AMD Ryzen Threadripper PRO 7995WX Emerges: 96 Cores, DDR5 Memory, and Over 5.0 GHz Boost Frequency

[Dr. Dro]

What? Quite LITERALLY current Zen3 (an architecture released November 2020) Threadripper Pro, is still faster than the brand new Sapphire Rapids HEDT/Workstation CPUs, so I would have to imagine that a Zen4 Threadripper will be performance leader for quite some time

That's quite situational, though. Zen 3 lacks Zen 4/SPR's advanced AVX-512 instructions entirely, which may very well matter to the target market (primarily VFX and renderfarms). Again keeping in mind that these aren't server chips so benching them as servers is counterproductive (there is Epyc for that).

 

[AnarchoPrimitiv]

Don't think these are gonna replace Sapphire Rapids chips anywhere, really... it's not in the same market segment, TR Pros aren't server chips. They fill this ultra-HEDT niche that persists despite the death of consumer-grade HEDT (since Intel hasn't released an HEDT chip since the i9-10980XE, which was already a third generation protracting the aging X299 chipset's service life and AMD abandoned Threadripper + aborted the TRX40 platform in the middle of its service life without a generational upgrade).

These are gonna be distributed only through OEMs in prebuilt workstations, perhaps with an exceptionally limited DIY channel release with stratospheric prices, just like the 5995WX.

They're talking about the Sapphire Rapids W-3XXX workstation/HEDT chips, not server chips. And they reason they're probably saying that is because if you, for example go to puget systems and check out their review of the Sapphire Rapids HEDT/Workstation Chips from a few months ago, you'll see that Zen3 Threadripper Pro is already faster, so it goes to reason that if a Threadripper Pro chip based on a 2-1/2 year old architecture is already faster than Intel's best, then a Zen4 Threadripper Pro should be light-years ahead

That's quite situational, though. Zen 3 lacks Zen 4/SPR's advanced AVX-512 instructions entirely, which may very well matter to the target market (primarily VFX and renderfarms). Again keeping in mind that these aren't server chips so benching them as servers is counterproductive (there is Epyc for that).

Who is talking about server chips? I'm not, and haven't so I'm confused as to why you keep mentioning epyc.

Well, at least in content creation, puget systems says Zen3 Threadripper Pro is better:

"Overall, the new Intel Xeon W-3400 series of processors are significantly faster than the previous generation, but this comes at the cost of much higher power consumption. And while the gen-over-gen performance is great in many cases, outside of a few isolated workloads, it is only enough to (at best) bring Intel on par with AMD’s Threadripper PRO 5000 WX-Series processors. In most cases, AMD maintains a small but measurable performance edge. This, combined with AMD’s lower power consumption, is going to make them a more attractive option than the new Xeon W-3400 processors in most content creation workflows."

 

[Assimilator]

People, I don't give a crap about most of TR's functionality.

I don't care about 20 billion CPU cores.
I don't care about the ability to support 30 sticks or 90 billion petabytes of RAM with ECC support.
I don't care about 300 SATA ports.
I don't care about 15 10GbE interfaces.

All I care about are at least two x16 PCIe slots, and lots of USB ports. That's it, yet it's something that no non-"HEDT" motherboard offers. Because of artificial market segmentation.

 

[Makaveli]

This is a workstation CPU and not lower tier HEDT.
Those who need it and can afford it will know its value in daily work.


Not necessarily. Sapphire Rapids has both HEDT (W2400) and workstation (W3400) segments. AMD is so far leaking only workstation, higher tier products, with 8 channels. If they do not release 4 channel HEDT SKUs, they will not have a competitor here.

I'm fairly certain there is two sku's for TR based on Zen 4.

a TR Pro that will be oem only and has 8 channel memory.

Regular TR with 4 channel memory.

 

[Dr. Dro]

Who is talking about server chips? I'm not, and haven't so I'm confused as to why you keep mentioning epyc.

I see, I just looked it up and indeed they released the w9-3495X to the workstation segment. I thought that they only had released SPR-112 to server segment. Cool beans. I stand corrected, did not think those had launched yet.

$5889. Gonna sell well with TR Pro's giving them that much of a beat down... no wonder it flew under my radar

 

[Dirt Chip]

Well, something need to eat through that 16k RAW video somehow..

This is AMD real advantage, let it shine!

 

[phanbuey]

Yes. 1000 to 3000-series non-PRO Threadrippers.

Not since they've taken the crown and abandoned that segment.

TBF they moved the consumer class into that segment, and finally broke the quad core stagnation, but still would be interesting to have that segment.

I wish companies were honest about why they cut certain segments -- would be interesting to see why they discontinued non-pro threadrippers.

 

[Dr. Dro]

Not since they've taken the crown and abandoned that segment.

TBF they moved the consumer class into that segment, and finally broke the quad core stagnation, but still would be interesting to have that segment.

I wish companies were honest about why they cut certain segments -- would be interesting to see why they discontinued non-pro threadrippers.

Answer is simple, it's money. This isn't cynicism, it's just market conditions. It took the same silicon to build a Threadripper 3990X (which would sell at a stratospheric $3,990 marketing stunt MSRP - far too high for consumers, comparably low for businesses), but the equivalent Epyc 7742 sold for $6,950 MSRP. Yet worse, the Threadripper had to clock reasonably high, the 7742? It's a 2.25 GHz CPU with a boost that barely reached the 3 GHz mark, so much, much easier to get dies for it.

This coincided with the pandemic and the demand for server hardware was absolutely enormous with the whole work from home thing, hosting, cloud and service companies were buying up these CPUs faster than AMD could make them and paying full price for them, too. Sometimes even more, so why would AMD allocate any silicon to this segment? It sells relatively poorly, has very high maintenance costs... best to just discontinue, pretend that TRX40 never existed, and ignore/deflect the complaints of anyone who was unfortunate enough to have bought into this platform. Which is exactly what AMD has done.

 

[vmarv]

What a monster! I doubt that this will become a 50€ chip anytime soon.

 

[phanbuey]

Answer is simple, it's money. This isn't cynicism, it's just market conditions. It took the same silicon to build a Threadripper 3990X (which would sell at a stratospheric $3,990 marketing stunt MSRP - far too high for consumers, comparably low for businesses), but the equivalent Epyc 7742 sold for $6,950 MSRP.

This coincided with the pandemic and the demand for server hardware was absolutely enormous with the whole work from home thing, hosting, cloud and service companies were buying up these CPUs faster than AMD could make them and paying full price for them, too. Sometimes even more, so why would AMD allocate any silicon to this segment? It sells relatively poorly, has very high maintenance costs... best to just discontinue, pretend that TRX40 never existed, and ignore/deflect the complaints of anyone who was unfortunate enough to have bought into this platform. Which is exactly what AMD has done.

Right that makes sense, I just wish they would say it.

"Sorry losers you're too cheap, and we don't really need the market or mind share anymore" etc. -- they don't have to word it that way but the "Oh well the 5950x has 16 cores so that's the new HEDT" just made me feel like I was getting lied to.

Same for the gaslit excuse we will get for no high end Radeons next cycle because all the silicon is getting dumped into datacenter? No idea.

 

[HBSound]

It would be great if small form factor motherboards could support Threadripper Pro. While big motherboard vendors support Intel Xeon in all shapes and sizes, the same cannot be said for AMD. It would be nice to see some love for the AMD community as well.

 

[Dr. Dro]

Right that makes sense, I just wish they would say it.

"Sorry losers you're too cheap, and we don't really need the market or mind share anymore" etc. -- they don't have to word it that way but the "Oh well the 5950x has 16 cores so that's the new HEDT" just made me feel like I was getting lied to.

Same for the gaslit excuse we will get for no high end Radeons next cycle because all the silicon is getting dumped into datacenter? No idea.

That's precisely because you were.

The same way that they initially gaslit X470 owners saying they wouldn't get Zen 3, and lied to X370 owners that the BIOS size was too small (see: all it took was Alder Lake destroying the budget-end for Ryzen for them to backtrack and start releasing CPUs below the $300 5600X and re-launch their previously unlaunched 5700X), the key lesson to learn from all this is that AMD is a multi-billion-dollar corporation, not an underdog worthy of being our darling. The second they get reasonably ahead you better believe that all of the consumer-friendly façade will subside and they will start charging and pulling products to maximize their profits with the same shameless greed that Intel and especially Nvidia displays when they're ahead. It just happens they're ahead most of the time, while Intel... isn't, the 13900K/KS and the 7950X/X3D usually tie in amount of wins, some workloads favoring each design but never a step ahead of the other.

 

[Tek-Check]

People, I don't give a crap about most of TR's functionality.

I don't care about 20 billion CPU cores.
I don't care about the ability to support 30 sticks or 90 billion petabytes of RAM with ECC support.
I don't care about 300 SATA ports.
I don't care about 15 10GbE interfaces.

All I care about are at least two x16 PCIe slots, and lots of USB ports. That's it, yet it's something that no non-"HEDT" motherboard offers. Because of artificial market segmentation.

If you don't care, why are you posting in TR PRO article?
On consumer motherboards, we already have two x16 slots working as x8/x8 PCIe 5.0!!!

I'm fairly certain there is two sku's for TR based on Zen 4.
a TR Pro that will be oem only and has 8 channel memory.
Regular TR with 4 channel memory.

Yes, but there hes not been any benchmark leak of any non-PRO TR with 4 channels so far. Those are either going to come out later or not at all.

 

[Patriot]

Right that makes sense, I just wish they would say it.

"Sorry losers you're too cheap, and we don't really need the market or mind share anymore" etc. -- they don't have to word it that way but the "Oh well the 5950x has 16 cores so that's the new HEDT" just made me feel like I was getting lied to.

Same for the gaslit excuse we will get for no high end Radeons next cycle because all the silicon is getting dumped into datacenter? No idea.

It is annoying and frustrating, that said there is still a manufacturing shortage, the data centers eat up everything, so if a consumer doesn't like a launched price of something, or doesn't buy something, they will reduce output or not make a next version of it... and sell more server chips and or GPGPU accelerators. AMD didn't make a Zen3 threadripper non-pro and it pisses me off... but almost no one would have bought it. It was too late in the cycle.

 

[Dr. Dro]

If you don't care, why are you posting in TR PRO article?
On consumer motherboards, we already have two x16 slots working as x8/x8 PCIe 5.0!!!

The only problem with this logic is that even though modern platforms support bifurcation, devices are still restricted to their generational bandwidth.

A bifurcated Gen 5 x4 link with a Gen 3 device will still result in a Gen 3 x2 + whatever x2 link; and performance on existing devices will still degrade. Which is why this is less optimal than having a healthy supply of physical lanes available, something that is not even on advanced creator motherboards such as my MSI MEG Z690 ACE.

Note the liberal use of up to, for example, M2_4 slot bandwidth is shared with PCIE_3, x2 each or x4 fully directed to each slot, disabling the other. There's simply not enough resources in the system to enable all ports at once.

 

[Patriot]

The only problem with this logic is that even though modern platforms support bifurcation, devices are still restricted to their generational bandwidth.

A bifurcated Gen 5 x4 link with a Gen 3 device will still result in a Gen 3 x2 + whatever x2 link; and performance on existing devices will still degrade. Which is why this is less optimal than having a healthy supply of physical lanes available, something that is not even on advanced creator motherboards such as my MSI MEG Z690 ACE.

View attachment 309570

Note the liberal use of up to, for example, M2_4 slot bandwidth is shared with PCIE_3, x2 each or x4 fully directed to each slot, disabling the other. There's simply not enough resources in the system to enable all ports at once.

A switch would add latency but solve this. PLX switches disappeared from consumer boards when Broadcom acquired them and jacked the prices up 4x.
But other options like Microchip exist... that said the current chipset for amd is essentially a pcie switch...

 

[Tek-Check]

The only problem with this logic is that even though modern platforms support bifurcation, devices are still restricted to their generational bandwidth.

A bifurcated Gen 5 x4 link with a Gen 3 device will still result in a Gen 3 x2 + whatever x2 link; and performance on existing devices will still degrade. Which is why this is less optimal than having a healthy supply of physical lanes available, something that is not even on advanced creator motherboards such as my MSI MEG Z690 ACE.

Devices are by default restricted by their generational bandwidth. People need to choose their devices carefully.

Note the liberal use of up to, for example, M2_4 slot bandwidth is shared with PCIE_3, x2 each or x4 fully directed to each slot, disabling the other. There's simply not enough resources in the system to enable all ports at once.

Glass of water is either half empty or half full. It's a metter of perception. Z690 chipset provides certain number of lanes by design. Motherboard vendors often create more interfaces with shared lanes in order to give choice to consumers. You have a choice in this case to attach a full x4 PCIe device or yet another M.2 drive. You can even do both, at half speed.

If you know what the chipset provides in the spec, you will know what to expect on a motherboard, without being surprised if there is overprovision of interfaces with shared lanes. Simple.

 

[Dr. Dro]

Devices are by default restricted by their generational bandwidth. People need to choose their devices carefully.

Glass of water is either half empty or half full. It's a metter of perception. Z690 chipset provides certain number of lanes by design. Motherboard vendors often create more interfaces with shared lanes in order to give choice to consumers. You have a choice in this case to attach a full x4 PCIe device or yet another M.2 drive. You can even do both, at half speed.

If you know what the chipset provides in the spec, you will know what to expect on a motherboard, without being surprised if there is overprovision of interfaces with shared lanes. Simple.

Normally I agree, but like, replacing your entire stock of SSDs just to optimize bandwidth allocation (work around platform restrictions) is quite the costly endeavor

It's usually something we happily carry on from a previous build, and NVMe drives have been around for long enough that unless one keeps their machine for a decade plus, chances are you already have a NVMe M.2 drive with some of your stuff onto it.

 

[A Computer Guy]

We're going to need a bigger motherboard...

What! I was hoping AsRock would make an STX motherboard version for this CPU.

People, I don't give a crap about most of TR's functionality.

I don't care about 20 billion CPU cores.
I don't care about the ability to support 30 sticks or 90 billion petabytes of RAM with ECC support.
I don't care about 300 SATA ports.
I don't care about 15 10GbE interfaces.

All I care about are at least two x16 PCIe slots, and lots of USB ports. That's it, yet it's something that no non-"HEDT" motherboard offers. Because of artificial market segmentation.

I'm pining for more PCIe slots too and it would be nice if they put two good CCD's in the dual CCD's chips (instead of one good + meh).
The could even classify it LHEDT. (Lightweight HEDT+)
For example: 8950x LHEDT+ (Lisa Sue Special Edition)

It would be great if small form factor motherboards could support Threadripper Pro. While big motherboard vendors support Intel Xeon in all shapes and sizes, the same cannot be said for AMD. It would be nice to see some love for the AMD community as well.

I thought there was an ITX threadripper board? Now I have to check just because curiosity.

(edit) I found this https://www.asrockrack.com/general/productdetail.asp?Model=ROMED4ID-2T#Specifications
(edit2 - oof sorry that was EPYC - those darn sockets all look the same to me)

and found this but unfortunately not a real product.

 

[HBSound]

I thought there was an ITX threadripper board? Now I have to check just because curiosity.

(edit) I found this https://www.asrockrack.com/general/productdetail.asp?Model=ROMED4ID-2T#Specifications
(edit2 - oof sorry that was EPYC - those darn sockets all look the same to me)

and found this but unfortunately not a real product.
View attachment 309613

Exactly, I have looked all over the globe. No one, has anything. I saw the older PCI G3 PCI lane MTAX motherboard, but now I must step back in time.

Thank you so much.

 

[Tek-Check]

Normally I agree, but like, replacing your entire stock of SSDs just to optimize bandwidth allocation (work around platform restrictions) is quite the costly endeavor

I don't like overstuffed motherboards.

It's usually something we happily carry on from a previous build, and NVMe drives have been around for long enough that unless one keeps their machine for a decade plus, chances are you already have a NVMe M.2 drive with some of your stuff onto it.

There's already too many NVMe slots on increasing number of boards, in expense of PCIe slots. I'd prefer to put M.2 AIC into second x8 slit and call it a day

 

[Assimilator]

There's already too many NVMe slots on increasing number of boards, in expense of PCIe slots. I'd prefer to put M.2 AIC into second x8 slit and call it a day

I have a quad M.2 AIC. BUT I CAN'T USE IT unless I'm willing to put it in the primary PCIe slot and then move my GPU to the second slot WHICH IS ONLY x4. Because AM4 was designed by chumps and AM5 is the same.

 

[TumbleGeorge]

Because AM4 was designed by chumps and AM5 is the same.

Then they have to hire someone to teach them how to do it right. Are you available for this job?

 

[Tek-Check]

I have a quad M.2 AIC. BUT I CAN'T USE IT unless I'm willing to put it in the primary PCIe slot and then move my GPU to the second slot WHICH IS ONLY x4. Because AM4 was designed by chumps and AM5 is the same.

That depends on a board. On better boads, x16 slot 1 and slot 2 operate as x8 when both are populated, for example on my B550 ProArt has that solution.

The same applies for AM5 on most Asrock and Asus boards, but not Gigabyte boards that mostly come with one Gen5 x16 slot.

Buying a board with second slot wired x4 is asking for trouble if you want to connect more ambitious AIC, such as quad NVMe. It is your responsibility to match appropriate board with appropriate AIC, so that they both match in capability.

AIC vendors know that the best desktop boards can offer is x8 on the second slot osmf the first is populated with GPU, and they should be designing AIC with x8 electrical interface rather than releasing only x16 versions for workstations. So, there is that too.

 

[Dr. Dro]

That depends on a board. On better boads, x16 slot 1 and slot 2 operate as x8 when both are populated, for example on my B550 ProArt has that solution.

The same applies for AM5 on most Asrock and Asus boards, but not Gigabyte boards that mostly come with one Gen5 x16 slot.

Buying a board with second slot wired x4 is asking for trouble if you want to connect more ambitious AIC, such as quad NVMe. It is your responsibility to match appropriate board with appropriate AIC, so that they both match in capability.

AIC vendors know that the best desktop boards can offer is x8 on the second slot osmf the first is populated with GPU, and they should be designing AIC with x8 electrical interface rather than releasing only x16 versions for workstations. So, there is that too.

That still doesn't avoid the fact that if the main link is x8, it can only bifurcate into four x2 links, which would still reduce the SSD's maximum performance and incur the generational penalty to older devices. So we're back to square one, it's either two x4 or four x2, and there's no way to work around that unless a x16 link to the chipset or CPU is available