I have 2 7402's and every 6 months compare the price of 7H12's or something else and a new mb (using DDR4 because that's $20k of RAM) to exceed the performance difference enough to provide a significant value delta, not just an expensive marginal one.
I can't really tell what conclusion OP's trying to get to, but I think it's that the price/perf of his (already purchased, devaluing) 7402 CPUs are still holding strong against newer EPYC chips, thus it doesn't make sense for his use case to upgrade to a newer platform, especially since it requires buying a whole new motherboard as well.
A used pair of (24c/48t 2.8GHz) 7402s are ~$700. A single new 7643P (48c/96t 2.3GHz) is $2722.
That's a big difference for 48 cores if you're paying for it yourself. The old ones are circa 2019, so not ancient by any means. Same net L3 cache as well.
> The new Milan chips are not for everyone, of course, and are designed for those who are looking for the best price/performance, not the absolute best performance or lowest performance per watt that AMD can deliver in its now wide and deep CPU lineup. But if you are doing HPC on the cheap, you need to take a look at these new Milan CPUs
So middle of the road, like most things for the general customer base.
Who's looking for the lowest performance per watt?
The article makes these processors look like specialty things.
Lowest performance per watt has been the domain of x86 for a while now ;-)
Where they shone has always been highest single-thread performance per capex dollar (not necessarily per opex, which is highly influenced by power consumption).
When you add everything, the CPU is a relatively minor part of the server - a lot of energy is going to storage, memory, networking and other non-application activities, as well as memory and vector pipelines in GPUs (often a lot more than whatever goes into the CPU).
Being less burdened by compatibility, ARM is in a much better place WRT performance per watt. I expect RISC-V to be in a similar position soon.
x86 is a legacy platform, ia64 could've served the niche this article is catering toward, f.e. HPC and data center. Sadly intel squandered Itanium by being actively hostile towards developers
I would guess the "actively hostile" part had a lot to do with all the time and effort to get the compiler to make good use of Itanium. Hard to focus on developers without that foundation. And they never got there.
To give them some grace, it compiled "made for VLIW" code for Itanium well, where it would reap the benefits. It couldn't do much with existing code to make it efficient.
Some combination of strategies. Grabbing an old Itanium Developer's manual, the suggested rewrites include making loops branchless where possible, balancing instruction depth for branches that can't be avoided, doing more comparisons in parallel rather than if X -> then if Y, and so on.
Obviously, this doesn't apply to every general purpose problem. Just noting there were some narrow cases where writing for Itanium produced something close to the promise.
>ia64 could've served the niche this article is catering toward
No it can't. Itanium is a terrible design due to combining both static scheduling and pointless complexity. The worst of both worlds.
If you really insisted on static scheduling, you would implement an EDGE [0] architecture. The idea is quite simple. Programs consist of instruction blocks with a finite size. The CPU performs some very basic scheduling on the block level so the expectation is that the compiler optimizes one block at a time. Additionally, these blocks can reference other blocks and form a graph to make up a complete program. The processor then schedules the instruction blocks dynamically.
Tachyum built such a processor and it turns out they built it out of order anyway.
I have a rack with a few old Dell PowerEdge R710 with dual Intel Xeon X5650 processors (6 cores 12 threads per CPU). I use them to test the changes on my chess engine in winter and they help keeping the house warm. They are very cheap.
The slightly more recent Dell PowerEdge R720 can have CPUs with more cores and use a bit less electricity and noise but they are much more expensive where I live.
I've been on the lookout for a more modern platform with more cores per $100 but couldn't find any yet.
The best performance per dollar in a server for now is if it is built with desktop AMD Ryzen 7000 CPUs, e.g. Ryzen 9 7950X, whose performance per dollar is vastly better than that of any Intel Sapphire Rapids or AMD Genoa or Bergamo or Threadripper. (There are easily available AM5 motherboards with ECC support, and network cards up to those with dual 25 Gb/s ports in SFP connectors are also cheap.)
Ryzen 7000 CPUs are not the best only when the performance of the CPU is less important than having a lot of PCIe lanes and/or memory concentrated in a single box, instead of being distributed in multiple boxes (for the latter case multiple Ryzen 7000 boxes are cheaper than a single big server with the same total amount of memory and total count of PCIe lanes).
For a lot of memory and PCIe lanes in a single box, there are some Sapphire Rapids models with few cores that are quite cheap, but of course the Intel marketing has decided to cripple them by limiting their memory speed to DDR5-4400. The ubiquitous crippling of the cheap Intel SKUs is intended to force their customers to buy the SKUs with outrageous prices, but all that this has ever done to me was to force me to buy from their competition.
For the same purpose, AMD Siena CPUs are expected to be competitive, but only if they will become available at retail, which is uncertain.
Is it nefarious marketing or just that a lower end CPU with 8 memory channels doesn't need a whizzbang memory controller to get the memory capacity/performance it needs? Besides, doesn't the platinum version only go to 4800 MHz anyways? (split comment to separate topics)
Decreasing the DDR5 transfer rate from 4800 MHz to 4400 MHz may provide a minuscule improvement in Intel's CPU die yields, supposing that their manufacturing process is bad and it barely allows to reach the standard 4800 MHz frequency, which might be justified as allowing lower prices for the low-end SKUs.
The users certainly do not want such a speed reduction, especially taking into account that they will usually buy standard DDR5-4800 modules, which will never be used at the performance that they have been paid for.
It is very unlikely that the manufacturing cost reduction due to a relaxed specification for the memory controller is important. Cutting whatever features they could in the cheap SKUs, to annoy the customers (like disabling instruction subsets, and in Sapphire Rapids disabling the accelerators), has been the MO of Intel already for decades.
I'm not saying people directly want lower memory clocks rather I'm saying 4000/4400/4800 MHz doesn't seem to be a decision based on intentionally aiming to create upsell segmentation because 400 MHz of memory bandwidth on an 8 channel system is a completely irrelevant way to incentivize people while using lower quality binning and manufacturing to compete on price IS an incentive people in that range want, even if it comes at a minor cost in performance (that's the point of the low end SKUs!). E.g. Epyc segments memory performance in the Edge line by cutting the number of memory channels in HALF, the equivalent of dropping to 2400 MHz speed in a loaded server. Now there are other additional reasons that's actually attractive to buyers as well, namely the Edge class is expected to be lower power and driving 12 channels of high speed memory is counter to that goal, but it gives an idea of how miniscule 400 MHz of memory bandwidth is in terms of things that matter for low tier server performance.
If anything I'd say that Epyc has support for 12 memory channels (win of the dedicated I/O die) is leaps and bounds more relevant. The Intel SKUs would need to run their 8 channels at 7200 MHz just to break even and here we are talking about 4400 MHz instead of 4800 MHz as a deciding reason to not buy an Intel server!
Every manufacturer has the MO of artificially cutting features to segment things. Intel, AMD, Nvidia, Apple - all of them. Now maybe the folks at Intel marketing were smoking something funny that day and genuinely though this limitation would actually create incentive to go to a higher tier, who knows. Even then, why would that particular (poor) segmentation be a driver be worth choosing to exclude a brand over when the meaningful segments from all players are all much larger?
Except for Intel, the others, like AMD or NVIDIA, cut features that reduce the performance without having a software impact, and which indeed correspond to real reductions in the manufacturing cost, i.e. the number of cores, the size of the cache memories and the maximum clock frequencies.
Only Intel has always disabled features that have negligible impact on the manufacturing cost, but which force various expenses upon the customers, like the disabling of various instructions or of internal peripherals. Due to this Intel policy, there are still plenty of programs that run slower than they should on the latest CPUs, because they may have been built for instruction sets that have become obsolete more than a decade ago, like SSE.
The 10% memory throughput reduction is not negligible. There are 2 main scenarios for using a cheap Sapphire Rapids SKU, one would be to use it to host a very large number of SSDs, and the other would be to host some application that cannot be distributed over multiple servers and which wants a lot of memory. In the latter case, if the application has a performance dependent on the amount of memory, it is likely that it would also benefit from a faster memory, but with Sapphire Rapids if you want a 10% faster memory, which costs nothing extra in the price of the DIMMs, you may need to pay double or triple for the CPU.
I think you'd be surprised by how many identical style limitations others Nvidia has done over the years, from slightly lower clocked memory all the way down to programs that run slower than they should on other vendors because Nvidia code used by games was intentionally deoptimized for other platforms. AMD too, though to a lesser extent since it's been a while since it had its time in the sun from a market penetration perspective.
10% memory bandwidth is actually extremely negligible, particularly for the lower core count SKUs in the tier. The desktop gear you mentioned this all for has 1/3 the memory bandwidth, 1/4 if you plan on putting lots of RAM in, and can get away with a great deal many more tasks than you're describing something with 10% less memory bandwidth as able to do. We're talking a maximum worst case difference, for going between the lowest end Xeon to highest end Xeon, of 10% for the "my workload is to benchmark memory bandwidth only" use case. The reasoning for why it's supposed to be bad seems more focused with the starting assumption anything Intel does must somehow be bad than actually explaining what actual workload impact is had vs other options, and that's the red flag for me here.
You can get a 14900k with better performance (single core and multi core) for cheaper. Between my 7950X system and that I've found the Intel memory controller to handle higher RAM capacities speeds as well. It also comes with ECC support if you get the right motherboard and you get additional usable PCIe bandwidth over what the 7950X will expose.
It is true that the desktop Intel CPUs have lower idle consumption, usually by 10 to 15 W, than the desktop AMD CPUs.
On the other hand, for any given performance at full load, the AMD CPUs have a much lower power consumption, as low as two thirds.
Which is preferable depends on the application. For a personal computer, idle power consumption may be the most important. For a server, the power consumption when used is normally more important.
I have several servers in my home, which are used intermittently. Their idle power consumption does not matter much, because whenever they are not used they are shut down. Wake-on-Ethernet is enabled on them, and whenever I use them Ethernet magic packets wake them immediately.
The Intel E-cores are quite good for mostly integer workloads, like code compilation, or for file servers or Web servers, so for such tasks the Intel CPUs are adequate (but ECC-supporting MBs for them are quite expensive).
On the other hand their floating-point performance is quite poor (and the P-cores are few and intentionally crippled), so for such workloads AMD Zen 4 is much better, especially for running non-legacy programs, which have been updated and optimized for the modern ISAs.
I do not remember now, but there have been a large number published.
The wall plug idle power consumption that I have seen published for recent Intel-based desktops was usually around 30 to 35 W, about the same as of the older Kaby Lake or Coffee Lake based computers that I have.
The wall plug idle power consumption of Ryzen-based desktops, both as published and also valid for a couple of Ryzens that I use, is about 40 W to 50 W. (The idle Ryzen cores consume a negligible total power of less than 1 W, but the I/O die consumes 20 to 30 W, even when idle.)
The desktops that have GPU cards have higher idle power consumptions than these.
The (well-designed) small computers like Intel NUC or similar, which use mobile CPUs, have a wall plug idle power consumption of 5 to 10 W, regardless whether they use Intel Core or AMD Ryzen.
When the idle power consumption is a concern, the real solution is to use a small computer with a volume of less than 1 liter and a mobile CPU as the workstation (there are now models with Ryzen 9 7940HS or Intel Core i9-13900H, which are more powerful than most big desktops from just a few years ago), and to use big desktop computers only as servers that sleep when they are not used.
With a W680 motherboard, the Raptor Lake Refresh becomes significantly more expensive than 7950X.
The multi-threaded performance of 14900K is good only for mostly-integer workloads (e.g. for software compilation it works fine), and even then only with the price of a much higher power consumption, e.g. by consuming 50 W to 100 W extra at equal performance.
For floating-point workloads or for other workloads that can exploit the vector instructions, 7950X can be much faster than 14900K.
Arguably you can buy the mobile variants of the Ryzen 7000 series, like 7940HS/7840HS, as a package, cheaper per core than the 7950X. Depending on the use case it may be a better fit. Definitely if footprint and/or power consumption are main drivers it may be much better.
If you need the I/O or full ECC of the bigger systems then you just need the bigger systems.
Do these new CPUs contain mitigations against Zenbleed?
Unfortunately, there are still cloud providers who haven't rolled out mitigations either because their upstream vendors are slow or because they're afraid of the 15% performance hit.
I wonder how much else these providers are skimping out on to save a few % there too. Maybe a testable thing like this is the litmus test of "this place is nuts, stay away".
41 comments
[ 3.6 ms ] story [ 97.6 ms ] threadThat's a big difference for 48 cores if you're paying for it yourself. The old ones are circa 2019, so not ancient by any means. Same net L3 cache as well.
Isn't it "bring your friends"? I could be wrong (the last time this song played on the radio was in the 90s sometime).
Original: "Load up on guns, bring your friends ..."
So middle of the road, like most things for the general customer base.
Who's looking for the lowest performance per watt?
The article makes these processors look like specialty things.
Where they shone has always been highest single-thread performance per capex dollar (not necessarily per opex, which is highly influenced by power consumption).
When you add everything, the CPU is a relatively minor part of the server - a lot of energy is going to storage, memory, networking and other non-application activities, as well as memory and vector pipelines in GPUs (often a lot more than whatever goes into the CPU).
Being less burdened by compatibility, ARM is in a much better place WRT performance per watt. I expect RISC-V to be in a similar position soon.
I think this is a mistake from the author.
If you can't run actual software faster than your competition, your theoretical benefits remain just that.
Obviously, this doesn't apply to every general purpose problem. Just noting there were some narrow cases where writing for Itanium produced something close to the promise.
I remember their Xeon Phi ones were very nice as well. I miss those machines.
No it can't. Itanium is a terrible design due to combining both static scheduling and pointless complexity. The worst of both worlds.
If you really insisted on static scheduling, you would implement an EDGE [0] architecture. The idea is quite simple. Programs consist of instruction blocks with a finite size. The CPU performs some very basic scheduling on the block level so the expectation is that the compiler optimizes one block at a time. Additionally, these blocks can reference other blocks and form a graph to make up a complete program. The processor then schedules the instruction blocks dynamically.
Tachyum built such a processor and it turns out they built it out of order anyway.
[0] https://en.wikipedia.org/wiki/Explicit_data_graph_execution
The slightly more recent Dell PowerEdge R720 can have CPUs with more cores and use a bit less electricity and noise but they are much more expensive where I live.
I've been on the lookout for a more modern platform with more cores per $100 but couldn't find any yet.
Ryzen 7000 CPUs are not the best only when the performance of the CPU is less important than having a lot of PCIe lanes and/or memory concentrated in a single box, instead of being distributed in multiple boxes (for the latter case multiple Ryzen 7000 boxes are cheaper than a single big server with the same total amount of memory and total count of PCIe lanes).
For a lot of memory and PCIe lanes in a single box, there are some Sapphire Rapids models with few cores that are quite cheap, but of course the Intel marketing has decided to cripple them by limiting their memory speed to DDR5-4400. The ubiquitous crippling of the cheap Intel SKUs is intended to force their customers to buy the SKUs with outrageous prices, but all that this has ever done to me was to force me to buy from their competition.
For the same purpose, AMD Siena CPUs are expected to be competitive, but only if they will become available at retail, which is uncertain.
Hetzner, OVH and Contabo are some of the only few I’m aware of who does.
The users certainly do not want such a speed reduction, especially taking into account that they will usually buy standard DDR5-4800 modules, which will never be used at the performance that they have been paid for.
It is very unlikely that the manufacturing cost reduction due to a relaxed specification for the memory controller is important. Cutting whatever features they could in the cheap SKUs, to annoy the customers (like disabling instruction subsets, and in Sapphire Rapids disabling the accelerators), has been the MO of Intel already for decades.
If anything I'd say that Epyc has support for 12 memory channels (win of the dedicated I/O die) is leaps and bounds more relevant. The Intel SKUs would need to run their 8 channels at 7200 MHz just to break even and here we are talking about 4400 MHz instead of 4800 MHz as a deciding reason to not buy an Intel server!
Every manufacturer has the MO of artificially cutting features to segment things. Intel, AMD, Nvidia, Apple - all of them. Now maybe the folks at Intel marketing were smoking something funny that day and genuinely though this limitation would actually create incentive to go to a higher tier, who knows. Even then, why would that particular (poor) segmentation be a driver be worth choosing to exclude a brand over when the meaningful segments from all players are all much larger?
Only Intel has always disabled features that have negligible impact on the manufacturing cost, but which force various expenses upon the customers, like the disabling of various instructions or of internal peripherals. Due to this Intel policy, there are still plenty of programs that run slower than they should on the latest CPUs, because they may have been built for instruction sets that have become obsolete more than a decade ago, like SSE.
The 10% memory throughput reduction is not negligible. There are 2 main scenarios for using a cheap Sapphire Rapids SKU, one would be to use it to host a very large number of SSDs, and the other would be to host some application that cannot be distributed over multiple servers and which wants a lot of memory. In the latter case, if the application has a performance dependent on the amount of memory, it is likely that it would also benefit from a faster memory, but with Sapphire Rapids if you want a 10% faster memory, which costs nothing extra in the price of the DIMMs, you may need to pay double or triple for the CPU.
10% memory bandwidth is actually extremely negligible, particularly for the lower core count SKUs in the tier. The desktop gear you mentioned this all for has 1/3 the memory bandwidth, 1/4 if you plan on putting lots of RAM in, and can get away with a great deal many more tasks than you're describing something with 10% less memory bandwidth as able to do. We're talking a maximum worst case difference, for going between the lowest end Xeon to highest end Xeon, of 10% for the "my workload is to benchmark memory bandwidth only" use case. The reasoning for why it's supposed to be bad seems more focused with the starting assumption anything Intel does must somehow be bad than actually explaining what actual workload impact is had vs other options, and that's the red flag for me here.
As much as I like AMD, I find this a significant advantage for a workstation / small server.
On the other hand, for any given performance at full load, the AMD CPUs have a much lower power consumption, as low as two thirds.
Which is preferable depends on the application. For a personal computer, idle power consumption may be the most important. For a server, the power consumption when used is normally more important.
I have several servers in my home, which are used intermittently. Their idle power consumption does not matter much, because whenever they are not used they are shut down. Wake-on-Ethernet is enabled on them, and whenever I use them Ethernet magic packets wake them immediately.
The Intel E-cores are quite good for mostly integer workloads, like code compilation, or for file servers or Web servers, so for such tasks the Intel CPUs are adequate (but ECC-supporting MBs for them are quite expensive).
On the other hand their floating-point performance is quite poor (and the P-cores are few and intentionally crippled), so for such workloads AMD Zen 4 is much better, especially for running non-legacy programs, which have been updated and optimized for the modern ISAs.
The wall plug idle power consumption that I have seen published for recent Intel-based desktops was usually around 30 to 35 W, about the same as of the older Kaby Lake or Coffee Lake based computers that I have.
The wall plug idle power consumption of Ryzen-based desktops, both as published and also valid for a couple of Ryzens that I use, is about 40 W to 50 W. (The idle Ryzen cores consume a negligible total power of less than 1 W, but the I/O die consumes 20 to 30 W, even when idle.)
The desktops that have GPU cards have higher idle power consumptions than these.
The (well-designed) small computers like Intel NUC or similar, which use mobile CPUs, have a wall plug idle power consumption of 5 to 10 W, regardless whether they use Intel Core or AMD Ryzen.
When the idle power consumption is a concern, the real solution is to use a small computer with a volume of less than 1 liter and a mobile CPU as the workstation (there are now models with Ryzen 9 7940HS or Intel Core i9-13900H, which are more powerful than most big desktops from just a few years ago), and to use big desktop computers only as servers that sleep when they are not used.
The multi-threaded performance of 14900K is good only for mostly-integer workloads (e.g. for software compilation it works fine), and even then only with the price of a much higher power consumption, e.g. by consuming 50 W to 100 W extra at equal performance.
For floating-point workloads or for other workloads that can exploit the vector instructions, 7950X can be much faster than 14900K.
If you need the I/O or full ECC of the bigger systems then you just need the bigger systems.
Unfortunately, there are still cloud providers who haven't rolled out mitigations either because their upstream vendors are slow or because they're afraid of the 15% performance hit.