with their e and zen 5c cores. Saying nobody cares about it is insane.

 

[Kamen Rider Blade]

Yes, to get the performance of the 7950x you need twice the transistors of the 7700x and therefore it's going to cost twice as much. Saying it's 2xCCD's is irrelevant. At some point theyll end up packing 16core in one ccd, the 16 core part is still going to be more expensive than the 8core part.

They already did that in Turin Dense. They packed 16x Zen 5C cores in one CCD.
But it's a EPYC / Server level product, unlikely to come over to Client side anytime soon.

That's just you saying it. Everyone is obsessed by it. That's why their chasing node shrinks etc. I mean literally both Intel and AMD trying to find ways to extract more performance per transistor with their e and zen 5c cores. Saying nobody cares about it is insane.

Nobody measures things with "Transistor" to performance the way you're trying to do it.
The way you measure performance isn't how everybody else is doing things, no where even close.

Go read some reviews, and ask yourself why nobody cares about the Transistor/Performance metric that you seem to care about.

Yes, everybody in the Fab Space is shrinking their nodes, to benefit from the advances in packaging density, lower travel distance, and the fact that Electricity travel hits the near speed of light limit within the silicon.

Everybody wants the performance bumps that come with Process Node Shrinkage.

But not everybody is going to utilize it in the way you're wanting to utilize it, just to throw transistors at the problem.

There are many good reasons why there are Die-Area, Transistor Budgets, Transistor per Core limits, and Transistor Budget for "Non-Core" related stuff.

It all feeds into costs / yields / Die Area / Heat Mangement / Power Costs / Areal Density / Energy Efficiency / etc.

There are a WHOLE hosts of metrics that matter, but you're misusing transistors as a measuring point for performance.

Something nobody else in the industry really uses the way you've been arguing with @bit_user about.

 

[TheHerald]

Nobody measures things with "Transistor" to performance the way you're trying to do it.
The way you measure performance isn't how everybody else is doing things, no where even close.

Well sure, reviews aren't directly talking about transistor counts cause the important part for a product is the price. But the price is directly derived by the size of the chip. inadvertently everyone is talking about transistor counts, since that's what it comes down to.

It's way more evident for example with GPUs. When the 4080 12gb was announced everyone was mad at nvidia cause it is such a small die at such a high price point. Cause obviously, a smaller die leads to less performance.