Intel Readies Xeon W-2500 Series with 4-channel Memory to Square Off Against Threadripper 7000

[Chaitanya]

Hahaha! Better to have what you would actually use. Higher number in this case is not necessarily better.
Your future 5090, and then 6090, will never saturate the bandwidth of x16 Gen4 slot, let alone Gen5 slot. Rest assured it will not happen. PCIe Gen5 arrived to client and HEDT segment from server too soon. It has been largely a gimmick to feed our Ego.

Motheboards have had x16 Gen5 slot since Alder Lake in 2021. Two years later, nobody in the world is using the bandwidth capability of that slot. It's silly, isn't it? Even next year, when AMD and Nvidia release GPUs with PCIe Gen5 support, those GPUs will not saturate what Gen4 currently provides. It will take another 3-4 generations of GPUs to start using meaningfully what x16 Gen5 slot has to offer. By that time, you will have a new platform.

By having two DIMMs per channel, your RAM speed will drop below basic 4800 MT/s DDR5 speed. Something to have in mind about disadvantages of having two DIMMs per channel. There will be 64GB and 96GB memory sticks next year, as Samsung has developed 32Gbit RAM modules, so one DIMM per channel should be enough, plus its higher speeds.

So, 64 Gen5 lanes on Sapphire/Emerald Rapids is a total overkill driving away attention from what actually matters, which is CPU cores and performance. Zen4 TR is unrivaled here.

Even 48 Gen5 lanes on TRX50 is overkill for a few years to come, as the only device that could actually utilize this bandwidth are x16 Gen5 AIC with four Gen5 SSDs. Now, do you really need Gen5 SSDs? That's another question.

The most important question is whether you plan CPU upgrade. If so, Emerald Rapids is the end of the road. That would be one-off purchase for several years. TRX50 and WRX90 platform should receive Zen5 TR upgrade, a simple CPU swap and easy upgrade without buying completely new system.

Also AMD does have massive advantage when it comes to lower power consumption unlike Xeon-W which suck electricity like there is no tomorrow for ~10% lower performance(against Pro-5000 series). I would really like to read how new TR 7000 compare against Xeon-W.

 

[unwind-protect]

Hahaha! Better to have what you would actually use. Higher number in this case is not necessarily better.
Your future 5090, and then 6090, will never saturate the bandwidth of x16 Gen4 slot, let alone Gen5 slot.[....]

For games, no. But if you have compute payloads on multiple GPUs those now have to communicate through PCIe since 4090 doesn't have NVLink anymore (why, anyway?).

 

[1d10t]

8 core 16 threads for workstation? In 2023 ?

 

[thestryker6]

Hahaha! Better to have what you would actually use. Higher number in this case is not necessarily better.
Your future 5090, and then 6090, will never saturate the bandwidth of x16 Gen4 slot, let alone Gen5 slot. Rest assured it will not happen. PCIe Gen5 arrived to client and HEDT segment from server too soon. It has been largely a gimmick to feed our Ego.

Somehow you don't seem to grasp that 64 lanes is physically more than 48 and that PCIe revision doesn't even need to come into play. It's not about bandwidth it's about phsycial lanes as you can't just make PCIe 5.0 lanes into 3.0/4.0 lanes without a PLX switch.

By having two DIMMs per channel, your RAM speed will drop below basic 4800 MT/s DDR5 speed. Something to have in mind about disadvantages of having two DIMMs per channel. There will be 64GB and 96GB memory sticks next year, as Samsung has developed 32Gbit RAM modules, so one DIMM per channel should be enough, plus its higher speeds.

There are already 128GB RDIMMs (which is what Zen 4 TR and SPR/EMR Xeons use), but they're insanely expensive. ECC DDR5 just started using 24gbit IC, but these sticks appear to be limited to OEM/business as they're not on the consumer market so 32gbit based ones will be a long while. Not to mention every RDIMM over 5600 currently only comes in 16GB capacity (and even the 5600 availability is bad if you want more than 32GB). 1DPC is definitely a liability given the availability of RDIMMs unless things rapidly shift when the new TR hit the open market.

 

[Tek-Check]

For games, no. But if you have compute payloads on multiple GPUs those now have to communicate through PCIe since 4090 doesn't have NVLink anymore (why, anyway?).

There is no evidence that compute payloads need x16 Gen5 bandwidth to run fully on desktop and HEDT systems, even without NVLink.

Guide to PCIe Lanes: How many do you need for your workload?

PCIe Lanes can severly bottleneck your performance if you don't have enough for your workload. Here's all you need to know about PCIe Lanes.

www.cgdirector.com

Best Hardware for GPU Rendering in Octane - Redshift - Vray (Updated)

Here's everything you need to know about the Best Hardware for GPU Rendering in Octane, Redshift, V-Ray and other modern GPU Render Engines.

www.cgdirector.com

For compute workloads on GPU, it's far more important that GPU has enough bus width and VRAM capacity to process workloads within itself than bandwidth needed to send and receive queries to/from CPU and other peripherals.

Somehow you don't seem to grasp that 64 lanes is physically more than 48 and that PCIe revision doesn't even need to come into play. It's not about bandwidth it's about phsycial lanes as you can't just make PCIe 5.0 lanes into 3.0/4.0 lanes without a PLX switch.

Somehow you forgot that TRX50 platform has more physical PCIe lanes directly wired with CPU than Sapphire/Emerald Rapids, which deals with any bottleneck created by traffic via chipset. On TRX50, only necessary I/O peripherals are wired into chipset, everything else has a direct access to CPU.

As said, both platforms are overkill in terms of Gen5 lanes, Intel's platform more so. It's a marketing gimmick to drive away attention from simple fact that Emerald Rapids is end-of-life platform with no CPU upgrades.

 

[lexluthermiester]

the W5-2545

FINALLY!! Was about to pull the trigger on a Threadripper until I saw this. Let's see what the prices will be though!

 

[thestryker6]

Somehow you forgot that TRX50 platform has more physical PCIe lanes directly wired with CPU than Sapphire/Emerald Rapids, which deals with any bottleneck created by traffic via chipset. On TRX50, only necessary I/O peripherals are wired into chipset, everything else has a direct access to CPU.

The only board I'd paid attention to PCIe wise was Gigabyte's and clearly I should have looked at ASRock/Asus. I didn't realize they were each using 2 PCIe 4.0 x16 lanes (Gigabyte only uses 1) from the chipset on their boards.

 

[thesmokingman]

Also AMD does have massive advantage when it comes to lower power consumption unlike Xeon-W which suck electricity like there is no tomorrow for ~10% lower performance(against Pro-5000 series). I would really like to read how new TR 7000 compare against Xeon-W.

Don't forget the optional under table chiller!!

 

[Tek-Check]

Don't forget the optional under table chiller!!

New Noctua for TR will do the job just fine. Check it out.

 

[thesmokingman]

New Noctua for TR will do the job just fine. Check it out.

Think you missed the joke.

Intel: We 'Forgot' to Mention 28-Core, 5-GHz CPU Demo Was Overclocked

We met with Intel and asked a few pointed questions about its now-infamous 28-core CPU demo.

www.tomshardware.com

 

[Tek-Check]

I'm mostly curious if Intel is going to change pricing at all since the 7960X MSRP is $1500 and the 2495X is $2200. Intel's main advantage here is 2DPC (RDIMMs are extremely expensive as capacity rises) and the extra PCIe lanes, but that and 2 extra CPU cores don't fill the price gap in my opinion.

2DPC is not really a significant advantage, as it runs slower, at 4400 MT/s. W790 platform has literally x4 Gen4 lanes more, which is negligible. And the price of this Xeon looks very silly now that TR CPUs have been announced. Those Xeons will quickly go down in price even before Emerals Rapids launches.

The only board I'd paid attention to PCIe wise was Gigabyte's and clearly I should have looked at ASRock/Asus. I didn't realize they were each using 2 PCIe 4.0 x16 lanes (Gigabyte only uses 1) from the chipset on their boards.

I am very disappointed by Asrock's TRX50 board. Nonsense release, terrible rear I/O, lack of ambition, horrible fans. They should sack the board chief designer. They used to have brilliant Creator line of boards a few years ago. Not anymore. Asrock, are you ok?

Gigabyte Aero D is decent, I/O with Thunderbolt4/USB4 ports, DP IN is good, WiFi7 is good, but it needs at least one more PCIe slot.

Asus TRX50 also has unambitious rear I/O, no USB4, no DP IN, PCIe slots are better.

I would not buy any of those three boards. All vendors need to release something more mature, with better I/O. They treated TRX50 platform as guiney pig. Boards are not mature.

 

[thestryker6]

2DPC is not really a significant advantage, as it runs slower, at 4400 MT/s. W790 platform has literally x4 Gen4 lanes more, which is negligible. And the price of this Xeon looks very silly now that TR CPUs have been announced. Those Xeons will quickly go down in price even before Emerals Rapids launches.

It is if you need capacity and are on a budget as 512GB with 64GB RDIMMs is about $2000 whereas with 128GB RDIMMs it's about $5600. That being said I absolutely agree that the 24C Zen 4 TR makes the Intel pricing look like a joke which is why I said the features don't fill the gap.

I am very disappointed by Asrock's TRX50 board. Nonsense release, terrible rear I/O, lack of ambition, horrible fans. They should sack the board chief designer. They used to have brilliant Creator line of boards a few years ago. Not anymore. Asrock, are you ok?

Gigabyte Aero D is decent, I/O with Thunderbolt4/USB4 ports, DP IN is good, WiFi7 is good, but it needs at least one more PCIe slot.

Asus TRX50 also has unambitious rear I/O, no USB4, no DP IN, PCIe slots are better.

I would not buy any of those three boards. All vendors need to release something more mature, with better I/O. They treated TRX50 platform as guiney pig. Boards are not mature.

This is just a guess on my part because I haven't seen a chipset diagram, but the TB/USB4 might be the price paid for the extra PCIe 4.0 x16 slot.

I agree with the really weird tons of fans design out of ASRock though it's super confusing.

 

[chrcoluk]

Do they supply a portable nuclear plant to keep it powered as well? Although been serious this will probably be way more efficient than the normal chips as likely wont have the crazy factory overclock configuration.

 

[Tek-Check]

This is just a guess on my part because I haven't seen a chipset diagram, but the TB/USB4 might be the price paid for the extra PCIe 4.0 x16 slot.

TB4 chip connects to x4 Gen3 PHY and USB4 connects to x4 Gen4 PHY, so it should not be a trade off against seperate PCIe slot, as there are plenty of lanes.

 

[lemans24]

Hahaha! Better to have what you would actually use. Higher number in this case is not necessarily better.
Your future 5090, and then 6090, will never saturate the bandwidth of x16 Gen4 slot, let alone Gen5 slot. Rest assured it will not happen. PCIe Gen5 arrived to client and HEDT segment from server too soon. It has been largely a gimmick to feed our Ego.

Motheboards have had x16 Gen5 slot since Alder Lake in 2021. Two years later, nobody in the world is using the bandwidth capability of that slot. It's silly, isn't it? Even next year, when AMD and Nvidia release GPUs with PCIe Gen5 support, those GPUs will not saturate what Gen4 currently provides. It will take another 3-4 generations of GPUs to start using meaningfully what x16 Gen5 slot has to offer. By that time, you will have a new platform.

By having two DIMMs per channel, your RAM speed will drop below basic 4800 MT/s DDR5 speed. Something to have in mind about disadvantages of having two DIMMs per channel. There will be 64GB and 96GB memory sticks next year, as Samsung has developed 32Gbit RAM modules, so one DIMM per channel should be enough, plus its higher speeds.

So, 64 Gen5 lanes on Sapphire/Emerald Rapids is a total overkill driving away attention from what actually matters, which is CPU cores and performance. Zen4 TR is unrivaled here.

Even 48 Gen5 lanes on TRX50 is overkill for a few years to come, as the only device that could actually utilize this bandwidth are x16 Gen5 AIC with four Gen5 SSDs. Now, do you really need Gen5 SSDs? That's another question.

The most important question is whether you plan CPU upgrade. If so, Emerald Rapids is the end of the road. That would be one-off purchase for several years. TRX50 and WRX90 platform should receive Zen5 TR upgrade, a simple CPU swap and easy upgrade without buying completely new system.

I want 64 PCIE Gen 5.0 lanes because I want to run multiple graphics cards and especially two or more M.2 Gen5 Quad cards.

PCIE 5.0 will also help with latency and less CPU usage when moving data from GPU to host memory if you are doing high performance CUDA programming, which is the only reason I am using a 4090 card
that I have already. This is definitely not a gaming machine but for AI/algorithmic options dev/trading which is higly sensitive to system latency.

Since I plan on keeping this machine for at least another 5+ years, I want the best performance that I can afford with todays pricing.
The threadripper 7000x HEDT series is way better overall than the w-2500 series except for one chip, w-2495x.
Starting with 4 sticks of memory and then upgrading to another 4 sticks of memory without throwing out memory is well worth it to me.
For 32 cores or more, threadripper Pro or HEDT is the only choice but I will see what the best price performance is for 256GB memory (dont really care about mem speed but price is important) between
the threadripper 7960x vs w-2595x.

I have never upgraded my motherboard to another chip but always waited until performance has at least doubled or more which invariable is at least 2 generations.
Even if w-2595x is a dead end, it will meet my needs for the next 5 years for dev work.

Definitely will compare next year when both are available for retail order and IF i can afford 32 cores, will definitely go with threadripper 7975x...

 

[R-T-B]

I agree. HEDT is all about cores and maybe cores/$.

I mean, it's also about IO but I hear you, that usually plays second fiddle to core count.

 

[lemans24]

There is no evidence that compute payloads need x16 Gen5 bandwidth to run fully on desktop and HEDT systems, even without NVLink.

Guide to PCIe Lanes: How many do you need for your workload?

PCIe Lanes can severly bottleneck your performance if you don't have enough for your workload. Here's all you need to know about PCIe Lanes.

www.cgdirector.com

Best Hardware for GPU Rendering in Octane - Redshift - Vray (Updated)

Here's everything you need to know about the Best Hardware for GPU Rendering in Octane, Redshift, V-Ray and other modern GPU Render Engines.

www.cgdirector.com

For compute workloads on GPU, it's far more important that GPU has enough bus width and VRAM capacity to process workloads within itself than bandwidth needed to send and receive queries to/from CPU and other peripherals.


Somehow you forgot that TRX50 platform has more physical PCIe lanes directly wired with CPU than Sapphire/Emerald Rapids, which deals with any bottleneck created by traffic via chipset. On TRX50, only necessary I/O peripherals are wired into chipset, everything else has a direct access to CPU.
View attachment 320185
As said, both platforms are overkill in terms of Gen5 lanes, Intel's platform more so. It's a marketing gimmick to drive away attention from simple fact that Emerald Rapids is end-of-life platform with no CPU upgrades.

Ok. The workstation I am building is NOT for gaming but strictly for algorithmic options dev/trading!!

I run risk management every time my trade changes value and it takes around 100ms round trip to send/compute/receive using a 4090 card.
The transfer time for send/receive data alone is over 90% of that time and that is with a Gen 3.0 PCIE x16 computer (1950x threadripper 16 core circa 2017).
The 4090 is incredibly fast but I am sure i will improve latency by running it with Gen 5.0 even with 4090 which is Gen4 and especially when 5090 Gen5 x16 comes out in 2025.

Ideally I would like to get this down to 1ms round trip within the next few years if possible with more optimization and at least 4 gpu (5090+) cards running on a single dedicated server.

2DPC is not really a significant advantage, as it runs slower, at 4400 MT/s. W790 platform has literally x4 Gen4 lanes more, which is negligible. And the price of this Xeon looks very silly now that TR CPUs have been announced. Those Xeons will quickly go down in price even before Emerals Rapids launches.


I am very disappointed by Asrock's TRX50 board. Nonsense release, terrible rear I/O, lack of ambition, horrible fans. They should sack the board chief designer. They used to have brilliant Creator line of boards a few years ago. Not anymore. Asrock, are you ok?

Gigabyte Aero D is decent, I/O with Thunderbolt4/USB4 ports, DP IN is good, WiFi7 is good, but it needs at least one more PCIe slot.

Asus TRX50 also has unambitious rear I/O, no USB4, no DP IN, PCIe slots are better.

I would not buy any of those three boards. All vendors need to release something more mature, with better I/O. They treated TRX50 platform as guiney pig. Boards are not mature.

You can go upto 6000 MT/s DDR5 ecc rdimm on overclocked 2DPC w-2495x today with Asus/Asrock w-790 motherboards.
This definitely could be a significant/only advantage over 1DPC for threaderipper 7000x HEDT series depending on price/performance.
If I can afford 32 core, will definitely go with threadripper 7970x but w-2595x is still a good possibility for my needs.

 

[unwind-protect]

I want 64 PCIE Gen 5.0 lanes because I want to run multiple graphics cards and especially two or more M.2 Gen5 Quad cards.

PCIE 5.0 will also help with latency and less CPU usage when moving data from GPU to host memory if you are doing high performance CUDA programming, which is the only reason I am using a 4090 card
that I have already. This is definitely not a gaming machine but for AI/algorithmic options dev/trading which is higly sensitive to system latency.

Since I plan on keeping this machine for at least another 5+ years, I want the best performance that I can afford with todays pricing.
The threadripper 7000x HEDT series is way better overall than the w-2500 series except for one chip, w-2495x.
Starting with 4 sticks of memory and then upgrading to another 4 sticks of memory without throwing out memory is well worth it to me.
For 32 cores or more, threadripper Pro or HEDT is the only choice but I will see what the best price performance is for 256GB memory (dont really care about mem speed but price is important) between
the threadripper 7960x vs w-2595x.

I have never upgraded my motherboard to another chip but always waited until performance has at least doubled or more which invariable is at least 2 generations.
Even if w-2595x is a dead end, it will meet my needs for the next 5 years for dev work.

Definitely will compare next year when both are available for retail order and IF i can afford 32 cores, will definitely go with threadripper 7975x...

Why not look into EPYC systems since you seem not to need too much CPU speed, with the GPU doing the heavy lifting for you?

 

[Tek-Check]

I have never upgraded my motherboard to another chip but always waited until performance has at least doubled or more which invariable is at least 2 generations.
Even if w-2595x is a dead end, it will meet my needs for the next 5 years for dev work.

Definitely will compare next year when both are available for retail order and IF i can afford 32 cores, will definitely go with threadripper 7975x...

You also want to check which platform performs better in specific workloads you need it for. It might be the case that Intel is on par of even faster in those specific workloads, despite the fact that TR is overall much faster.

I run risk management every time my trade changes value and it takes around 100ms round trip to send/compute/receive using a 4090 card.

Is this not linked to network latency?

 

[lemans24]

You also want to check which platform performs better in specific workloads you need it for. It might be the case that Intel is on par of even faster in those specific workloads, despite the fact that TR is overall much faster.

Is this not linked to network latency?

I should have been more clear regarding risk management with 4090.
I run my risk management calculations directly on the 4090 using Monte Carlo simulations which is pure math compute but there is some latency between the 4090 and main memory.
I have measured it and it is definitely there as compared to how long the computations take on the 4090 which definitely take way smaller than the time it takes to read/write data thru the PCIE channel.
Extremely fast and very effiicient and barely taxes the 4090. In fact I don't see temps greater than 50F when running 23 hours daily, 5 days a week.
Latency is only an issue during very heavy market days and cpu useage can spike easily and slows the system for a short time.
Going from PCIE Gen3 to Gen5 will definitely make some latency difference too when running next gen 5090.
But the biggest improvement will be getting an up to date workstation, either a threadripper 7000 HEDT or xeon w-2595x!!

Also, Intel Sapphire chips have much large L2 cache than threadripper and if heavily overclocked, my assumption is that it would perform higher than a similar over clocked threadripper,
which in this case would be threadripper 7960x vs xeon w-2595x only.