AMD Gen 4 Epyc up to 96 cores.

Some benchmarking was done at Level1Techs:


If I understood AMD seem to be able to down to around 1 watt for the actual core computtation (a lot of power goes into the 12 lane of DDR-5 and PCI-E lane controller), thus the very low difference of power betweeen 16 and 96 core possible.

Seem to some strategy around using AVX-512 really well instead of making specific AI-other hardware going on that could be interesting.
 
Last edited:
A shame these don't have good turbo. The giant 9654 is a whole 2.1 GHz (!!!) below the 7950X and 2.3 GHz below the 13900KS which is an absurd amount of single-threaded performance to give up to get your 96 cores. I guess we're back to the days of needing two computers if you really want to go fast.
 
Last edited:
A shame these don't have good turbo. The giant 7654 is a whole 2.1 GHz (!!!) below the 7950X and 2.3 GHz below the 13900KS which is an absurd amount of single-threaded performance to give up to get your 96 cores. I guess we're back to the days of needing two computers if you really want to go fast.

These go into server racks... database applications... virtualization.... containerization.... clusters. You know... You know: the big pieces of the internet. Parallelized workloads. Not your anemic/pathetic gaming machine.

Go ahead... fire up a serous database on your gaming machine. Do it... have fun. God help you if you actually put it under load. All hail the soft errors....

It's so tasty when some dude with some gaming machines plays "systems engineer".

Wait.... actually- it's kind of cute....

From now on we'll just call you "Pookie".
 
These go into server racks... database applications... virtualization.... containerization.... clusters. You know... You know: the big pieces of the internet. Parallelized workloads. Not your anemic/pathetic gaming machine.
Well, obviously. It's a server processor. Also I will have you know one of my gaming machines has 64 cores in it so, there you go (it's not a very good gaming machine but is very fast for work).

Go ahead... fire up a serous database on your gaming machine. Do it... have fun. God help you if you actually put it under load. All hail the soft errors....
I've done plenty of serious work on "gaming" machines (the actual ones, with only 8 cores). While I'm not recommending running at 100C and 6GHz for server workloads, the evidence that gaming CPUs are unstable under sustained load is anecdotal at best; the only story that comes to mind is a singular TR 3000 sample that failed AVX Prime95. There are plenty of professional applications that aren't infinitely scalable and need good single-threaded (or 4-core, 8-core, etc) performance. If you have evidence otherwise I'm sure I'm not the only person on [H] who would be interested...

It's so tasty when some dude with some gaming machines plays "systems engineer".

Wait.... actually- it's kind of cute....

From now on we'll just call you "Pookie".
I could fire so many insults back here its not even funny, but in the interest of not getting banned by Kyle I'm not going to.

Personal shittery aside, my point is we're now in a dilemma where there won't be a good high-end workstation solution for 6-12 months. If you buy into TR 5000 you're getting an obsolete socket and taking a 15% hit in IPC. If you buy Epyc 9004 you get modern IPC and platform features but take a 2GHz hit in clock speeds. If you go for one of the consumer platforms, you get state-of-the-art single-core performance for day-to-day work but fall 80 cores short of the state of the art (you also lose RDIMM support which is a big deal).

I'm sure TR 6000 will launch eventually but even then its a bit concerning, the amount of favored core tuning they do at the factory on the consumer parts is crazy (thanks competition!). It may be that the favored core can't run at 100% load for the warranty period at Tjmax, but then again I've yet to hear of any electromigration or other degradation at "stock" speeds on ADL.
 
A shame these don't have good turbo. The giant 9654 is a whole 2.1 GHz (!!!) below the 7950X and 2.3 GHz below the 13900KS which is an absurd amount of single-threaded performance to give up to get your 96 cores. I guess we're back to the days of needing two computers if you really want to go fast.
Eh. Different goals. The top end Epyc from Rome/etc also had shitty turbo speeds - ended up facepalming with a lot of customers that bought them for VDI and realized they performed like garbage for that task (Blast / HDX encoding is single threaded and highly clock-speed dependent - 2.5ghz is the ~minimum~ sustained for 30FPS on a desktop, as it also doesn't scale well with IPC improvements). That's why you wait for the other chips to come out - and pick the right one for the right job. Snowflakes exist in the DC as well (in terms of unique hardware and requirements, not the politically charged version of the word).
These go into server racks... database applications... virtualization.... containerization.... clusters. You know... You know: the big pieces of the internet. Parallelized workloads. Not your anemic/pathetic gaming machine.

Go ahead... fire up a serous database on your gaming machine. Do it... have fun. God help you if you actually put it under load. All hail the soft errors....
~looks at a couple of his gaming boxes that double as ESXi hosts running DBs~ ... ~looks confused~. Then again, I've built a "gaming box" with 3T of RAM, so... I'm weird? RAM capacities and the like are limited in consumer kit - but a core is a core for a lot of this stuff, and skylake plus (or Zen2+) will do just fine in a pinch for server workloads if you ask it to.
Well, obviously. It's a server processor. Also I will have you know one of my gaming machines has 64 cores in it so, there you go (it's not a very good gaming machine but is very fast for work).
Eww. :p
I've done plenty of serious work on "gaming" machines (the actual ones, with only 8 cores). While I'm not recommending running at 100C and 6GHz for server workloads, the evidence that gaming CPUs are unstable under sustained load is anecdotal at best; the only story that comes to mind is a singular TR 3000 sample that failed AVX Prime95. There are plenty of professional applications that aren't infinitely scalable and need good single-threaded (or 4-core, 8-core, etc) performance. If you have evidence otherwise I'm sure I'm not the only person on [H] who would be interested...
Yup. Lots out there that gives a shit about core speeds. DBs are a big one - aside from licensing by core, there's a reason that the "database specific" processors from both vendors tend to have much higher clock speeds - some transactional loads still depend on speed AND a broad base of cores to work with. VDI is another (and my specialty historically), as it's very clock speed dependent - but also scales linearly, so the idea of a 64core processor seems great (~192 desktops per socket!) until you realize that you have to somehow feed a GPU to all of those to make them usable (and that doesn't work - there aren't GRID GPUs dense enough for that even today, unless you're doing single-socket, in which case... just get 2x 32?). Web servers, basic app servers - those feed off of these massive CPUs, but there are others that do not.
Personal shittery aside, my point is we're now in a dilemma where there won't be a good high-end workstation solution for 6-12 months. If you buy into TR 5000 you're getting an obsolete socket and taking a 15% hit in IPC. If you buy Epyc 9004 you get modern IPC and platform features but take a 2GHz hit in clock speeds. If you go for one of the consumer platforms, you get state-of-the-art single-core performance for day-to-day work but fall 80 cores short of the state of the art (you also lose RDIMM support which is a big deal).
Eh. I've never been a fan of using Xeon/Epyc in workstations - I get why people do it, but as you said, it's often not the best solution. I have high hopes for Sapphire Rapids (want to replace at least one X299 box), and I'm dreaming of a non-pro threadripper 7000, but even if I can't get those, there will at least be a Xeon-W style of option at some point... I hope. If not, well - there will be non "high core" server CPUs too.

Please?
 
Yea... though I despise the guy doing the review.

In my company, I manage the servers- and we are running Gen 2 Epyc. It has performed very well.

We'll go Epyc in the next server I order as well.
That is refreshing. I heard the mentality a couple of years ago is nobody ever gets fired for ordering Intel so IT for most companies go with Intel even when AMD is outperforming.
I guess the performance gap has gotten so rediculously big people can't ignore EPYC anymore.
 
A shame these don't have good turbo. The giant 9654 is a whole 2.1 GHz (!!!) below the 7950X and 2.3 GHz below the 13900KS which is an absurd amount of single-threaded performance to give up to get your 96 cores. I guess we're back to the days of needing two computers if you really want to go fast.
Threadripper pro at the desk, Epyc in the server room seem to be the model, yes, I imagine server room hardware being different than desktop have historically being more the norm than the exception too ?

https://www.amd.com/en/press-releas...ro-5000-wx-series-processors-are-the-ultimate
“This year we've began migrating all global studios production systems to AMD, with EPYC processors powering our render farms and Threadripper PRO processors powering our workstations,” said Syama Mishra, Chief DCC Technology Officer, Binyan Studios

That said they do boost to 4.4ghz for those you need Epyc connectivity and high frequency on the lower core counts:

https://cdn.wccftech.com/wp-content...-9004-Zen-4-Genoa-CPUs-Official-Launch-13.png

Which could make the trick for a single machine in some use case.
 
Last edited:
That is refreshing. I heard the mentality a couple of years ago is nobody ever gets fired for ordering Intel so IT for most companies go with Intel even when AMD is outperforming.
I guess the performance gap has gotten so rediculously big people can't ignore EPYC anymore.
Trick is migrations - Intel to Intel for most environments is dead simple. Intel to AMD tends to require an outage - and depending on who you are, that can be a problem (even if it shouldn't be), so they default back to "buy what we can migrate to seamlessly." Even if that outage is in the order of seconds.
 
Well, obviously. It's a server processor. Also I will have you know one of my gaming machines has 64 cores in it so, there you go (it's not a very good gaming machine but is very fast for work).

Therefor not a gaming machine. That's a workstation- maybe.

I've done plenty of serious work on "gaming" machines (the actual ones, with only 8 cores). While I'm not recommending running at 100C and 6GHz for server workloads, the evidence that gaming CPUs are unstable under sustained load is anecdotal at best;

Never had to run a high availability workload 24/7/365 huh?

Please. Anecdotal. Any SE who's ever tried to cheap out with non enterprise server kit can tell you the horror stories. It's not anecdotal: It's fact.

You show me a gaming machine that can host 1TB of ECC RAM and I'll show you a server.

I could fire so many insults back here its not even funny, but in the interest of not getting banned by Kyle I'm not going to.

Listen Pookie... I'm criticizing your lame statements not attacking you personally. And be happy... Pookie is a cool name.

Personal shittery aside, my point is we're now in a dilemma where there won't be a good high-end workstation solution for 6-12 months. If you buy into TR 5000 you're getting an obsolete socket and taking a 15% hit in IPC. If you buy Epyc 9004 you get modern IPC and platform features but take a 2GHz hit in clock speeds. If you go for one of the consumer platforms, you get state-of-the-art single-core performance for day-to-day work but fall 80 cores short of the state of the art (you also lose RDIMM support which is a big deal).

I'm sure TR 6000 will launch eventually but even then its a bit concerning, the amount of favored core tuning they do at the factory on the consumer parts is crazy (thanks competition!). It may be that the favored core can't run at 100% load for the warranty period at Tjmax, but then again I've yet to hear of any electromigration or other degradation at "stock" speeds on ADL.

OK... please. Server processors are not a substitute for extreme HEDT setups. They just are not. You're trying to make the a point in an argument that has a false premise.

Boost clock on a basic 32 core, 64 thread, enterprise server? What for? I mean what is the point? None of the workloads are single thread sensitive. And most racked servers never run at their highest rated speed across more than a few cores... And those instances are sporadic s8ince threads are always being reassigned..

So your comments make no sense. This is enterprise computing. No one cares about boost clock. All we care about is threaded performance. Cores. Threads. That is all there is.
 
That is refreshing. I heard the mentality a couple of years ago is nobody ever gets fired for ordering Intel so IT for most companies go with Intel even when AMD is outperforming.
I guess the performance gap has gotten so rediculously big people can't ignore EPYC anymore.

Well, I'm director level. And while my hands are still on the systems themselves (by choice) I can call the shots.

Going to a systems integrator for commodity Supermicro kit, configuration, and burn in saved a lot of money compared to the corporate go-to vendors of Dell and HP. Additionally, an integrator, like Silicon Mechanics, gave me more choice over what went into the server. So there was better performance in the end product because it was tailored to our exact use case.

There was much rejoicing when my solution came in a lot less than Dell or HP per 2U deployment.

Epyc was a no brainer.
 
Threadripper pro at the desk, Epyc in the server room seem to be the model, yes, I imagine server room hardware being different than desktop have historically being more the norm than the exception too ?

https://www.amd.com/en/press-releas...ro-5000-wx-series-processors-are-the-ultimate
“This year we've began migrating all global studios production systems to AMD, with EPYC processors powering our render farms and Threadripper PRO processors powering our workstations,” said Syama Mishra, Chief DCC Technology Officer, Binyan Studios

That said they do boost to 4.4ghz for those you need Epyc connectivity and high frequency on the lower core counts:

https://cdn.wccftech.com/wp-content...-9004-Zen-4-Genoa-CPUs-Official-Launch-13.png

Which could make the trick for a single machine in some use case.

Right, that's what you're supposed to do, AMD is weird because they segment out workstation Epyc as a different brand with dedicated boards - Intel just lumps them all together as "3rd generation Scalable" and leaves it to the customer to figure out what they want. The issue is AMD launches their workstation-targeted SKUs a year after their rack-oriented SKUs, leaving us in this bad situation where Milan Threadripper barely launched into the channel and Genoa Epyc is rolling out already.

I was really hoping Sapphire Rapids would level the playing field and give us some competition in the high end again but looks like that's not happening for another generation.
 
  • Like
Reactions: DPI
like this
Right, that's what you're supposed to do, AMD is weird because they segment out workstation Epyc as a different brand with dedicated boards - Intel just lumps them all together as "3rd generation Scalable" and leaves it to the customer to figure out what they want. The issue is AMD launches their workstation-targeted SKUs a year after their rack-oriented SKUs, leaving us in this bad situation where Milan Threadripper barely launched into the channel and Genoa Epyc is rolling out already.

I was really hoping Sapphire Rapids would level the playing field and give us some competition in the high end again but looks like that's not happening for another generation.
Intel also has/had the Xeon-W line, which is both similar but different from full scalable - we just forget about them since they weren't marketed that much (and the boards were weird - the only really consumer oriented one was the Dominus Extreme, and you had to use a hair dryer to start it some days). They also had LGA2066 Xeon-W chips, so the lines REALLY blur at times on that side.
 
Well, I'm director level. And while my hands are still on the systems themselves (by choice) I can call the shots.

Going to a systems integrator for commodity Supermicro kit, configuration, and burn in saved a lot of money compared to the corporate go-to vendors of Dell and HP. Additionally, an integrator, like Silicon Mechanics, gave me more choice over what went into the server. So there was better performance in the end product because it was tailored to our exact use case.

There was much rejoicing when my solution came in a lot less than Dell or HP per 2U deployment.

Epyc was a no brainer.
That works when you have use cases that can’t be delivered in time frame or the vendor straight up doesn’t have what you need as a deliverable sku.

I’ve found some of the VARs the vendors use as local storage and showcases can be more flexible.
Especially when I’ve walked them thru a colo with a poc built and running.

The real issue is getting credible input about my future dumpster load of burned gear before the contracts have been signed.

That’s when I’ve learned to defer to the colo techs running Regions for public cloud providers and getting an up to date read on what is and isn’t resilient for a given use case.
 
Last edited:
I was really hoping Sapphire Rapids would level the playing field and give us some competition in the high end again but looks like that's not happening for another generation.

Sapphire Rapids feels like it's approaching Running Joke status at this point.

Right, that's what you're supposed to do, AMD is weird because they segment out workstation Epyc as a different brand with dedicated boards - Intel just lumps them all together as "3rd generation Scalable" and leaves it to the customer to figure out what they want. The issue is AMD launches their workstation-targeted SKUs a year after their rack-oriented SKUs, leaving us in this bad situation where Milan Threadripper barely launched into the channel and Genoa Epyc is rolling out already.
All of this. And this seems like such a common complaint among esoteric hardware enthusiast forums, and it's definitely one of my own frustrations, but then I've resigned that it's not that these corporations have 'abandoned' us so much as they're not going to abandon shareholders for us. Meaning HEDT/WS/DIY/Enthusiast segment is no longer a significant enough demographic not to shift resources elsewhere from.

And yet here we are almost 2023 and still stuck with an anemic 16-lanes for GPU, 4 lanes for an NVME SSD and 4 lanes for the chipset -- even on Z790/X670E. And the next best option for anyone not interested in splitting compute and storage into separate systems, is weird combinations of scarce WRX80 motherboards already abandoned by most AIB's, and $100-per-core Zen3 5000WX CPU's already obsoleted the moment the Lenovo exclusivity bullshit expired.

Maybe 2023 will finally be the magical year of > 24 PCIe lanes for under 10K.
 
Last edited:
Back
Top