Tek-CheckHahaha! 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.[....]

Tek-CheckHahaha! 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.

Tek-CheckBy 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.

unwind-protectFor 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?).

thestryker6Somehow 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.

btarunrthe W5-2545

Tek-CheckSomehow 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.

ChaitanyaAlso 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.

thesmokingmanDon't forget the optional under table chiller!!

Tek-CheckNew Noctua for TR will do the job just fine. Check it out.

thestryker6I'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.

thestryker6The 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.

Tek-Check2DPC 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.

Tek-CheckI 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.

thestryker6This 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.

Tek-CheckHahaha! 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.

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

Tek-CheckThere is no evidence that compute payloads need x16 Gen5 bandwidth to run fully on desktop and HEDT systems, even without NVLink.
www.cgdirector.com/guide-to-pcie-lanes/
www.cgdirector.com/best-hardware-for-gpu-rendering-in-octane-redshift-vray/

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.

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.

Tek-Check2DPC 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.

lemans24I 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...

lemans24I 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...

lemans24I run risk management every time my trade changes value and it takes around 100ms round trip to send/compute/receive using a 4090 card.

Tek-CheckYou 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?