Fantastic article with beautiful visualisations, one of the many reasons why I love Anandtech, however I really wish Anandtech would be a little bit more active, oh well, hopefully things will improve in the future in some way or another. Also finally first comment, what a great day, yeah!
Check out the announcement post on techpowerup about the launch of the Sapphire Rapids Workstation Platform and compare it to this one on Anandtech, there really is no comparison which is better and the former is not:). IMHO the slides are very nice and make the article more visually appealing and give a more rounded impression, this is also the reason for my previous comment, after all it still makes a real difference how visually appealing and well structured the same content comes across as both improve memorability and readability.
What's wrong with that? Most of us don't subscribe to the press releases from large tech companies, and rely on sites like AnandTech to give us this sort of news.
I think the point is only that credit is inappropriately being given to Anandtech for these 'beautiful visualizations'. Which were created and published by Intel and reposted here.
Your guess is way off. Threadripper is Zen3 while this one had Golden Cove. Golden Cove has much better IPC than Zen3. Also these are priced lower than Threadripper. I am sure AMD will be forced to lower Threadripper prices due to these just like they were forced to do the same for client Zen3 and Zen4 due to Alder Lake and Raptor Lake.
It most likely can't and the article says it only supports 4ch memory so it's pretty clear this is aimed at the 2400 series though it technically can support the 3400 as well.
Finally Intel launched their SPR XEON based W class HEDT replacements.
AMX - Something that is not a dedicated block like ARM junk has and an x86 HW and not some Apple type garbage. I wonder what's Intel plans are for AGILEX FPGA esp when Xilinx AMD is going to definitely integrate these, I think they incorporated them in those BGA processors for some AI workloads and here AMX is Deep Learning AI workloads without FPGA.
MCC is 24C so the P core scaling is higher, BUT I did not see any mention of RING or MESH. What is the Cache Ring these processors are using ? Mesh was a mess with SKL-X as it was a power hog. I wonder how this fares. Shame this is the only way we can get P cores in Intel CPU lineup.
AVX did not change, so Intel is still inferior to AMD I guess, their AVX512 blocks are order of 2 so they do not need to reduce the Clock ratio. This needs to change from Intel side to be honest.
Next up is the DMI, why does a damn XEON W class chipset is having literally same bandwidth speed as Raptor Lake Z790 chipset, it also has DMI4.0x8, this thing has having 112 PCIE lanes ffs and yet the saturation exists. Shame. AMD's Zen 4 X670E is also on the same side, it has same link speed as X570 series. On the PCIe controllers Intel has 80 in HPC vs 112 in Prosumer lol, that is definitely a sign of how multiple Stepping revisions had to take place to fix the issues in SPR platform, did not think that it could have limitations with high socket scalability.
Now the cost it's just insane. Like TR Pro, these are unobtanium no way anyone can put a Xeon 3400 series, really unfortunate BS that there's no more HEDT, well we have stupid normies who are like MUH GAMING and only focus on that garbage so we get 8P cores max on Mainstream and this sort of garbage pricing for HEDT. AMD is the only one at-least having that high cores but damn their new X3D is a joke for Zen 4 again muh gaming.. Moving on it's interesting how Intel kept same W790 chipset for both Xeon classes of processors unlike AMD's old socket disaster of sTRX platforms. They gave a huge L to the consumers esp TR 3000 series owners.
Interesting to see how far these processors can OC and esp what's the damn performance given that they do not have the garbage E core nonsense, those low core W5 parts which can be OCed are the focus as they scale over 8 cores esp with the new 24C MCC and EMIB involved >24C XCC parts, but will anyone of the YTers will review these SKUs, they are super high cost.
Historically Xilinx and Altera FPGAs were not good at AI, so I dont think Intel is doing that.
Then again, AMD specifically mentioned "reconfigurable" Xilinx IP in their laptop 7000 series AI block, which is really strange. No one knows how it works yet, but my impression is its another device (like a GPU or the Apple AI cores) rather than something built into the CPU cores like AMX or the ARM/RISC-V matrix extensions.
Also... I am very much a fan of the E cores. In fact, I think Intel shot themselves in the foot by not making an E-core only tile. Cloud providers woild *love* that.
"AVX did not change, so Intel is still inferior to AMD I guess, their AVX512 blocks are order of 2 so they do not need to reduce the Clock ratio. This needs to change from Intel side to be honest." -- this is huge misunderstanding IMHO. Intel's AVX512 is basically twice of AMD width so more performant.
RKL gets destroyed in AVX512 vs Ryzen 7000. There's no contest in AVX512 with AMD, Intel gets destroyed. ADL fused it off. Once the benches are out for SPR HPC XEON vs EPYC Genoa we will see how "Twice" Intel performs vs AMD.
@Silver: You seem to have no clue about what you’re talking about. RKL is 2 year older than 7000. Not a fair or intelligent comparison. Phoronix did a comparison of Genoa vs Sapphire Rapids and found that Intel’s implementation is much superior. Here is their summary:
“When taking the geometric mean of all the raw AVX-512 benchmark results in this article, the Xeon Platinum 8380 2P "Ice Lake" performance went up by 34% with the AVX-512 enabled run where as the Xeon Platinum 8490H 2P "Sapphire Rapids" AVX-512 performance improved by 44% and the EPYC 9654 2P "Genoa" performance was up by 21% with AVX-512. Sapphire Rapids not only enjoyed much larger uplift from AVX-512 compared to with Ice Lake, but also a big win for power efficiency.”
The OP said 2x performance. I said RKL is dead vs Ryzen 7000 (7600X 6C12T is almost equivalent to 11900K 8C16T and the TDP better not to talk and the heat as well along with the Clockspeed negative offset)
Since there's no ADL / RPL AVX512 that makes the only valid option to compare with the existing processors. Your "Intelligent" post is about Genoa vs SPR. And that is also saying how "Superior" it is from Intel perspective ? (From your link the top end Xeon SPR lost to Genoa)
What you're missing is that while Zen 4 splits most AVX-512 instructions into 2 parts and executes them sequentially, Zen 4 has more vector issue ports. This gives Zen 4 the same effective AVX-512 throughput as Sapphire Rapids, except for FMA (where Zen 4 has only one port and SPR has 2).
The benchmarks of AVX-512 performance on Genoa vs. Sapphire Rapids back this up. The per-core performance is virtually the same between the two (with SPR having a slight advantage).
Zen 4 executes each 256 bit half of an AVX-512 instruction in parallel, not sequentially (it’s split at the latest stage it can be). That’s what lets them execute 1 AVX-512 instruction per cycle. Only stores are executed sequentially (1 512-bit store per 2 cycles).
I am a bit sad that the 2400 series pricing has shaken out this way even though I expected it to. AMD reset the market pricing with the TR 3000 series and Intel is all too happy to join them. I was hoping the 2455X would fall into the $700-800 range meaning Intel would be doing their own price reset.
There are some motherboards which are designed for the 2400 series so that might mean their pricing will be better than the universal ones. Haven't seen any pricing yet but perhaps that will be enough to make up the difference though I doubt it.
I wish they put the 2400-series in a smaller socket. However, maybe they figured sales volumes would be low enough (i.e. with desktops eating into the low-end workstation market) that it wouldn't be much of a win, pricing-wise.
Yeah one would think they could cut off a significant number of pins given half the PCIe and memory channels. I wouldn't be surprised if it did come down to not making financial sense to make a third socket size.
I assume this is all we're getting, and there won't be i9-149#0X chips with somewhat fewer DDR5 channels appearing based on the MCC silicon for the chunky-desktop market. Pity, my i9-7940X is getting a bit long in the tooth and I'd quite like 24 cores for $1499 for a successor.
Desktop chips have eaten the lower workstation market. The 13900 (and the AMD 7950) will outperform most of the Sappire Rapids low end, unless you need a boatload of IO with a small core config, and those will get refreshed before the next HEDT update.
...because it's meant to be followed so soon by Granite Rapids that I wouldn't be surprised if they just skipped releasing Emerald Rapids, for a lot of market segments. Sort of like what happened with Broadwell.
The exception to that would be if Granite Rapids hits delays. I assume the main reason EMR is being kept alive is to hedge against that.
I can't see any reason for Intel to put out EMR workstation chips without future chip delays. I doubt they'd be enough more competitive against Zen 4 threadripper for that to be a driver and since they're just a refined SPR it makes sense to primarily use them for server/HPC.
You can have some idea what to expect by looking at Alder Lake vs. Ryzen 7000. Yes, these are the same P-cores & process node that were used in Alder Lake (not Raptor Lake), but they're saddled with the additional thermal overhead of a CPU with a lot more PCIe 5.0 lanes and DDR5 memory channels.
Happy to see at least one accelerator and CXL Support, thanks for the info. It makes this launch more exciting even if pricing went up more than i expected.
Oof, and a Xeon Max with 56 cores will set you back a cool $13k. Funny enough, it's not even the most expensive SPR Xeon. That distinction goes to the 60-core 8490H with all accelerators enabled, for $17k. But even the 48-core 8468H is $1k more expensive than the top Max model.
In spite of a 350 W TDP, the entire Xeon Max series has a Turbo limit of just 3.5 GHz, which is somewhat intriguing. Perhaps you wouldn't get a lot more out of it, even as a Xeon W.
I also find it interesting they disabled all the accelerators on Max, other than DSA (and AMX, of course). Market segmentation, I guess. Plus, those pesky accelerators would just eat into the TDP budget of the cores and further lower turbo limits.
I can finally move on from X299! Currently have an Asrock Creator board but looking forward to the new tech. I may even stick with Asrock but we'll see when the boards are released.
Hello , i am planning to build a server or workstation for home server with linux with kvm and 3-4 guest 8 gb ram each with networking /dpdk ,one of guest is firewall /paloaltonetworks panos which system/cpu i should buy
It’s impossible to answer your question without knowing what you intend to use it for. Are you studying networking or distributed computing or doing something else?
To get started, almost any PC CPU made in the last 10 years would be fine, people have run this kind of setup on laptops. Linux guests run fine on 2-4GB of ram each if you’re just studying networking and how to set up networks and firewalls and all that stuff.
If it’s for actual work or for hosting shared games that’s quite different.
Another wrinkle is if you’re studying virtualisation and hosting, then various CPUs have different virtualisation abilities. As a beginner any CPU will do to get started. When you’ve hit the limits of what your current machine can do, you’ll have a better idea of what the next step for you is.
Asus W790 Sage board seams to be very high end with : https://www.asus.com/motherboards-components/mothe... Here are some specs - 7 x PCIe 5.0 slot(s) at x16/x16/x16/x16/x16/x8/x16 (with Xeon 34xx CPUs) - 8 x DIMM slots, Max. 2048GB, overclockable up to DDR5 6800 (with limited RAM usage but still) and support some sort of RDIMM XMP... - 2 x m.2 at PCIe 5.0 x4 - 2 x SlimSAS - nice! - 2 x 10Gbit ethernet (intel native) - The ASUS OC tech (that ASUS also have on other boards - BIOS will have more feature for sure) Extreme OC Kit - FlexKey button - LN2 Mode - ReTry button - Safe boot button - Start button ---- - 4+1+1 power stages - usual latest USB 3.2 2x2, etc I/O - some workstation remote management h/w and s/w BMC (Baseboard Management Controller) - enterprise grade.
a very competent WS motherboard with (at least on paper) Enthusiast level OC capability.
Price? is it going to be that same as the corresponding Threadripper Pro WS WRX80E-SAGE SE WIFI II at $1,099.99?
price for a w3475 36 core based system is going to be far north of $5K (even with max reuse) - Ouch! I think I'll wait and hope for a price war :)
that's where the slimSAS posts comes in handy. 1 SlimSAS expands to 8 SATA with a slimSAS x8 SATA cable - $25 on amazon (may be included - we'll see...)
Ah I was not aware of that thank you. Still will wait for more options, especially the W-2400 only boards. The W-3400 series is too much for my needs and too expensive. I expect the W-2400 will perform great for gaming and video encoding.
That's not actually a guarantee, and most only support 4 when they support it at all. Per the tech specs listed by Asus on that product page: SlimSAS_1 port supports PCIe 4.0 x4 mode NVMe device SlimSAS_2 port supports PCIe 4.0 x4 mode NVMe device
This indicates no direct SATA support for those two ports.
I don't understand the PCIe relationship between the W790 chipset based Mobo and the CPU: CPU - Offers N PCIe 5.0 lanes, depending on type. Mobo - Only supports PCIe 4.0 and 3.0 peripherals via W790 chipset.
The CPU is connected to the chipset by 8x PCIe 4.0 lanes which is why there's the restriction there. The W790 chipset appears to be the same hardware as is in the Z790 so it shares all of the same base specifications while running different firmware.
Historically the connection between CPU and chipset hasn't been particularly high performance as it hasn't been needed and running the lanes there isn't necessarily cheap for motherboards. As we're seeing a rise in M.2 storage and high speed USB/TB this has changed so perhaps in another generation or two this will be run off of PCIe 5.0.
I wonder what happened to Core i9, ie, prosumers (consumer grade workstations), like how the original Threadripper was. Current TR & these Xeons are too expensive and complicated for prosumers. But they're appealing to professionals and corporate-level workstations.
Another thing about these is the long time to market compared to the original architecture (regular Desktop), Especially on the AMD side, they just released the Zen3-based TR just before the Zen4 release, which means Zen4-based TR will come later this year just before the Zen5 release !!
That market died when Intel released the 10 series HEDT and AMD started TR 3000 at $1400. X299 platform notoriously blew up in everyone's face and TR was too expensive for that market so everything afterwards has been aimed at the high performance workstation market.
It's very unfortunate, but seemingly isn't a large market as it was allowed to die. It's what I've always gotten due to the extra PCIe, but the price of entry just seems too high to justify relative performance wise so I'm working out how to cut PCIe usage.
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
Facebook
Twitter
RSS
61 Comments
Back to Article
ILove Anandtech - Wednesday, February 15, 2023 - link
Fantastic article with beautiful visualisations, one of the many reasons why I love Anandtech, however I really wish Anandtech would be a little bit more active, oh well, hopefully things will improve in the future in some way or another. Also finally first comment, what a great day, yeah!nandnandnand - Wednesday, February 15, 2023 - link
Beautiful visualisations? They're just Intel slides.ILove Anandtech - Wednesday, February 15, 2023 - link
Check out the announcement post on techpowerup about the launch of the Sapphire Rapids Workstation Platform and compare it to this one on Anandtech, there really is no comparison which is better and the former is not:). IMHO the slides are very nice and make the article more visually appealing and give a more rounded impression, this is also the reason for my previous comment, after all it still makes a real difference how visually appealing and well structured the same content comes across as both improve memorability and readability.boozed - Wednesday, February 15, 2023 - link
It's still just a reposted press release.name99 - Wednesday, February 15, 2023 - link
What's wrong with that? Most of us don't subscribe to the press releases from large tech companies, and rely on sites like AnandTech to give us this sort of news.DougMcC - Thursday, February 16, 2023 - link
I think the point is only that credit is inappropriately being given to Anandtech for these 'beautiful visualizations'. Which were created and published by Intel and reposted here.DannyH246 - Wednesday, February 15, 2023 - link
Yawn....late, hot, slow, expensive. Typical Intel.Jorgp2 - Wednesday, February 15, 2023 - link
As opposed to what exactly?ILove Anandtech - Wednesday, February 15, 2023 - link
The AMD Threadripper offerings I would gues...m53 - Thursday, February 16, 2023 - link
Your guess is way off. Threadripper is Zen3 while this one had Golden Cove. Golden Cove has much better IPC than Zen3. Also these are priced lower than Threadripper. I am sure AMD will be forced to lower Threadripper prices due to these just like they were forced to do the same for client Zen3 and Zen4 due to Alder Lake and Raptor Lake.jerrylzy - Saturday, February 18, 2023 - link
Golden Cove cores running at much lower clocks don't sound different from TR Zen 3 to me.mode_13h - Sunday, February 19, 2023 - link
You can already find Sapphire Rapids benchmarks. Just not on this site. I guess Anandtech doesn't do server CPU reviews, any more.m53 - Thursday, February 16, 2023 - link
@DannyH246: Now go back to your school, kidmeacupla - Wednesday, February 15, 2023 - link
I see that Asrock board can't even utilize all 112 lanes.Unless it has another 4x m.2 slots on the backside?
Monstieur - Wednesday, February 15, 2023 - link
It might have CPU-attached 10 GbE NICs, USB controllers, and U.2 ports.thestryker - Wednesday, February 15, 2023 - link
It most likely can't and the article says it only supports 4ch memory so it's pretty clear this is aimed at the 2400 series though it technically can support the 3400 as well.Silver5urfer - Wednesday, February 15, 2023 - link
Finally Intel launched their SPR XEON based W class HEDT replacements.AMX - Something that is not a dedicated block like ARM junk has and an x86 HW and not some Apple type garbage. I wonder what's Intel plans are for AGILEX FPGA esp when Xilinx AMD is going to definitely integrate these, I think they incorporated them in those BGA processors for some AI workloads and here AMX is Deep Learning AI workloads without FPGA.
MCC is 24C so the P core scaling is higher, BUT I did not see any mention of RING or MESH. What is the Cache Ring these processors are using ? Mesh was a mess with SKL-X as it was a power hog. I wonder how this fares. Shame this is the only way we can get P cores in Intel CPU lineup.
AVX did not change, so Intel is still inferior to AMD I guess, their AVX512 blocks are order of 2 so they do not need to reduce the Clock ratio. This needs to change from Intel side to be honest.
Next up is the DMI, why does a damn XEON W class chipset is having literally same bandwidth speed as Raptor Lake Z790 chipset, it also has DMI4.0x8, this thing has having 112 PCIE lanes ffs and yet the saturation exists. Shame. AMD's Zen 4 X670E is also on the same side, it has same link speed as X570 series. On the PCIe controllers Intel has 80 in HPC vs 112 in Prosumer lol, that is definitely a sign of how multiple Stepping revisions had to take place to fix the issues in SPR platform, did not think that it could have limitations with high socket scalability.
Now the cost it's just insane. Like TR Pro, these are unobtanium no way anyone can put a Xeon 3400 series, really unfortunate BS that there's no more HEDT, well we have stupid normies who are like MUH GAMING and only focus on that garbage so we get 8P cores max on Mainstream and this sort of garbage pricing for HEDT. AMD is the only one at-least having that high cores but damn their new X3D is a joke for Zen 4 again muh gaming.. Moving on it's interesting how Intel kept same W790 chipset for both Xeon classes of processors unlike AMD's old socket disaster of sTRX platforms. They gave a huge L to the consumers esp TR 3000 series owners.
Interesting to see how far these processors can OC and esp what's the damn performance given that they do not have the garbage E core nonsense, those low core W5 parts which can be OCed are the focus as they scale over 8 cores esp with the new 24C MCC and EMIB involved >24C XCC parts, but will anyone of the YTers will review these SKUs, they are super high cost.
brucethemoose - Wednesday, February 15, 2023 - link
Historically Xilinx and Altera FPGAs were not good at AI, so I dont think Intel is doing that.Then again, AMD specifically mentioned "reconfigurable" Xilinx IP in their laptop 7000 series AI block, which is really strange. No one knows how it works yet, but my impression is its another device (like a GPU or the Apple AI cores) rather than something built into the CPU cores like AMX or the ARM/RISC-V matrix extensions.
Also... I am very much a fan of the E cores. In fact, I think Intel shot themselves in the foot by not making an E-core only tile. Cloud providers woild *love* that.
JayNor - Wednesday, February 15, 2023 - link
The e-core only chip is Sierra Forest, coming on Intel-3. It taped in a year ago, prior to Granite Rapids.brucethemoose - Wednesday, February 15, 2023 - link
A right I thought that was after granite.TBH a workstation variant of that would sell too. I imagine lots of devs would love a thread-happy compilation monster.
The Von Matrices - Wednesday, February 15, 2023 - link
112 PCIe lanes are from the CPU, not the chipset. No bottleneck.Silver5urfer - Friday, February 17, 2023 - link
Thanks I overlooked that part.kgardas - Thursday, February 16, 2023 - link
"AVX did not change, so Intel is still inferior to AMD I guess, their AVX512 blocks are order of 2 so they do not need to reduce the Clock ratio. This needs to change from Intel side to be honest." -- this is huge misunderstanding IMHO. Intel's AVX512 is basically twice of AMD width so more performant.Silver5urfer - Friday, February 17, 2023 - link
RKL gets destroyed in AVX512 vs Ryzen 7000. There's no contest in AVX512 with AMD, Intel gets destroyed. ADL fused it off. Once the benches are out for SPR HPC XEON vs EPYC Genoa we will see how "Twice" Intel performs vs AMD.m53 - Friday, February 17, 2023 - link
@Silver: You seem to have no clue about what you’re talking about. RKL is 2 year older than 7000. Not a fair or intelligent comparison. Phoronix did a comparison of Genoa vs Sapphire Rapids and found that Intel’s implementation is much superior. Here is their summary:“When taking the geometric mean of all the raw AVX-512 benchmark results in this article, the Xeon Platinum 8380 2P "Ice Lake" performance went up by 34% with the AVX-512 enabled run where as the Xeon Platinum 8490H 2P "Sapphire Rapids" AVX-512 performance improved by 44% and the EPYC 9654 2P "Genoa" performance was up by 21% with AVX-512. Sapphire Rapids not only enjoyed much larger uplift from AVX-512 compared to with Ice Lake, but also a big win for power efficiency.”
https://www.phoronix.com/review/intel-sapphirerapi...
Silver5urfer - Saturday, February 18, 2023 - link
The OP said 2x performance. I said RKL is dead vs Ryzen 7000 (7600X 6C12T is almost equivalent to 11900K 8C16T and the TDP better not to talk and the heat as well along with the Clockspeed negative offset)Since there's no ADL / RPL AVX512 that makes the only valid option to compare with the existing processors. Your "Intelligent" post is about Genoa vs SPR. And that is also saying how "Superior" it is from Intel perspective ? (From your link the top end Xeon SPR lost to Genoa)
Give me a break.
mode_13h - Friday, February 17, 2023 - link
What you're missing is that while Zen 4 splits most AVX-512 instructions into 2 parts and executes them sequentially, Zen 4 has more vector issue ports. This gives Zen 4 the same effective AVX-512 throughput as Sapphire Rapids, except for FMA (where Zen 4 has only one port and SPR has 2).The benchmarks of AVX-512 performance on Genoa vs. Sapphire Rapids back this up. The per-core performance is virtually the same between the two (with SPR having a slight advantage).
MetalScythe - Tuesday, March 28, 2023 - link
Zen 4 executes each 256 bit half of an AVX-512 instruction in parallel, not sequentially (it’s split at the latest stage it can be). That’s what lets them execute 1 AVX-512 instruction per cycle. Only stores are executed sequentially (1 512-bit store per 2 cycles).mode_13h - Friday, February 17, 2023 - link
> I did not see any mention of RING or MESH.The slides show block diagrams which mention a mesh. For this many cores & other blocks per tile, ring wouldn't make much sense.
TEAMSWITCHER - Wednesday, February 15, 2023 - link
These prices make my 24 Core AMD Ryzen Threadripper CPU (3960X) seem like a bargain.It was only $1400.
thestryker - Wednesday, February 15, 2023 - link
I am a bit sad that the 2400 series pricing has shaken out this way even though I expected it to. AMD reset the market pricing with the TR 3000 series and Intel is all too happy to join them. I was hoping the 2455X would fall into the $700-800 range meaning Intel would be doing their own price reset.There are some motherboards which are designed for the 2400 series so that might mean their pricing will be better than the universal ones. Haven't seen any pricing yet but perhaps that will be enough to make up the difference though I doubt it.
mode_13h - Friday, February 17, 2023 - link
I wish they put the 2400-series in a smaller socket. However, maybe they figured sales volumes would be low enough (i.e. with desktops eating into the low-end workstation market) that it wouldn't be much of a win, pricing-wise.thestryker - Saturday, February 18, 2023 - link
Yeah one would think they could cut off a significant number of pins given half the PCIe and memory channels. I wouldn't be surprised if it did come down to not making financial sense to make a third socket size.TomWomack - Wednesday, February 15, 2023 - link
I assume this is all we're getting, and there won't be i9-149#0X chips with somewhat fewer DDR5 channels appearing based on the MCC silicon for the chunky-desktop market. Pity, my i9-7940X is getting a bit long in the tooth and I'd quite like 24 cores for $1499 for a successor.brucethemoose - Wednesday, February 15, 2023 - link
Desktop chips have eaten the lower workstation market. The 13900 (and the AMD 7950) will outperform most of the Sappire Rapids low end, unless you need a boatload of IO with a small core config, and those will get refreshed before the next HEDT update.JayNor - Wednesday, February 15, 2023 - link
Emerald Rapids is also coming this year.The desktop chips have no avx512, no AMX, no DSA, few memory channels, no User Interrupts.
mode_13h - Friday, February 17, 2023 - link
Has Intel announced an EMR refresh of their Xeon W line?mode_13h - Friday, February 17, 2023 - link
...because it's meant to be followed so soon by Granite Rapids that I wouldn't be surprised if they just skipped releasing Emerald Rapids, for a lot of market segments. Sort of like what happened with Broadwell.The exception to that would be if Granite Rapids hits delays. I assume the main reason EMR is being kept alive is to hedge against that.
thestryker - Saturday, February 18, 2023 - link
I can't see any reason for Intel to put out EMR workstation chips without future chip delays. I doubt they'd be enough more competitive against Zen 4 threadripper for that to be a driver and since they're just a refined SPR it makes sense to primarily use them for server/HPC.Tunnah - Wednesday, February 15, 2023 - link
If the perf/watt is less than double AMD I'll be surprised.Foeketijn - Thursday, February 16, 2023 - link
You think the P cores are more efficient than ZEN 4?I'm also curious, but am expecting a disappointment.
mode_13h - Friday, February 17, 2023 - link
You can have some idea what to expect by looking at Alder Lake vs. Ryzen 7000. Yes, these are the same P-cores & process node that were used in Alder Lake (not Raptor Lake), but they're saddled with the additional thermal overhead of a CPU with a lot more PCIe 5.0 lanes and DDR5 memory channels.ANORTECH - Wednesday, February 15, 2023 - link
Happy to see at least one accelerator and CXL Support, thanks for the info. It makes this launch more exciting even if pricing went up more than i expected.brucethemoose - Wednesday, February 15, 2023 - link
I guess HBM was too much to hope for.Then again, there aren't even any benches to suggest it will help outside of crazy AMX-dense workloads.
mode_13h - Friday, February 17, 2023 - link
If you want HBM, Intel will happily sell you a Xeon Max that you can put in a server board (once they launch).brucethemoose - Friday, February 17, 2023 - link
Yeah but not a frequency/TDP unlocked one, or one with the discount/extra lanes of the 1P HEDT SKUs.mode_13h - Sunday, February 19, 2023 - link
80 PCIe 5.0 lanes isn't enough for you?:D
Oof, and a Xeon Max with 56 cores will set you back a cool $13k. Funny enough, it's not even the most expensive SPR Xeon. That distinction goes to the 60-core 8490H with all accelerators enabled, for $17k. But even the 48-core 8468H is $1k more expensive than the top Max model.
In spite of a 350 W TDP, the entire Xeon Max series has a Turbo limit of just 3.5 GHz, which is somewhat intriguing. Perhaps you wouldn't get a lot more out of it, even as a Xeon W.
I also find it interesting they disabled all the accelerators on Max, other than DSA (and AMX, of course). Market segmentation, I guess. Plus, those pesky accelerators would just eat into the TDP budget of the cores and further lower turbo limits.
Demon of Elru - Wednesday, February 15, 2023 - link
I can finally move on from X299! Currently have an Asrock Creator board but looking forward to the new tech. I may even stick with Asrock but we'll see when the boards are released.shrikantp - Thursday, February 16, 2023 - link
Hello , i am planning to build a server or workstation for home server with linux with kvm and 3-4 guest 8 gb ram each with networking /dpdk ,one of guest is firewall /paloaltonetworks panos which system/cpu i should buyTomatotech - Saturday, February 18, 2023 - link
It’s impossible to answer your question without knowing what you intend to use it for. Are you studying networking or distributed computing or doing something else?To get started, almost any PC CPU made in the last 10 years would be fine, people have run this kind of setup on laptops. Linux guests run fine on 2-4GB of ram each if you’re just studying networking and how to set up networks and firewalls and all that stuff.
If it’s for actual work or for hosting shared games that’s quite different.
Another wrinkle is if you’re studying virtualisation and hosting, then various CPUs have different virtualisation abilities. As a beginner any CPU will do to get started. When you’ve hit the limits of what your current machine can do, you’ll have a better idea of what the next step for you is.
SystemsBuilder - Thursday, February 16, 2023 - link
Asus W790 Sage board seams to be very high end with :https://www.asus.com/motherboards-components/mothe...
Here are some specs
- 7 x PCIe 5.0 slot(s) at x16/x16/x16/x16/x16/x8/x16 (with Xeon 34xx CPUs)
- 8 x DIMM slots, Max. 2048GB, overclockable up to DDR5 6800 (with limited RAM usage but still) and support some sort of RDIMM XMP...
- 2 x m.2 at PCIe 5.0 x4
- 2 x SlimSAS - nice!
- 2 x 10Gbit ethernet (intel native)
- The ASUS OC tech (that ASUS also have on other boards - BIOS will have more feature for sure)
Extreme OC Kit
- FlexKey button
- LN2 Mode
- ReTry button
- Safe boot button
- Start button
----
- 4+1+1 power stages
- usual latest USB 3.2 2x2, etc I/O
- some workstation remote management h/w and s/w BMC (Baseboard Management Controller) - enterprise grade.
a very competent WS motherboard with (at least on paper) Enthusiast level OC capability.
Price? is it going to be that same as the corresponding Threadripper Pro WS WRX80E-SAGE SE WIFI II at $1,099.99?
price for a w3475 36 core based system is going to be far north of $5K (even with max reuse) - Ouch!
I think I'll wait and hope for a price war :)
SystemsBuilder - Thursday, February 16, 2023 - link
should be 14+1+1 power stages (not 4) - enough for heavy OC?Demon of Elru - Thursday, February 16, 2023 - link
No Thunderbolt 4 and less SATA and M.2 slots than the Asrock. I will wait until more manufacturers release their boards.SystemsBuilder - Friday, February 17, 2023 - link
that's where the slimSAS posts comes in handy. 1 SlimSAS expands to 8 SATA with a slimSAS x8 SATA cable - $25 on amazon (may be included - we'll see...)Demon of Elru - Saturday, February 18, 2023 - link
Ah I was not aware of that thank you. Still will wait for more options, especially the W-2400 only boards. The W-3400 series is too much for my needs and too expensive. I expect the W-2400 will perform great for gaming and video encoding.thestryker - Saturday, February 18, 2023 - link
That's not actually a guarantee, and most only support 4 when they support it at all. Per the tech specs listed by Asus on that product page:SlimSAS_1 port supports PCIe 4.0 x4 mode NVMe device
SlimSAS_2 port supports PCIe 4.0 x4 mode NVMe device
This indicates no direct SATA support for those two ports.
Demon of Elru - Sunday, February 19, 2023 - link
So far the Asrock board looks ideal. But I will wait until more manufacturers release their boards.SharpEars - Monday, February 20, 2023 - link
I don't understand the PCIe relationship between the W790 chipset based Mobo and the CPU:CPU - Offers N PCIe 5.0 lanes, depending on type.
Mobo - Only supports PCIe 4.0 and 3.0 peripherals via W790 chipset.
thestryker - Monday, February 20, 2023 - link
The CPU is connected to the chipset by 8x PCIe 4.0 lanes which is why there's the restriction there. The W790 chipset appears to be the same hardware as is in the Z790 so it shares all of the same base specifications while running different firmware.Historically the connection between CPU and chipset hasn't been particularly high performance as it hasn't been needed and running the lanes there isn't necessarily cheap for motherboards. As we're seeing a rise in M.2 storage and high speed USB/TB this has changed so perhaps in another generation or two this will be run off of PCIe 5.0.
Xajel - Wednesday, February 22, 2023 - link
I wonder what happened to Core i9, ie, prosumers (consumer grade workstations), like how the original Threadripper was. Current TR & these Xeons are too expensive and complicated for prosumers. But they're appealing to professionals and corporate-level workstations.Another thing about these is the long time to market compared to the original architecture (regular Desktop), Especially on the AMD side, they just released the Zen3-based TR just before the Zen4 release, which means Zen4-based TR will come later this year just before the Zen5 release !!
thestryker - Thursday, February 23, 2023 - link
That market died when Intel released the 10 series HEDT and AMD started TR 3000 at $1400. X299 platform notoriously blew up in everyone's face and TR was too expensive for that market so everything afterwards has been aimed at the high performance workstation market.It's very unfortunate, but seemingly isn't a large market as it was allowed to die. It's what I've always gotten due to the extra PCIe, but the price of entry just seems too high to justify relative performance wise so I'm working out how to cut PCIe usage.