In 10 years, these will be sold with knock-off motherboards and 128gb of RAM on AliExpress for less than 200 dollars. Imagine having a budget PC with that kind of power
Yikes only 128GB? Not even full memory channels! Usually folks start with 768GB per socket in these systems like we are using, and then go up from there
As a firmware engineer who spent the last 2.5 years working on the OpenBMC features for GNR/BHS, it's great to see this product ship and get positive reviews!
Intel is bleeding money, stock crashed 50%, and they've suspended the dividend. They desperately need to win customers back. If you're an investor, you might want to check their stock. It hasn't been this low since 2013. I can't imagine Intel dying; it might be an evil company but they generally make good products (last gen failures being the rare exception)
@@ServeTheHomeVideo Just to be clear: Real review: you buy the product with your own money and make an independent review. You do whatever you want with the product after you are done (keep, sell etc). If the above doesn't apply to your video, it's an Ad. Sorry to disappoint you 😊
@@MrIcanAmer an ad for who though? anyone buying these CPUs at scale doesn't care about this video and anyone else who is seriously thinking about buying a system with this class of hardware will be doing so much more than watching one video. I also hate the idea that talking about something you didn't buy automatically becomes an ad, he didn't try to sell anyone anything, there are no links to go buy one, and he acknowledges the shortcomings of his tests and the fact that a new generation for AMD is releasing soon. It is ridiculous to think that if someone didn't buy it, then it is automatically a fake review. also he said that it is a developer kit. this means you will never be able to get that server, in fact, you cannot buy an Intel 6 pcore server anywhere right now. this is not an ad pretending to be a review, hell it isn't even a review. he got a couple days with the server to run cursory tests the he had to send it back. this is showcase of awesome, new technology that we would never get to see if it weren't for people like sth!
@@reszie1014 Agree. I could tell by the tone and tenor he was excited that Intel is leaning forward again with eye watering performance while pretending to care about efficiency. Leaps in technology drives Patrick, not shilling for corporate America.
You are right, Intel strikes back not just in the consumer market with Lunar Lake, but also with Xeon in the server/datacenter enterprise market. Now what matters is price vs performance and TCO in comparison to AMD and others. AMD will surely strike back as well and responds to it... competition is good - customer wins!😉
1:36 -- "Challenges." I somewhat am piqued to watch STH's follow-up content and charts on this 6900P, too. Superbly crafted video, thank you 👍 Kindest regards, friends and neighbours.
If they didn't pull the trigger now, they wouldn't even have a couple weeks. The real question is, will these chips still be reliable in 6 months???????????
Looking at the phoronix numbers, it's not even really a win for intel. In performance sure, they're ahead some low double digit percentage so are competing in that sense, but the power budget is blown out of the water relative to AMD's 9754 to do it.
INTEL IS BACK FOLKS!!! As great as these are, I really want to see what they deliver on Clearwater with 18A next year. That's their "leadership" node and I'm hoping they can get the 192 cores on the 350W socket. That would be huge for them. Also excited to see Panther Lake and how that does on the client side. What's crazy is it feels like it wasn't even 18 months ago Intel had launched Sapphire Rapids. They've got 3 whole generations out in not even 2 years.
Don't forget that there are additional 256MB of L2 cache and that the l3 cache is shared between all cores (or at least between all cores on a tile) with Intel, compared to the 128MB L2 cache and the 32MB L3 cache per CCX on AMD. Intel is much better suited for shared memory workloads as the compute and memory is not as segmented.
Pre and eaely ryzen, I actively aimed to do business with anyone but Intel. But I feel like something has changed deep within Intel. Arc is chosing us consumers over the AI craze, their 18a and lunar lake have me excited again, they seem to hear us FOS folk genuinely. I'm actually excited for intel and wish the best
That's competition working as it should. They are not the only option in the market, so they need to care about other things. Now we should be excited about AMD's response!
If they were actually pro-consumer, we would see them actually make commitments to long-lasting sockets like AMD does with AM4 and AM5. Cheap prices don't mean they're all pro-consumer, it just means it's cheaper, but they could stifle upgrades by forcing people to buy a new motherboard every 2 years if they want the latest.
Sadly, and I mean that, sad that intel failed, their latest release is quite underwhelming. What I feel from Intel is they're like Boeing. Their best days are behind them.
6980P beats 9754 in performance, but not efficiency. Which sucks when you think about it, 9754 is on an older TSMC node so intel is still more than a generation behind (when you consider zen5 is round the corner).
True, bust that is not the case in every workload. Remember the 9754 is more of the Sierra Forest segment. This is a Genoa competitor for higher performance per core. Bergamo is just getting many vCPUs in a socket. It is a lot more nuanced than you might think. Microbenchmarks say Bergamo is faster than Genoa. In practice, the high value HGX H100 AI sockets went mostly to lower core count Intel and then Genoa/Genoa-X and not to Bergamo.
@@ServeTheHomeVideo And it costs %30 and the DOD backs it to keep Intel alive after Israeli R&D destroyed Icelake 10nm with their little "backdoor". Consult Chris Domas and "God Mode Unlocked".
Next intel will be on TSMC N3P and will have better efficiency than AMD. AMD time is gone. 2025+ will be intel years. AMD is again where it's normal - low end junk
I appreciate the Xeon line. I'll wait to be able to buy it 4 years later for pennies on the dollar. I'll pair it with a an Aliexpress mobo and make another home server for cheap.
What a beast of a CPU, even if it feels like Intel is taping their cores together with hopes & dreams. I'd really like to know who is using these specific sizes of CPUs and for what purpose. Are there really that many monolithic apps out there that need this many cores & terabytes of memory? Or is it primarily going to be used for huge numbers of virtualized instances? Just crazy.
Geo-data processing, a lot of data + a lot of parallel compute as well as a lot of data needs to be in RAM whole time. Not best example but still possible.
Hi Patrick, I know it's not my business but you should look after your health mate. We want to see you talking about many more future HW releases. take care
No mention that efficiency is still not comparable to AMD? Sure you can blast the cpu with 500W when AMD does 360W for the same core count, and win in charts, but is adding 40% more power to win by less than 30% on the best case worth it? I bet this will force AMD to go to 500W as well, and nobody wins.
I'm on am4. But if intel stops putting the screws to customers and competing I'm interested. Am4 is still a fantastic investment. Intel has yet to prove that they won't force immediate motherboard upgrades.
According to Phoronix Xeon 6 is ahead of current Epyc by around 11% on average (in some cases it is 50% faster, in other cases, it loses by 40%). But in a couple of weeks or less, the Epyc is gonna launch and it will perform way better than desktop Zen 5 as it is made on 3nm node vs 4nm. So I would say with 90% confidence, that AMD will win back the crown and lead by 10-15% on average.
As always with AMD, it will come down to code optimizations. Intel has always been ahead in schedulers and engineering support. That said, I generally track with AMD's engineering philosophy and AMD has decided to compete with Intel. From a market perspective, irreparable damage has been done to Intel's brand with the prior fabrication issues and they've lost a chunk of market share that represents several years of commitment to AMD by firms. I won't say benchmarks don't matter... But it is interesting to see people completely misjudge the market based on benchmarks. In most cases, the marginal difference in performance between flagship components is negligible next to the needs of the business committing to the upgrade. If you are in an environment where that 10% performance margin is your bread and butter, sure, it's worth the cost - but most of the market just needs a product that reliably meets their needs, even in the category of this type of processor. Worse, if I am looking to purchase a massive server farm, for example, I can read news and profit statements. The reason to buy Intel is not the 10% margin. It's to have someone to call to come make your software problem go away when you are trying to get your hundred million dollar water heater to work right. Intel's software engineering support is where they make their profit margins.... Which have gone from nearly 30% 10 years ago to 2% today and they are desperately slashing jobs and expenses to try and get things under control. If I make the commitment to buy an Intel based server environment and live with it for the next 5-8 years at the center of my business... Is Intel going to even be there in 5 years to service their hardware? Sure, maybe I have to hire some technicians and eat some development time up front with an AMD server system - but the tool chain is mostly open and I know what I am in for. I can't control whether or not Intel is going to be able to provide the support expected of their brand for the duration of my business commitment, but I can control what level of investment I make toward my own software and engineering support. Who knows - maybe I can have them consult or do business on the side if they are a bit too idle - or I could have them invest in good will initiatives by contributing to open source projects everyone loves (... Or the business uses). Difference between retainer pay and labor pay.
Intel is back in the competition! I look forward to see how Epyc Turin (not counting the Zen 5c dense version) does as it has the same max core count as GNR but with the newer core architecture.
I left intel as i felt abandoned by them always changing there sockets which led to no upgrade path, i also hate the big little core idea leave that for laptops if going to use it, i want BIG cores that can brutally punish maths, but upgrade path is very important intel has been neglecting this for years.
So these are big P-cores for higher-end compute sockets, and different CPUs with all E-cores for cloud native workloads. I think big.little was for Arm and a mix on the same CPU. A bit different in server land.
Big little is still a smart idea for desktop class chips. AMD is doing the same idea but different execution with their Zen 4C cores. It has already launched with Ryzen 8000 APUs and will certainly continue into the Zen5 generation
@@Dweller12Videos you said nothing about how.....but you most likely don't know you just go along with the slop they give out. Real world application it makes more issues and needs extra steps to get something done."simple" Thus many people turn off the little cores on Computers. Kinda like Windows UI more clicks is somehow better. Windows 11 is the worst for this. Real reasons, its cheaper for them and they get to say things like we have 12 core but really its not as the little cores don't do much work if any at all on a COMPUTER. Devices like mobile phones or routers / switches ect sure why not you don't need the extra instruction sets that General CPU's have and many functions may never get used so it makes sense to have ARM/RISC or even ASICS to do the computer for the tasks that need to get done. Example Mobile phone processors do not need SROIV function at all thus i be shocked to see a mobile phone processor that has that function, thus they can save on the silicon real estate or add functions they need with that saved real estate. I was hoping AMD would do its own thing regarding this, but yeah we might have to deal with the big / little core rubbish, for me when i am forced to buy such a cpu i would most likely turn it off.
@@ServeTheHomeVideo yes server's and there OS has more support for arm but still small compared to standard x86 thus more times then not arm just sit there idle.
Microsoft should change it's licensing formula for Windows Server & SQL Server. It was not designed for this much of cores. I can't imagine how much would cost just a server.
Epyc Naples (not Milan written initially) (Zen 1/1+) had the memory controller also on the CPU chiplet before AMD moved it to the I/O die from Zen 2 till now. There is a trade off by having the memory controller on the CPU die as it results in NUMA which requires software/OS to account for it, while memory controller on the I/O die gives you UMA but with a slightly higher latency.
Genoa is actually NUMA, each group of three CCDs have three local memory channels. You can switch between 1NPS and 4NPS in the UEFI, same as the toggle between 3NPS and 1NPS for GNR. The difference is really just which mode they default to on either platform.
the legacy BIOS compatibility support module (CSM) was dropped on the Whitley (Icelake) platform, but GNR still comes out of reset in real mode at 0xffff:0000. Eventually it'll all be 64-bit (including PEI which has to run out of cache), but not yet. I wouldn't be surprised if some BIOS vendors (really UEFI vendors these days) still support CSM, I don't think the hardware prevents it.
These are cool and all, but I'm actually more interested to see what the small socket P core options will be. Dual 350W chips are already a lot of heat to deal with in a chassis.
@@ServeTheHomeVideo IMO it'd be too much for HEDT. To me HEDT is 12 to 32 cores, quad channel memory and 48 PCIE lanes. Although in reality, I'd prefer consumer socket was more reasonable with 4 to 8 cores and then HEDT 8 to 16. Let workstation be 16 to 64. Basically - I just want the golden days back like X99. The days where if you got more cores than you actually needed for a desktop, you got the HEDT features to go with it. TR4 was such a great start for AMD and they had to go and muck it up. I would have absolutely bought Zen 2 on TR4 - but they went and put the core count I wanted (16c) on the socket that doesn't have the features I also wanted - basically doubling the price of everything, consumer socket and everything above it.
@@ILoveTinfoilHats AMD Threadripper is not HEDT anymore. It is workstation. HEDT is expensive but within reasonable reach for consumers. There are no Threadrippers that fit that bill anymore. TR4 was an incredibly promising start and I would have probably bought TWO of them by now if they had not killed HEDT.
Intel must be really struggling with yields on intel 3. The single tile designs get up to 48 cores, for intel to only enable 42 or 43 on the same die in their flagship part there must not be all that many fully working 48 core tiles. The existence of the 6979P is also informative, there's no reason they'd launch that SKU unless they are getting a lot of tiles with 40-42 working cores. All good for the consumer of course, if yields are amazing on a new process then the designers weren't ambitious enough!
In my experience China and the United States tend to have the cheapest prices for computer parts due to lots of competition and a desirable large market (also low taxes and tariffs until very recently). China has the advantage of Shenzhen basically being the scrappers paradise due to low labor cost and the used/refurbed stuff plus being right where the stuff is produced helps. Generally speaking Germany and most of Europe for that matter is more complicated tax wise and logistically speaking (due to multiple languages and distance) and much less competition (although in Germany Mind Factory prices are excellent) Mind Factory is basically the Micro Center of Germany and many Europeans buy and import from Mind Factory. The funny thing about this whole situation is that many South and Central Americans specifically go to Mexico to buy pc parts to import since it's cheaper that way even when accounting for the plane ticket. Due to the weaker currencies and lower demand plus higher tariffs or taxes it means most people not in Europe/North America basically rely on things like buying from Alibaba such as Zeons (basically intel server cpu) with outdated gpu from 3-8 years ago. Hell anyone around the world can benefit through Alibaba's massive discount on refurbs by buying some am4 cpu or 7500f (basically a down clocked ryzen 7600 without the igpu). Most pc gamers in poor countries tend to have to rely on playing older or easier to run titles like Deep Rock Galactic or free esports games like Leauge of Legends, Valorant, etc or take the risk of flying the seven seas and becoming a pirate.
No more than AMD epyc is 8 processors imo. Maybe less since latency here should be much lower. I’d say yes it still does, number of dies doesn’t seem different than multicore in that regard to me personally.
Well you buy in the unit of a CPU that goes into a socket, and there is a difference between on and off socket latency so I think it still makes sense. You are right though that this split into three is more than three Cascade Lake era CPU sockets
@@jolness1 its different, with AMD all CCDs have a direct connection to memory, here memory is divided into 4 channels per die like Zen 1. so its more like a 3 socket system in a single socket.
So AMD requires a hop every time off chip through package and to the IO die before going off package through pins to memory. That has been the case since Rome.
500W per is not a problem. We showed a NVIDIA GH200 server with a 900W chip in a 1U. www.servethehome.com/supermicro-ars-111gl-nhr-nvidia-gh200-grace-hopper-1u-server-review-arm/
@Masterchief-q8g All of the DC server rack's I've seen were air-cooled, often even without inrow rcs. And I've seen maybe a hundred racks. Emerald rapids were like 200 to 350 watts per socket. I can't even imagine the flow needed to cool 1000 watts from the cpus alone, and the power draw of such fans. Server cpus dont have to be this hot. I've dual socket 6148P scalable running less than a meter away from me when Im sleeping. Under full load the fans alone draw more power than the cpus, storage and networking all summed up.
@@ServeTheHomeVideo 900W chip is a one thing, but two 500w PL1, or 625W pl2 if intel didn't change anything, is a 1250W for dual cpu system like dl360 or dl380
But, these use the old (Meteor Lake class) Redwood Cove P-Cores and note the new Lion-Cove P-Cores that will be in the Arrow Lake CPUs that will be launched in a month or two. To get Lion Cove P-Cores, we have to wait at least another year or two (late 2025, early 2026, in the best case) , for the DIamond Rapids Xeons that _may_ finally use the Lion Cove architecture for its P-Core cpus (and support sixteen memory channels, across two sockets, hopefully).
I lost my trust in Intel they didn't deal properly witg their 13th and 14th gen issue and now what guarantees their new processors wont suffer from similar issues!
Hi, did not mention the max RAM design. Great overview. Intel is back, for me, they never fully left because in reality Intel always performed far better than the paper suggested compared to AMD. With this, it seems they are going to win on paper and in reality.
But how it works when cpu core from one tile needs data from other tile? In AMD EPYC case all memory is flat, and there is no different what core needs what data. In case of Intel it's like 3 sockets in one. It's ok when you running a lot of VMs, that fit in it's own tiles, but in case of some one big task that uses all cores in all tiles as I understand there will be b performance penalty
We showed the flat mode (called Hex mode) on this in the main site as well. Intel memory is die to die going over very fast EMIB. EPYC has to go chip to package to I/O die. EPYC has a similar feature that you will see "NPS=4". The I/O die on EPYC actually performs better in many cases with data locality as well.
@@ServeTheHomeVideo thnx, will read your article. When Anantech shut down, there is left not much places to read about architecture of modern silicon's design
Interesting, I wonder what kind of workload will be delegated to the E-cores most of the time. Edit: nvm it's not hybrid big.LITTLE systems like the consumer line since Alder Lake. So it's either all P-cores or all E-cores in each node
looks nice, but the more important question I have to ask "does intel still hate customers and refuse to replace all those defective processors?" If the answer is still "fuck our customers, buy a new 15th gen to replace the defective 13th gen we sold you."
@@ServeTheHomeVideo I understand that it is not, I see it as complementary technologies, RAM is still tied to only one type (DDRx), "OMI" makes it possible to have at the same time, DDR, GDDR 4/5/6, or another optimized one, even something exotic like HBM sticks (an OMI chip with HBM layers) or any other novelty that the market wants, as long as it is OMI the technology behind it is irrelevant, but connected directly to the CPU. The GB/s memory bandwidth scale would finally be broken, TB/s would be the new normal. I suggest the text "cxl-and-omi-competing-or-complementary" by Bryon Moyer (without a link so as not to cause problems)
@@ServeTheHomeVideo Not to mention the GB/s bandwidth per mm² that allows for many more channels and bandwidth per core. OMI achieves 29GB/s/mm/2 versus approximately 3.3GB/s/mm² for DDR4/5
For workloads that can benefit form high core count + the accelerators, these CPUs are real killers. Next, Intel needs to put similar effort into its software stack and unifies everything just like CUDA. I run LLMs on Nvidia and Intel Arc and I can say that the experience is day and night different (in favour of Nvidia).
I do have thoughts, but I don't have money for those kinds of systems. And I doubt many do. Not sure what this has to do with home though, as these things are clearly designed for data centers.
The STH main site has been the biggest server, storage, and networking review site for almost a decade. We were doing 8x GPU server reviews in 2016-2017.
In 10 years, these will be sold with knock-off motherboards and 128gb of RAM on AliExpress for less than 200 dollars. Imagine having a budget PC with that kind of power
Yikes only 128GB? Not even full memory channels! Usually folks start with 768GB per socket in these systems like we are using, and then go up from there
@@ServeTheHomeVideo When you go to the WC do you shit money ?
My consumer level desktop already has 128gb of RAM...
why wait you can get one for just $24,999 right now! lol
@@ServeTheHomeVideoyeah, pfffffffft! 1TB is where you start if you are serious about it
That is a Mammoth computer! The 90s are back! I think that chip set is bigger then my phone!😅
Huge
As a firmware engineer who spent the last 2.5 years working on the OpenBMC features for GNR/BHS, it's great to see this product ship and get positive reviews!
Thanks for your work! We used OpenBMC here
@@AlexSchendel same here ;) OpenBMC features as an software engineer
Thqs I worked on GPU unit along with the GPU Design team @ intel
It's barely 2 weeks and AMD has responded 😅
Mr. Patrick do you still run the Stockfish Chess Benchmarks?
The 8800MT/s memory support is a game-changer. That is bonkers 1.7 TB/s on a dual-socket server, ideal for CFD and other HPC workloads ;)
Yea!
Love your work Dr M
@@ServeTheHomeVideo Do these memory speeds also count as 'overclocking' / no warranty? :D
Am I the only one who pronounces them as "Mister. DIMM"???
Intel's back, baby! None of the CPU manufacturers are our friends, so some good ol-fashioned competition will do well for us all.
That’s why TSMC keep nerfing everything they build or Intel.
this is Patrick from STH and THIS is expensive!
I want one.
I'd run pihole on it.
For an entire enterprise network? At that level of usage you'd be better off using multiple piholes for load balancing and redundancy
That is a LOT of cores for just pihole, I hope you're planning to run it on Proxmox :)
CPU usage: .005%
Simplified product stack, all features enabled? Who is this company and what have they done with Intel
Intel is bleeding money, stock crashed 50%, and they've suspended the dividend. They desperately need to win customers back.
If you're an investor, you might want to check their stock. It hasn't been this low since 2013. I can't imagine Intel dying; it might be an evil company but they generally make good products (last gen failures being the rare exception)
@@shawn576 The last 10 generations of their CPUs have severe issues and their last 2 gens (minimum) are failing prematurely.
This didn’t age well 😂😂😂
FINALLY Intel learned the lesson on gating features behind SKUs!
This video is a AD. I would like to see real review.
It is not to be clear.
@@ServeTheHomeVideo Just to be clear: Real review: you buy the product with your own money and make an independent review. You do whatever you want with the product after you are done (keep, sell etc). If the above doesn't apply to your video, it's an Ad. Sorry to disappoint you 😊
@@MrIcanAmer an ad for who though? anyone buying these CPUs at scale doesn't care about this video and anyone else who is seriously thinking about buying a system with this class of hardware will be doing so much more than watching one video. I also hate the idea that talking about something you didn't buy automatically becomes an ad, he didn't try to sell anyone anything, there are no links to go buy one, and he acknowledges the shortcomings of his tests and the fact that a new generation for AMD is releasing soon. It is ridiculous to think that if someone didn't buy it, then it is automatically a fake review. also he said that it is a developer kit. this means you will never be able to get that server, in fact, you cannot buy an Intel 6 pcore server anywhere right now. this is not an ad pretending to be a review, hell it isn't even a review. he got a couple days with the server to run cursory tests the he had to send it back. this is showcase of awesome, new technology that we would never get to see if it weren't for people like sth!
@@reszie1014 Agree. I could tell by the tone and tenor he was excited that Intel is leaning forward again with eye watering performance while pretending to care about efficiency. Leaps in technology drives Patrick, not shilling for corporate America.
6900P
Nice
I wonder IF anybody was thinking of this model: 6969P
Ha....
Or 6969PP 😅
@@SpoonHurlerif they release a 6969PP I'm buying one, I don't care if I have to get a second mortgage
Is it faster than a Celeron G6900?
@@SpoonHurler Or 6942P
This video aged like milk left out of the fridge...
Finally the division by three mystery has been solved! thanks Patrick !
You should have heard the gasps when I asked this live
@@ServeTheHomeVideo I can only imagine LOL
You are right, Intel strikes back not just in the consumer market with Lunar Lake, but also with Xeon in the server/datacenter enterprise market. Now what matters is price vs performance and TCO in comparison to AMD and others. AMD will surely strike back as well and responds to it... competition is good - customer wins!😉
Nice work Patrick & Crew !! Watched every second !!
Awesome, thank you!
1:36 -- "Challenges." I somewhat am piqued to watch STH's follow-up content and charts on this 6900P, too.
Superbly crafted video, thank you 👍
Kindest regards, friends and neighbours.
Thanks
It would be short lived since 5th generation Epyc would release in October 2024
Aww, let Intel have one PR win, they need it...
128P core Xeon should compete well against 128 core Zen5 Epyc. As for the 192 core Zen5c Epyc we will see how well the 288E core Xeon performs.
@@lharsay lol
Great win for Intel! ... For not even three weeks. Turin is out on 10.10 with 192 cores per socket. But yes, good to see that competition is back :)
We are under an AMD embargo so cannot talk in detail about that
@@ServeTheHomeVideowe know
If they didn't pull the trigger now, they wouldn't even have a couple weeks. The real question is, will these chips still be reliable in 6 months???????????
Looking at the phoronix numbers, it's not even really a win for intel. In performance sure, they're ahead some low double digit percentage so are competing in that sense, but the power budget is blown out of the water relative to AMD's 9754 to do it.
@@brianwelch1579relative to what standard? The competition fail rate is double the average...
INTEL IS BACK FOLKS!!!
As great as these are, I really want to see what they deliver on Clearwater with 18A next year. That's their "leadership" node and I'm hoping they can get the 192 cores on the 350W socket. That would be huge for them. Also excited to see Panther Lake and how that does on the client side.
What's crazy is it feels like it wasn't even 18 months ago Intel had launched Sapphire Rapids. They've got 3 whole generations out in not even 2 years.
Can't wait to pick one up on ebay for 30 bucks in 10 years.
Don't forget that there are additional 256MB of L2 cache and that the l3 cache is shared between all cores (or at least between all cores on a tile) with Intel, compared to the 128MB L2 cache and the 32MB L3 cache per CCX on AMD. Intel is much better suited for shared memory workloads as the compute and memory is not as segmented.
Great points
Pre and eaely ryzen, I actively aimed to do business with anyone but Intel. But I feel like something has changed deep within Intel. Arc is chosing us consumers over the AI craze, their 18a and lunar lake have me excited again, they seem to hear us FOS folk genuinely. I'm actually excited for intel and wish the best
That's competition working as it should. They are not the only option in the market, so they need to care about other things.
Now we should be excited about AMD's response!
If they were actually pro-consumer, we would see them actually make commitments to long-lasting sockets like AMD does with AM4 and AM5.
Cheap prices don't mean they're all pro-consumer, it just means it's cheaper, but they could stifle upgrades by forcing people to buy a new motherboard every 2 years if they want the latest.
Sadly, and I mean that, sad that intel failed, their latest release is quite underwhelming. What I feel from Intel is they're like Boeing. Their best days are behind them.
6980P beats 9754 in performance, but not efficiency. Which sucks when you think about it, 9754 is on an older TSMC node so intel is still more than a generation behind (when you consider zen5 is round the corner).
True, bust that is not the case in every workload. Remember the 9754 is more of the Sierra Forest segment. This is a Genoa competitor for higher performance per core. Bergamo is just getting many vCPUs in a socket. It is a lot more nuanced than you might think. Microbenchmarks say Bergamo is faster than Genoa. In practice, the high value HGX H100 AI sockets went mostly to lower core count Intel and then Genoa/Genoa-X and not to Bergamo.
@@ServeTheHomeVideo And it costs %30 and the DOD backs it to keep Intel alive after Israeli R&D destroyed Icelake 10nm with their little "backdoor". Consult Chris Domas and "God Mode Unlocked".
Next intel will be on TSMC N3P and will have better efficiency than AMD. AMD time is gone. 2025+ will be intel years. AMD is again where it's normal - low end junk
@@hristobotev9726 The next Intel Xeon CPU's (Clearwater Forest in 2025 and Diamond Rapids in 2026) are on the Intel 18A node instead of TSMC nodes.
Clearwater Forest
Great breakdown. This proc looks good. Thanks
finally caught up with AMD EPYC™ 9754, good for intel.
I appreciate the Xeon line. I'll wait to be able to buy it 4 years later for pennies on the dollar. I'll pair it with a an Aliexpress mobo and make another home server for cheap.
if you can't buy it it's not out yet. when is it actually SHIPPING? or buyable??
With server CPUs it depends on your customer class. If you are a large customer you have been able to get them for some time
Intel tops AMD *shows off Intel slides as proof* :)
Unlike AMD slides, intel slides are usually telling the truth lol
Intel needed some good news lately
What a beast of a CPU, even if it feels like Intel is taping their cores together with hopes & dreams.
I'd really like to know who is using these specific sizes of CPUs and for what purpose. Are there really that many monolithic apps out there that need this many cores & terabytes of memory? Or is it primarily going to be used for huge numbers of virtualized instances? Just crazy.
I went into it a bit. Even with lower core counts than AMD, Intel had been winning high end AI system sockets because of memory and PCIe
Geo-data processing, a lot of data + a lot of parallel compute as well as a lot of data needs to be in RAM whole time. Not best example but still possible.
I want to see Linus tech tips play Minecraft on this
Hi Patrick, I know it's not my business but you should look after your health mate. We want to see you talking about many more future HW releases. take care
Yes
No mention that efficiency is still not comparable to AMD? Sure you can blast the cpu with 500W when AMD does 360W for the same core count, and win in charts, but is adding 40% more power to win by less than 30% on the best case worth it?
I bet this will force AMD to go to 500W as well, and nobody wins.
He said in the video that AMD is also going towards 500W/socket.
AMD is doing the same thing
We have already shown AMD 500W designs for Turin on the main site.
AMD is also going 500W with Epyc Turin lol
@@TitanPlakInside he just wants a reason to shit on intel :D
so you're telling me that Intel is gluing their chips together now?
I was sitting in the room when that was said.
I'm on am4. But if intel stops putting the screws to customers and competing I'm interested. Am4 is still a fantastic investment. Intel has yet to prove that they won't force immediate motherboard upgrades.
Do the MRDIMM' use the same ram slot as DDR5? could you upgrade a server from DDR5 to MRDIMM in the future?
Yes
Good video Patrick
Thank you Mr. Arm
Finally - it was nice to see AMD rise up, but they've started to get a bit complacent. Nice to see Intel woke up 👍
According to Phoronix Xeon 6 is ahead of current Epyc by around 11% on average (in some cases it is 50% faster, in other cases, it loses by 40%). But in a couple of weeks or less, the Epyc is gonna launch and it will perform way better than desktop Zen 5 as it is made on 3nm node vs 4nm. So I would say with 90% confidence, that AMD will win back the crown and lead by 10-15% on average.
We are under the AMD Turin embargo so we cannot go into details right now
As always with AMD, it will come down to code optimizations. Intel has always been ahead in schedulers and engineering support.
That said, I generally track with AMD's engineering philosophy and AMD has decided to compete with Intel.
From a market perspective, irreparable damage has been done to Intel's brand with the prior fabrication issues and they've lost a chunk of market share that represents several years of commitment to AMD by firms.
I won't say benchmarks don't matter... But it is interesting to see people completely misjudge the market based on benchmarks. In most cases, the marginal difference in performance between flagship components is negligible next to the needs of the business committing to the upgrade. If you are in an environment where that 10% performance margin is your bread and butter, sure, it's worth the cost - but most of the market just needs a product that reliably meets their needs, even in the category of this type of processor.
Worse, if I am looking to purchase a massive server farm, for example, I can read news and profit statements. The reason to buy Intel is not the 10% margin. It's to have someone to call to come make your software problem go away when you are trying to get your hundred million dollar water heater to work right. Intel's software engineering support is where they make their profit margins.... Which have gone from nearly 30% 10 years ago to 2% today and they are desperately slashing jobs and expenses to try and get things under control.
If I make the commitment to buy an Intel based server environment and live with it for the next 5-8 years at the center of my business... Is Intel going to even be there in 5 years to service their hardware?
Sure, maybe I have to hire some technicians and eat some development time up front with an AMD server system - but the tool chain is mostly open and I know what I am in for. I can't control whether or not Intel is going to be able to provide the support expected of their brand for the duration of my business commitment, but I can control what level of investment I make toward my own software and engineering support. Who knows - maybe I can have them consult or do business on the side if they are a bit too idle - or I could have them invest in good will initiatives by contributing to open source projects everyone loves (... Or the business uses). Difference between retainer pay and labor pay.
@@ServeTheHomeVideo Totally understand
Intel is back in the competition!
I look forward to see how Epyc Turin (not counting the Zen 5c dense version) does as it has the same max core count as GNR but with the newer core architecture.
New stuff at STTH, with 256 pounds of threads and fat 504MB of cache they must be feeling great.
The one that generates a lot of curiosity to me is the Intel that has 288 e-cores.
Yes. I am very excited for Sierra Forest-AP
At some point there'll be a CPU so powerful that if you buy one, agents break your door down and drag you away for questioning
I think they should call it "6900 PP", but I guess that would be too much on the nose of a CMO :D
I watched this whole thing and all I wanted to see was a cinebench run and what task manager looks like with all those threads.. lol
I left intel as i felt abandoned by them always changing there sockets which led to no upgrade path, i also hate the big little core idea leave that for laptops if going to use it, i want BIG cores that can brutally punish maths, but upgrade path is very important intel has been neglecting this for years.
So these are big P-cores for higher-end compute sockets, and different CPUs with all E-cores for cloud native workloads. I think big.little was for Arm and a mix on the same CPU. A bit different in server land.
Big little is still a smart idea for desktop class chips. AMD is doing the same idea but different execution with their Zen 4C cores. It has already launched with Ryzen 8000 APUs and will certainly continue into the Zen5 generation
@@Dweller12Videos you said nothing about how.....but you most likely don't know you just go along with the slop they give out.
Real world application it makes more issues and needs extra steps to get something done."simple"
Thus many people turn off the little cores on Computers.
Kinda like Windows UI more clicks is somehow better.
Windows 11 is the worst for this.
Real reasons, its cheaper for them and they get to say things like we have 12 core but really its not as the little cores don't do much work if any at all on a COMPUTER.
Devices like mobile phones or routers / switches ect sure why not you don't need the extra instruction sets that General CPU's have and many functions may never get used so it makes sense to have ARM/RISC or even ASICS to do the computer for the tasks that need to get done.
Example Mobile phone processors do not need SROIV function at all thus i be shocked to see a mobile phone processor that has that function, thus they can save on the silicon real estate or add functions they need with that saved real estate.
I was hoping AMD would do its own thing regarding this, but yeah we might have to deal with the big / little core rubbish, for me when i am forced to buy such a cpu i would most likely turn it off.
@@ServeTheHomeVideo yes server's and there OS has more support for arm but still small compared to standard x86 thus more times then not arm just sit there idle.
Microsoft should change it's licensing formula for Windows Server & SQL Server. It was not designed for this much of cores. I can't imagine how much would cost just a server.
Also think of VMware where the licensing is no longer aligned to modern cloud native CPUs
Epyc Naples (not Milan written initially) (Zen 1/1+) had the memory controller also on the CPU chiplet before AMD moved it to the I/O die from Zen 2 till now. There is a trade off by having the memory controller on the CPU die as it results in NUMA which requires software/OS to account for it, while memory controller on the I/O die gives you UMA but with a slightly higher latency.
Yes. Zen 1 was Naples though :) Milan was EPYC 7003
@@ServeTheHomeVideo Right, so many Italian cities they start to blur together.
Granite Rapid have a UMA configuration that they enable using specialized cache design. So you can either run it with NUMA domains or without it
@@SY-zv2cz interesting, need to read up on that
Genoa is actually NUMA, each group of three CCDs have three local memory channels. You can switch between 1NPS and 4NPS in the UEFI, same as the toggle between 3NPS and 1NPS for GNR. The difference is really just which mode they default to on either platform.
I would love to see this machine render and run some flip sims
Excellent. Does it still support real mode though?
the legacy BIOS compatibility support module (CSM) was dropped on the Whitley (Icelake) platform, but GNR still comes out of reset in real mode at 0xffff:0000. Eventually it'll all be 64-bit (including PEI which has to run out of cache), but not yet.
I wouldn't be surprised if some BIOS vendors (really UEFI vendors these days) still support CSM, I don't think the hardware prevents it.
@@poofygoof thanks, a few more years of life left in my turbo pascal 6 programming skills yet then.
These are cool and all, but I'm actually more interested to see what the small socket P core options will be. Dual 350W chips are already a lot of heat to deal with in a chassis.
Totally, and the R1S offerings
Great now buy one of these systems for me to game on.
Will they consider adding any overclocking options for that CPU, or at least options to increase the power limit?
Nope, hence the paltry 3.9 GHz max speed per core at the top end. So much for world's "fastest" CPU.
The slumbering giant awakens
Competition is really a great thing!
Cool - now take one compute tile and one IO tile and make HEDT great again.
There are two I/O tiles and one compute tile on Xeon 6700E, but that would not be a super HEDT part
@@ServeTheHomeVideo IMO it'd be too much for HEDT. To me HEDT is 12 to 32 cores, quad channel memory and 48 PCIE lanes. Although in reality, I'd prefer consumer socket was more reasonable with 4 to 8 cores and then HEDT 8 to 16. Let workstation be 16 to 64.
Basically - I just want the golden days back like X99. The days where if you got more cores than you actually needed for a desktop, you got the HEDT features to go with it. TR4 was such a great start for AMD and they had to go and muck it up. I would have absolutely bought Zen 2 on TR4 - but they went and put the core count I wanted (16c) on the socket that doesn't have the features I also wanted - basically doubling the price of everything, consumer socket and everything above it.
@@cracklingice well you're falling behind, because AMD thread ripper is plenty more than that
@@ILoveTinfoilHats AMD Threadripper is not HEDT anymore. It is workstation. HEDT is expensive but within reasonable reach for consumers. There are no Threadrippers that fit that bill anymore. TR4 was an incredibly promising start and I would have probably bought TWO of them by now if they had not killed HEDT.
@@cracklingice How about you Google "What is Threadripper" and tell me what the first result says buddy
Intel must be really struggling with yields on intel 3. The single tile designs get up to 48 cores, for intel to only enable 42 or 43 on the same die in their flagship part there must not be all that many fully working 48 core tiles. The existence of the 6979P is also informative, there's no reason they'd launch that SKU unless they are getting a lot of tiles with 40-42 working cores.
All good for the consumer of course, if yields are amazing on a new process then the designers weren't ambitious enough!
That also leaves open the possibility of 144 core sku if/once yield gets good enough and if there is a demand.
You'll pay more $ for the next generation with same silicon with the cores enabled.
Nice to see intel show up to the party, lets see where the prices go in a month or two when both intel and amd have 128 pcore parts out.
Stay tuned to STH
Welcome back Intel!
In my experience China and the United States tend to have the cheapest prices for computer parts due to lots of competition and a desirable large market (also low taxes and tariffs until very recently). China has the advantage of Shenzhen basically being the scrappers paradise due to low labor cost and the used/refurbed stuff plus being right where the stuff is produced helps. Generally speaking Germany and most of Europe for that matter is more complicated tax wise and logistically speaking (due to multiple languages and distance) and much less competition (although in Germany Mind Factory prices are excellent) Mind Factory is basically the Micro Center of Germany and many Europeans buy and import from Mind Factory.
The funny thing about this whole situation is that many South and Central Americans specifically go to Mexico to buy pc parts to import since it's cheaper that way even when accounting for the plane ticket. Due to the weaker currencies and lower demand plus higher tariffs or taxes it means most people not in Europe/North America basically rely on things like buying from Alibaba such as Zeons (basically intel server cpu) with outdated gpu from 3-8 years ago. Hell anyone around the world can benefit through Alibaba's massive discount on refurbs by buying some am4 cpu or 7500f (basically a down clocked ryzen 7600 without the igpu). Most pc gamers in poor countries tend to have to rely on playing older or easier to run titles like Deep Rock Galactic or free esports games like Leauge of Legends, Valorant, etc or take the risk of flying the seven seas and becoming a pirate.
But..... Will it run Handbrake?
Does it even make sense to talk about processors and sockets anymore? Or isn't this more like 3CPUs in one socket?
No more than AMD epyc is 8 processors imo. Maybe less since latency here should be much lower.
I’d say yes it still does, number of dies doesn’t seem different than multicore in that regard to me personally.
Well you buy in the unit of a CPU that goes into a socket, and there is a difference between on and off socket latency so I think it still makes sense. You are right though that this split into three is more than three Cascade Lake era CPU sockets
@@jolness1 its different, with AMD all CCDs have a direct connection to memory, here memory is divided into 4 channels per die like Zen 1. so its more like a 3 socket system in a single socket.
So AMD requires a hop every time off chip through package and to the IO die before going off package through pins to memory. That has been the case since Rome.
Is it PL1 500W? How do you air-cool this thing? What the sound level?
500W per is not a problem. We showed a NVIDIA GH200 server with a 900W chip in a 1U. www.servethehome.com/supermicro-ars-111gl-nhr-nvidia-gh200-grace-hopper-1u-server-review-arm/
@Masterchief-q8g All of the DC server rack's I've seen were air-cooled, often even without inrow rcs. And I've seen maybe a hundred racks.
Emerald rapids were like 200 to 350 watts per socket.
I can't even imagine the flow needed to cool 1000 watts from the cpus alone, and the power draw of such fans.
Server cpus dont have to be this hot. I've dual socket 6148P scalable running less than a meter away from me when Im sleeping. Under full load the fans alone draw more power than the cpus, storage and networking all summed up.
@Masterchief-q8g And you will probably have 2-4 of them on the board, while its also two separate hotspots on board
@@ServeTheHomeVideo 900W chip is a one thing, but two 500w PL1, or 625W pl2 if intel didn't change anything, is a 1250W for dual cpu system like dl360 or dl380
At last a good news from Intel!
Most people are never going to afford a setup for this processor setup.
This couldn't be more Intel fanboyish
But, these use the old (Meteor Lake class) Redwood Cove P-Cores and note the new Lion-Cove P-Cores that will be in the Arrow Lake CPUs that will be launched in a month or two. To get Lion Cove P-Cores, we have to wait at least another year or two (late 2025, early 2026, in the best case) , for the DIamond Rapids Xeons that _may_ finally use the Lion Cove architecture for its P-Core cpus (and support sixteen memory channels, across two sockets, hopefully).
Yeah let's hope these xeons don't explode after running them for three months.
Have any Xeon’s ever “exploded” after three months? Intel has already RMA’d and extended the warranty on problem 13 and 14th gen CPU’s.
Something tells me I need this
What would a low-end build using that 16-core chip look like? Would that be afforable for a home setup?
We will hopefully get to show you in Q1 when the chips are out.
Imagine... Is a Beowulf cluster of these still a thing?
Where cinnebench at?
I lost my trust in Intel they didn't deal properly witg their 13th and 14th gen issue and now what guarantees their new processors wont suffer from similar issues!
If it has 500mb of l3 cache, couldn't i run a operating system without cache?
with a lot slower, cache is designed to compensate low memory bandwidth. when ram is getting faster, l3 cache will become pointless.
what about price/performance metrics
pls put timestamps in the description
there is only 00:00 intro timestamp in the description
Hi, did not mention the max RAM design. Great overview. Intel is back, for me, they never fully left because in reality Intel always performed far better than the paper suggested compared to AMD. With this, it seems they are going to win on paper and in reality.
Ill wayt for the perf/w testing to be done, looks like a powerhog..........
We talked about power on the main site a bit more. Modern dual socket 500W from AMD and Intel automatically put you into 1kW+ per node systems
How much would it set sumone back for the whole server
all of the features on all of the CPUs - we know why they did that, cuz their competition does that.
Lets first see how many microcode updates it needs to work reliably and not destroy itself. The last time it took 2 years (so far)
But how it works when cpu core from one tile needs data from other tile? In AMD EPYC case all memory is flat, and there is no different what core needs what data. In case of Intel it's like 3 sockets in one. It's ok when you running a lot of VMs, that fit in it's own tiles, but in case of some one big task that uses all cores in all tiles as I understand there will be b performance penalty
We showed the flat mode (called Hex mode) on this in the main site as well. Intel memory is die to die going over very fast EMIB. EPYC has to go chip to package to I/O die. EPYC has a similar feature that you will see "NPS=4". The I/O die on EPYC actually performs better in many cases with data locality as well.
@@ServeTheHomeVideo thnx, will read your article. When Anantech shut down, there is left not much places to read about architecture of modern silicon's design
Interesting, I wonder what kind of workload will be delegated to the E-cores most of the time.
Edit: nvm it's not hybrid big.LITTLE systems like the consumer line since Alder Lake. So it's either all P-cores or all E-cores in each node
Exactly
Even money I’ll be able to afford this CPU before I can afford to replace my GTX1070.
looks nice, but the more important question I have to ask "does intel still hate customers and refuse to replace all those defective processors?"
If the answer is still "fuck our customers, buy a new 15th gen to replace the defective 13th gen we sold you."
I keep seeing fear of innovation, where for example is the use of OMI for memory? Always faster memories, and more cores, "more of the same".
In many ways that is CXL Type 3 right?
@@ServeTheHomeVideo I understand that it is not, I see it as complementary technologies, RAM is still tied to only one type (DDRx), "OMI" makes it possible to have at the same time, DDR, GDDR 4/5/6, or another optimized one, even something exotic like HBM sticks (an OMI chip with HBM layers) or any other novelty that the market wants, as long as it is OMI the technology behind it is irrelevant, but connected directly to the CPU.
The GB/s memory bandwidth scale would finally be broken, TB/s would be the new normal.
I suggest the text "cxl-and-omi-competing-or-complementary" by Bryon Moyer (without a link so as not to cause problems)
@@ServeTheHomeVideo Not to mention the GB/s bandwidth per mm² that allows for many more channels and bandwidth per core.
OMI achieves 29GB/s/mm/2 versus approximately 3.3GB/s/mm² for DDR4/5
i fREAKIN love server partsss
Yes they are super cool
For workloads that can benefit form high core count + the accelerators, these CPUs are real killers. Next, Intel needs to put similar effort into its software stack and unifies everything just like CUDA. I run LLMs on Nvidia and Intel Arc and I can say that the experience is day and night different (in favour of Nvidia).
what do you mean they led with quad core count?
More interested in AMD epyc release, costs, powerusage etc. Intel did so many damage to their brand, it seems its beyond repairable.
The cpu chip is as huge as Patrick's tummy. 😳😳
If the product is unreliable, no one buy it.
Wait for ten years to prove the product is stable.
For now, AMD will answer.
You do realise that the cycle will go on and on for the foreseeable future.
@@mikelay5360 Yeah sure, it's a never ending cycle.
We are under the AMD embargo
I do have thoughts, but I don't have money for those kinds of systems. And I doubt many do. Not sure what this has to do with home though, as these things are clearly designed for data centers.
The STH main site has been the biggest server, storage, and networking review site for almost a decade. We were doing 8x GPU server reviews in 2016-2017.
@ServeTheHomeVideo yep, and all those years I've been wondering what these big servers have to do with home 😀
Intel owns the means of production so is NOT Peasantry. all the other ONES are peasants
Good! This might lower the price on threadripper pro
this have nothing to do with threadripper, and threadripper already on sales
Whats memory bandwidth like compared to Zen5?
We cannot share Zen 5 yet due to embargo
Probably higher, mrdimm boosts speed by a lot. Unless zen5 can do mrdimm, that would be sweet.
Desperately needed Intel W
Yes
Waiting Turin to completive. Just a powerful comeback from Intel at Server/HEDT
Still under embargo for that