AMD EPYC 9754 CPU V-Ray Benchmarks Show Single Bergamo 42% Faster Than Dual Sapphire Rapids Chips

erek

erek

[H]F Junkie
Joined
Dec 19, 2005
Messages
10,900
Nice

"The AMD EPYC 9754 "Bergamo" CPU alone was able to beat two AMD EPYC 9554 64-core Genoa CPUs with a small lead of 3.2% The Genoa CPUs feature a much higher boost clock of 3.75 GHz, 21% higher than Bergamo. The CPU also offered a 41% performance uplift over the EPYC 9654, the flagship Genoa chip which also offers a higher boost clock of 3.7 GHz. The CPU was also able to beat a dual Intel Xeon Platinum 8490H platform with a lead of 42%.

But that's just one chip, what happens when two AMD EPYC 9754 "Bergamo" CPUs are used? Well, the answer is a massive 50% lead over dual EPYC 9654 CPUs and a monstrous 2.37x increase over the dual Intel Sapphire Rapids flagship option. Even compared to the current fastest Threadripper Pro CPUs, the 5995WX in a 2-way configuration, the EPYC Bergamo dual-socket solution delivered a 2.42x increase in performance.

That's a mighty performance boost and what's more impressive is that the 50% performance boost over standard Genoa chips comes from a 33% increase in cores while featuring a much lower boost clock. AMD's EPYC Bergamo already being much faster than Intel's Sapphire Rapids may spell bad news for the upcoming Sierra Forest E-Core-only chips which are still a whole year away."

1686807399027.png

Source: https://wccftech.com/amd-epyc-9754-...gamo-obliterates-intel-sapphire-rapids-genoa/
 
We will see how much of a buy-in these chips get.
There are problems with increasing core counts that much, as then adding the ram and IO necessary to feed them all becomes exceedingly expensive and AMD isn't promoting these to those who would want one or 2 but hundreds.
So they are essentially trying to sell these to data centers that are about to switch to ARM but haven't yet because they were waiting for their existing infrastructure to become EoL to do the replacement.
Depending on what you are primarily hosting ARM vs x86 means nothing, a web host or payment processor does not care either way because the work has been done and there is no more labor involved in provisioning a new server of one or the other, it is a Linux install and some libraries, PHP & SQL run equally well in both (realistically at that scale IO will be the limiting factor not any CPU architecture differences)
But this won't convince ARM customers to come back, and any large enough facility that was going to make the switch would already be well into the planning and development stages, and it's likely too late to change directions. Could this prevent a few from moving to ARM 2-3 years down the road? perhaps? But I think that AMD may have missed the mark with this by being too late to market with the concept.
 
Lakados said:
There are problems with increasing core counts that much, as then adding the ram and IO necessary to feed them all becomes exceedingly expensive
Click to expand...
Is it significantly more expensive to feed a 256 cores system with ram via CXL and what not (after the 24 channels are not enough) than feed in ram 2x128 cores system ?
 
  • Like
Reactions: erek
like this
LukeTbk said:
Is it significantly more expensive to feed a 256 cores system with ram via CXL and what not (after the 24 channels are not enough) than feed in ram 2x128 cores system ?
Click to expand...
Yes, sort of. Because you would not go from a 2x128 configuration to a 1x256 one, you would go from a 2x128 to a 2x256. So you are doubling the core counts which in turn requires at least if not more than double the memory density. Which then requires at least double the networking IO and close to 4x the storage IO.
So costs increase greatly with density. If they are only going to keep their core counts the same then they aren’t really adding any capacity. Things these are aimed at almost never pin a CPU, individual VMs or containers may pin their allocated ones but almost never the whole stack at once that’s why AMD can decrease clock speeds because they don’t matter much when your average load is 30-40% the generational increase in IPC will cover that spread. At that scale, you pin your IO long before you pin the CPU.
And due to how VMs deal with PCIe sharing you can start encountering packet collision on the network cards so there’s a whole world of potential hurt so in the end the CPU is not the thing the data center is primarily interested in but the platforms package as a whole. And I think AMD is coming up short here.

Edit:
This was recently posted in another thread here and sums it up similarly but from an authority on the subject.
https://www.theregister.com/2023/06/15/amd_aws_bergamo/
 
Last edited:
  • Like
Reactions: erek
like this
Lakados said:
Yes, sort of. Because you would not go from a 2x128 configuration to a 1x256 one, you would go from a 2x128 to a 2x256.
Click to expand...
Or you would go for a more dense solution (specially if you buy new). Because saying 512 cores worth of ram (and others) is higher than 256 cores worth of them, is not saying much, I mean of course no ? That was also true for for driving the 2x128 instead of going 2x32 where maybe your simple entry 8 ram channel was enough

Lakados said:
So costs increase greatly with density.
Click to expand...
In some ways in other not (smaller building, cooling solution, everything simpler in general), does at equal performance cost really increase greatly with density (and the only value of density is high performance fitting in the same building but at significantly higher cost) ?
 
LukeTbk said:
Or you would go for a more dense solution (specially if you buy new). Because saying 512 cores worth of ram (and others) is higher than 256 cores worth of them, is not saying much, I mean of course no ? That was also true for for driving the 2x128 instead of going 2x32 where maybe your simple entry 8 ram channel was enough


In some ways in other not (smaller building, cooling solution, everything simpler in general), does at equal performance cost really increase greatly with density (and the only value of density is high performance fitting in the same building but at significantly higher cost) ?
Click to expand...
Ram is an ungodly expense at that scale so it is a very big concern, but going that dense is a massive undertaking in Networking and storage, Amazon has spoken out about how they feel those chips are not the solution they are looking for because of many of the reasons I already talked about. Meta seems interested in them though, but I believe they are in the process of building a whole new data center.
 
erek said:
Nice

"The AMD EPYC 9754 "Bergamo" CPU alone was able to beat two AMD EPYC 9554 64-core Genoa CPUs with a small lead of 3.2% The Genoa CPUs feature a much higher boost clock of 3.75 GHz, 21% higher than Bergamo. The CPU also offered a 41% performance uplift over the EPYC 9654, the flagship Genoa chip which also offers a higher boost clock of 3.7 GHz. The CPU was also able to beat a dual Intel Xeon Platinum 8490H platform with a lead of 42%.

But that's just one chip, what happens when two AMD EPYC 9754 "Bergamo" CPUs are used? Well, the answer is a massive 50% lead over dual EPYC 9654 CPUs and a monstrous 2.37x increase over the dual Intel Sapphire Rapids flagship option. Even compared to the current fastest Threadripper Pro CPUs, the 5995WX in a 2-way configuration, the EPYC Bergamo dual-socket solution delivered a 2.42x increase in performance.

That's a mighty performance boost and what's more impressive is that the 50% performance boost over standard Genoa chips comes from a 33% increase in cores while featuring a much lower boost clock. AMD's EPYC Bergamo already being much faster than Intel's Sapphire Rapids may spell bad news for the upcoming Sierra Forest E-Core-only chips which are still a whole year away."

View attachment 576907
Source: https://wccftech.com/amd-epyc-9754-...gamo-obliterates-intel-sapphire-rapids-genoa/
Click to expand...

AMD EPYC 9754 “Bergamo” CPU With 128 Zen 4C Cores Being Sold For $5500 US, Less Than Half Its Official Price​

https://wccftech.com/amd-epyc-9754-bergamo-cpu-with-128-zen-4c-cores-being-sold-for-5500-us/
 
Just $21.50 USD per thread.... not that far from regular 7900x-7950x $16.50-17
 
  • Like
Reactions: erek
like this
when it comes to data center servers ... sometimes the cpu's are free
there was (maybe still is) a time not long ago that ibm would rent the cpu for free...but the ram was $1000 per gig ... and that was usually the show stopper
 
  • Like
Reactions: erek
like this
sunruh said:
when it comes to data center servers ... sometimes the cpu's are free
there was (maybe still is) a time not long ago that ibm would rent the cpu for free...but the ram was $1000 per gig ... and that was usually the show stopper
Click to expand...
You see that with Dell these days, go build out a server, and some how, 16GB of DDR5 ECC is $1k ...or more.. and you can even change models, choose the exact same size and speed of ram, and another server the cost is thousands more or hundreds less, for literally the same ram...
 
You must log in or register to reply here.
Back
Top