Discussion AWS Graviton2 64 vCPU Arm CPU Heightens War of Intel Betrayal

[exquisitechar]

https://www.servethehome.com/aws-graviton2-64-core-arm-cpu-heightens-war-of-intel-betrayal/

Pretty big deal for ARM in servers. Interested in seeing a comparison between this and Rome.

 

[Hitman928]

I agree. You can't compare cost@22nm and 5 nm (not that he used those exactly). The point is, just for the mfg part, the cost goes up exponentially for each node. He is saying things like "take the retail, divide by 2 and thats the high discounted price", etc... No real numbers.

I got to this quote and then couldn't listen anymore:

What Intel's strategy seems to be is they are going to always, when AMD moves to the next process node, the leading edge of the next process node and incurs greater costs, Intel will continue to compete on the legacy node, on the trailing edge of the legacy node for a cost advantage.

Yep, Intel stayed on 14 nm so long so that they could have a (virtually non-existant) cost advantage against AMD. Yeah, I'm done. I'll reserve full judgement because I'm not going to listen to all of it, but seems like another accountant type who tries to analyze Intel/AMD as if they are Coke/Pepsi. Hint: this doesn't work.

 

[Hitman928]

Alright, last comment on this then I'm moving on. According to his estimates, this is the breakdown from lowest core count to highest core count:

4 core -> $55
28 core -> $178 - $200

So to get a 7x increase in core count, there's a 3.24x - 3.63x increase in cost. That's pretty magical. I wonder why Intel hasn't come out with a monolithic 40 core design yet, I mean, it should only cost them like $275 to make, right?

 

[RetroZombie]

I'm not going to listen to this for over an hour, can you point out the relevant timestamp?

Sorry you are right.
The relevant part is at 7:30. But the discussion started before around 1:26 about the capability of amd starting a price war with intel? Which by the way according to them is no.

There was another video before this one with more detailed intel prices by models (but to OEMs):
How Much Cheaper can intel get?

 

[Hitman928]

Sorry you are right.
The relevant part is at 7:30. But the discussion started before around 1:26 about the capability of amd starting a price war with intel? Which by the way according to them is no.

There was another video before this one with more detailed intel prices by models (but to OEMs):
How Much Cheaper can intel get?

AMD is much more prepared for a price war than Intel is.

 

[Nothingness]

Sorry for not responding.
I just used the AT Graviton2 review, and the AT 7742 2P analysis divided by two.
However, I definitely made a huge error in Excel - somehow the formula didn't copy and it didn't divide by two. Should have double checked.
Graviton2 is 80% the speed of 1/2 of a 2P 7742 by my calculations.

Thanks a lot for double-checking

 

[lobz]

AMD is much more prepared for a price war than Intel is.

While I doubt that per se, but they are definitely much more prepared than Intel think they are.

 

[Richie Rich]

Alright, last comment on this then I'm moving on. According to his estimates, this is the breakdown from lowest core count to highest core count:

4 core -> $55
28 core -> $178 - $200

So to get a 7x increase in core count, there's a 3.24x - 3.63x increase in cost. That's pretty magical. I wonder why Intel hasn't come out with a monolithic 40 core design yet, I mean, it should only cost them like $275 to make, right?

There is probably some fixed cost (testing, packaging etc.) which makes smaller silicon a bit more expensive (assuming good yields).
Also don't forget 4 core Intel has GPU which is doubling the area similar to 8 core w/o GPU. Then those numbers looks reasonable.


This means that Graviton2 estimated cost could be 300$ for 350mm2 silicon. And now compare G2 to Rome (7500$, 1005mm2).
If Rome offers +50% performance for 2.5 times higher power consumption (90W vs. 225W) this still means Graviton2 saves 40% of electricity cost per same performance. This allows AWS to set price much lower and no x86 systems can compete with that. IMHO x86 in cloud business is done.

 

[Hitman928]

There is probably some fixed cost (testing, packaging etc.) which makes smaller silicon a bit more expensive (assuming good yields).
Also don't forget 4 core Intel has GPU which is doubling the area similar to 8 core w/o GPU. Then those numbers looks reasonable.

He's comparing the Xeon server chips only, no GPU to add. Packaging cost will also increase with larger die so. . .

This means that Graviton2 estimated cost could be 300$ for 350mm2 silicon. And now compare G2 to Rome (7500$, 1005mm2).
If Rome offers +50% performance for 2.5 times higher power consumption (90W vs. 225W) this still means Graviton2 saves 40% of electricity cost per same performance. This allows AWS to set price much lower and no x86 systems can compete with that. IMHO x86 in cloud business is done.

Honestly, not even going to bother with this, just going to say, let's see how much market share Ampere, Altra, and TX3 can take before claiming x86 is done.

 

[moinmoin]

While I doubt that per se, but they are definitely much more prepared than Intel think they are.

What is there to doubt? Intel and its stock rating literally lives off of its high margin. Starting a price war would quickly deteriorate their margin. AMD on the other hand is still coming from a lower margin point, still not pushing for the max possible margin (which would be the level Intel demands) and as such has much more room to maneuver in a price war.

 

[Richie Rich]

Honestly, not even going to bother with this, just going to say, let's see how much market share Ampere, Altra, and TX3 can take before claiming x86 is done.

You don't bother because the math and numbers are clear and there almost no single argument to buy x86. Economy crisis is comming and nobody wants pay double the price due to power hungry x86 systems. If x86 will not improve power efficiency by at least 50% by architecture then they will loose server market majority in 4 years. Thanks to crisis maybe much sooner because there will be super high pressure for cost cutting. This can accelerate ARM expansion everywhere. Remember during crisis the weak ones dyeing first - and x86 is the weak one here....

 

[DrMrLordX]

You don't bother because the math and numbers are clear and there almost no single argument to buy x86.

. . . except that Rome is objectively faster per socket than anything the ARM server vendors are selling right now. All the big players have their Milan samples too, and likely have access to early commercial silicon through ODM contracts. You seem to reshape numbers to fit whatever broken narrative suits you.

Right now, today, there is exactly one ARM server product anyone would even consider, and those are Graviton2 instances through AWS. That's it. ThunderX3? Not available yet. Ampere anything? Not available yet. A64FX? Really niche and . . . availability unknown. There were some wafers sighted late last year. Oh wait, you can get the Huawei 64c chip, but I still haven't seen Kunpeng benched against anything.

 

[RetroZombie]

improve power efficiency by at least 50% by architecture

I think it's not the architecture who is wasting the power, it's the uncore part of the chip that was never designed to extreme low power levels like the chips that are going into the smartphones.

 

[Nothingness]

. . . except that Rome is objectively faster per socket than anything the ARM server vendors are selling right now. All the big players have their Milan samples too, and likely have access to early commercial silicon through ODM contracts. You seem to reshape numbers to fit whatever broken narrative suits you.

Right now, today, there is exactly one ARM server product anyone would even consider, and those are Graviton2 instances through AWS. That's it. ThunderX3? Not available yet. Ampere anything? Not available yet. A64FX? Really niche and . . . availability unknown. There were some wafers sighted late last year. Oh wait, you can get the Huawei 64c chip, but I still haven't seen Kunpeng benched against anything.

I fully agree with you. Except the A64FX is available, there's a system ranked 159 on top500: https://www.top500.org/site/50811
OK it's named A64FX prototype, but still that's much more than wafers

 

[Richie Rich]

. . . except that Rome is objectively faster per socket than anything the ARM server vendors are selling right now. All the big players have their Milan samples too, and likely have access to early commercial silicon through ODM contracts. You seem to reshape numbers to fit whatever broken narrative suits you.

Right now, today, there is exactly one ARM server product anyone would even consider, and those are Graviton2 instances through AWS. That's it. ThunderX3? Not available yet. Ampere anything? Not available yet. A64FX? Really niche and . . . availability unknown. There were some wafers sighted late last year. Oh wait, you can get the Huawei 64c chip, but I still haven't seen Kunpeng benched against anything.

Rome is faster per socket, that's right. But at the cost of lower performance per thread and 250 225W TDP. That's not clear win.
Graviton2 is also faster per socket than any Intel which is the majority of market right now.
80 core Ampere Altra comes soon this year - it will match Rome in terms of performance per socket.

If next year will come Graviton3 or other A78 based systems at 5nm (they can put easily 128-160 cores in monolith) then x86 is done.
And Amazon is 40x bigger company than AMD, they have money to adopt 5nm next year.

And don't forger A77 was about 20% INT IPC jump and 35% FPU jump. Just A77 based 128 core ARM will beat Zen3 Milan bad. 160 core A78 is the true nightmare for AMD and Intel. If ARM vendors will adopt chiplet design and start putting more dies per socket, that's game over for x86.

AMD's 5nm 96 core Zen4 is schedulled when? 2022? Outch... x86 is done.

 

[Markfw]

Rome is faster per socket, that's right. But at the cost of lower performance per thread and 250W TDP. That's not clear win.
Graviton2 is also faster per socket than any Intel which is the majority of market right now.
80 core Ampere Altra comes soon this year - it will match Rome in terms of performance per socket.

If next year will come Graviton3 or other A78 based systems at 5nm (they can put easily 128-160 cores in monolith) then x86 is done.
And Amazon is 40x bigger company than AMD, they have money to adopt 5nm next year.

And don't forger A77 was about 20% INT IPC jump and 35% FPU jump. Just A77 based 128 core ARM will beat Zen3 Milan bad. 160 core A78 is the true nightmare for AMD and Intel. If ARM vendors will adopt chiplet design and start putting more dies, that's game over for x86.

AMD's 5nm 96 core Zen4 is schedulled when? 2022? Outch... (too late, too weak)

200 watt or 225, NOT 250. The 7742 that I have is 225 watt.

As far as all your other estimations, I have stopped reading most of your posts as bad fiction.

 

[Thunder 57]

200 watt or 225, NOT 250. The 7742 that I have is 225 watt.

As far as all your other estimations, I have stopped reading most of your posts as bad fiction.

I don't know if he's delusional or trolling.

 

[DrMrLordX]

I fully agree with you. Except the A64FX is available, there's a system ranked 159 on top500: https://www.top500.org/site/50811
OK it's named A64FX prototype, but still that's much more than wafers

Okay, so someone can get ahold of it! Reviewers haven't though, which is a shame. I think that is a chip that is not meant to be reviewed for the public though (a pity).

Rome is faster per socket, that's right. But at the cost of lower performance per thread and 250 225W TDP. That's not clear win.

Your "per thread" metric is completely meaningless. The only way Graviton2 makes any sense is running large clusters of small VM instances. Make the entire chip try to do something at once on a larger workload and it chokes. It's clearly not suitable for a lot of workloads that Rome handles with aplomb. And as I mentioned, Milan is on the way (read: actually already here). Never mind the costs involved moving your software over to the platform, and that's if it can be moved. Proprietary software vendors may not make the switch unless they see a lot of green.

 

[Richie Rich]

Your "per thread" metric is completely meaningless. The only way Graviton2 makes any sense is running large clusters of small VM instances. Make the entire chip try to do something at once on a larger workload and it chokes. It's clearly not suitable for a lot of workloads that Rome handles with aplomb. And as I mentioned, Milan is on the way (read: actually already here). Never mind the costs involved moving your software over to the platform, and that's if it can be moved. Proprietary software vendors may not make the switch unless they see a lot of green.

For some type of workload which is hard to parallelize it matters a lot. There will have a G2 significant advantage.
Especially for desktop, laptop and gaming is performance per thread important. Imagine ARM competition for AMD Renoir 8c/16t for laptops. They can use 16x core A77 (ARM core has half area so resulting in same total area) running it at 2.5GHz and still get performance similar to Renoir clocked at 5GHz (and that's not even possible) while having 4x lower TDP. And that's the ARM power for which x86 has no answer.

A77 is big step forward (20% int IPC and 35% FPU IPC) and probably bigger than Zen3. Hypothetical G3 with 64x A77 cores paired with 128MB L3 cache would be very dangerous for Milan system. Especially with ability to put 2 or 4 dies into one socket (like Xeon or Naples) boosting performance per socket. Your assumption that all ARM server CPUs has to stay monolithic with small L3$ is wrong.

 

[DrMrLordX]

For some type of workload which is hard to parallelize it matters a lot. There will have a G2 significant advantage.

Such as? Be specific.

 

[chrisjames61]

You don't bother because the math and numbers are clear and there almost no single argument to buy x86. Economy crisis is comming and nobody wants pay double the price due to power hungry x86 systems. If x86 will not improve power efficiency by at least 50% by architecture then they will loose server market majority in 4 years. Thanks to crisis maybe much sooner because there will be super high pressure for cost cutting. This can accelerate ARM expansion everywhere. Remember during crisis the weak ones dyeing first - and x86 is the weak one here....

This post is pure fantasy.

 

[chrisjames61]

I don't know if he's delusional or trolling.

A heaping helping of both.

 

[lobz]

You don't bother because the math and numbers are clear and there almost no single argument to buy x86. Economy crisis is comming and nobody wants pay double the price due to power hungry x86 systems. If x86 will not improve power efficiency by at least 50% by architecture then they will loose server market majority in 4 years. Thanks to crisis maybe much sooner because there will be super high pressure for cost cutting. This can accelerate ARM expansion everywhere. Remember during crisis the weak ones dyeing first - and x86 is the weak one here....

I try to refrain from answering you nowadays, but I have to ask this here:
Oh mighty being, on which plain of existence are you living in, where all these things are happening you're talking about???
The universe is indeed a vast place.

 

[NTMBK]

Look at it this way- the ARM fanaticism makes a nice change from the old days, where Intel fans would explain how their fab advantage made it 100% impossible that AMD would ever catch up. I think AMD were six months from bankruptcy for about 5 years, according to this forum.

 

[Nothingness]

Look at it this way- the ARM fanaticism makes a nice change from the old days, where Intel fans would explain how their fab advantage made it 100% impossible that AMD would ever catch up. I think AMD were six months from bankruptcy for about 5 years, according to this forum.

I know you were being sarcastic. But fanaticism is never a good thing. The ARM fanatics here brought some x86 (both Intel and even more AMD) fanatics that are no better. That makes sensible discussion a pain.

 

[CHADBOGA]

I don't know if he's delusional or trolling.

It is always good to have an Internet Strong Man around.