Best CPUs for Workstations
anandtech.com
anandtech.com
It is still rough though, mostly around the mobo. The Gigabyte Aorus Xtreme mobo doesn't work out of the gate: https://www.reddit.com/r/gigabyte/comments/enatk1/gigabyte_t..., https://www.reddit.com/r/linuxquestions/comments/ankjfx/inte...
Ubuntu 18.04 also has an issue with the dated kernel on 3970x: https://www.phoronix.com/scan.php?page=news_item&px=Linux-Bo... that can be workaround by mce=off. Due to combination of that and NVIDIA driver, I did 20.04, and it has been stable so far.
The only Ryzen/Linux issue I've heard of is on launch week of Zen2 some distros of Linux were broken by an AMD bug which was fixed pretty quick.
In fact performance is better for many cored Ryzens than windows usually.
If you want something that just works, buy something with Linux preinstalled and supported. If no vendor offers the distro you want, buy from a Linux vendor anyway and install what you want. When people buy Windows hardware then complain that Linux doesn't work it sends the wrong message to the market.
System76 is known for simply rebranding cheap Clevo/Sager laptops.
Figuring out which Clevo laptops can run Linux and which ones can't is work that is worth money.
But when I bought it a couple of years ago, I experienced some lock-ups, about once a month, especially when the CPU was idle. I replaced the "gamer" motherboard with [1] and the problems disappeared. I think a lot, but possibly not all, of problems you see online are caused by gaming community itself. For whatever reason, that culture/market is plagued by premature/buggy product launches (motherboard+UEFI) and oftentimes CPU has nothing to do with user issues.
[1] https://www.asrockrack.com/general/productdetail.asp?Model=X...
But when you connect to a Windows VM via RDP, there's an option to pipe sound back to the client via a virtual sound card. On a gigabit ethernet, you can watch Youtube videos this way. :)
No-drama install on Ubuntu 19.10 (18.04 didn't boot due to drivers?). No kernel panics or even X lockups. It's wonderful to have such a fast development box.
If the stock cooler is too loud, the Noctua NH-D15 is silent until full load when it just audibly whispers.
(I'd been developing PhotoStructure on older hardware to make inefficient code paths more immediately painful, but I needed a beefer box to replicate parallelism issues that some of my beta users were seeing. Thankfully, those issues were immediately apparent, and library syncs are dramatically faster now).
[0] https://www.asus.com/us/Motherboards/TUF-GAMING-X570-PLUS/
Linux problems only crop up with heavily integrated custom systems. Typically laptops that cram a ton of shit into the smallest possible space, and then try to work around incompatibilities with extensive driver customization. Typically, if you can wipe a system's hard disk, install Windows on it, and only install drivers from the chipset manufacturer rather than from the brand of the laptop, and everything works, you'll be fine. But if your Lenovo laptop has a Realtek wireless chip in it, and the driver downloaded directly from Realtek is flaky, but the driver downloaded from Lenovo works, it will be flaky in Linux too. The laptop brand isn't going to spend the money to monkey with some weird driver because they like spending money - they're spending that money because they're doing something the chipset maker doesn't support.
These problems almost never happen on desktops with full sized ATX motherboards. First of all, the expectation is that you'll buy a copy of Windows and the motherboard manufacturer doesn't have the opportunity to slip in their pile of custom drivers and other assorted crapware into a custom Windows install. Second, with plenty of space on the motherboard there's no need for fancy tricks that fall down when you look at them funny.
PopOS has been rock solid on it. I have 19.10 but will be upgrading to 20.04 this weekend. As someone who doesn't want to spend all day futzing around with my system and just get work done, I'm very happy with this distro.
I assume the more common consumer processors have been tested more than the thread ripper ones, being they are more common and out longer.
Registered ECC memory is for larger amounts, mostly used in servers or extreme workstations.
https://en.wikipedia.org/wiki/Registered_memory
In reality, registered non-ECC is uncommon because the situations that create the need for RDIMMs typically coincide with the need for ECC.
For TR and Ryzen you have to buy unregistered DIMMs, aka UDIMMs. This is documented on the data sheets for the CPUs as well as the motherboards.
Never had any issues with any of my desktop hardware before. Did a 3900X+Gigabyte Aorus Elite X570 build in december. Aside from overheating and throttling issues (which I mostly sorted out), everything was fine for a while, but then my mobo straight up died in February. I tried everything possible with no success, it wouldn't even get to the BIOS screen, not even talking about the OS load screen. The error code on the mobo was pointing at the GPU, but that definitely wasn't it (that GPU worked fine for years before, and currently works fine in my build after I got rid of that Gigabyte mobo).
Then I saw a friend doing a build in March with the same components. I didn't think much of my mobo dying, since random accidents happen, so I didn't really discourage him from getting the same mobo. His mobo died on day 2 after the build.
I got a X570 asus mobo instead as a replacement, and had zero issues since. Neither with stability/perf nor with overheating. Aside from swapping the mobo and adding an extra SSD, I had made zero changes to my build, so I blame solely the mobo.
Things are running stable, and I didn't had any overheating issue (even though people told me that board would never work due to too week VRMs on B450 series).
My bad, I should've specified that I was talking about their X570 AORUS lineup, not all AORUS mobos in general, that one is on me. Since yours is B450, I wouldn't expect the same issues to surface as in the X570 mobos, makes sense.
I had no install issues. It ran rock solid under several days of memtest86 and prime95, the latter on both Windows and Linux.
My only "complaint" is that it cannot maintain full turbo indefinitely on even a beefy air cooling setup in a big case with massive airflow. I have 7x140mm fans in the case and on the cooler. The case maintains ambient temperature, except for the (short) stream of air from the CPU cooler to the rear exhaust. Under sustained synthetic load, the CPU will run at >4.4GHz all-core until the heat sink is saturated (usually ~10 minutes), and then slowly stabilize down to a 3.9GHz-4.1GHz range for all cores and stay there. An eventual project will be a water cooling loop with much radiator surface area.
Sure, it has half as many cores as the 3960X Threadripper listed, but 12C/24T is still a lot.
Also, it's much easier to cool -- 105W TDP vs 280W on the Threadrippers.
I built one, a Noctua NH-D15 can cool it at 16*4ghz at 700RPM, so basically almost inaudible even at full load.
No ECC, though.
Is ECC useful for a workstation? I thought it was useful for shared servers (VM hosts) to protect against rowhammer & co, but does is have other uses?
Without ECC, modern systems (>= a few GB RAM) will have bits flipping pretty much daily.
Ultimately these flips are caused by charged particles hitting the memory module (a so called 'single event upset' or SEU), and the number of charged particles hitting the modules has not increased with density (although the modules are more sensitive to SEU, it's a smaller effect).
https://storage.googleapis.com/pub-tools-public-publication-...
I think the environment in racks for nodes in whichever formfactor in racks is toxic. Be it EMF interference, power distribution issues, vibrations, and/or temperature. You don't have that in a single system at home, when it is built in a good way, and has stable power. Or at least to a lesser degree.
Until they happen more often that expected. It's something that varies based on building, temperature, etc. environmental factors. And of course on your luck with memory module lottery.
https://storage.googleapis.com/pub-tools-public-publication-...
I assume the 3950X is the same way. The hardware is there, the official blessing is not.
edit: I mean near the DIMMs. Don't try this if you are afraid, or need your hardware. This is a test i'm doing with every build, or any system i have to fix. And it crashes every system, though not at the same places and in the same ways. So far it didn't damage anything permanently, but no guaranties it won't.
edit: By searching for networks i mean for the telcos, not wifi. Then it uses all bands it is capable of.
But I'm of course not suggesting to run it in this state whil doing anything important.
Ryzen has always been very "sensitive" to low-clocked RAM and shows some pretty good scaling as you increase frequency and tighten timings and the ECC UDIMMs on the market are just not binned to do that.
Or convince AMD to stop artificially not supporting registered memory on Ryzen/Threadripper -- the early Athlons supported both.
I wonder if those would run on my Haswell-E xeons. They support RDIMMs, and they are unlocked so nominally they should support higher RAM clocks too (although I've never tried it with RDIMMs, only UDIMMs).
More recently, the W-3175X (Threadripper's real competitor) supports RDIMM as well. So Threadripper actually offers less features than the competition here.
I really would like to see the Threadripper series merged with the 7002P series though. Right now there is no option that gives you both Threadripper-level clocks and large quantities of RAM, the frequency-optimized parts are still a large compromise in performance and aren't widely supported. There is no part that gets you to 4 GHZ on 64 cores with 512+ GB of RAM, despite all of these capabilities existing in their respective segments.
When you are shipping a 64C HEDT processor there is really no longer any meaningful distinction between that and a server processor. And 256GB simply is not enough RAM for a 64C processor. AMD could still upcharge on the multi-socket models, but they are drawing lines between product segments that no longer make sense now that they have pushed core counts so high.
Basically, give us a modern successor to the Intel X5670 series. It's the top of the line, you're paying $4000 for a workstation processor, you should get the whole processor and not have a bunch of disabled features.
The really funny thing is, it's not even a money thing, Epyc isn't even priced significantly higher. 7702P is $4150 at Provantage (HPE p/n P16660-B21) vs $3990 for the 3990X. The 7402P (HPE p/n P16664-B21) is actually $250 cheaper than the 3960X. AMD is no longer using the "traditional" model where workstation is cheaper than server. Workstation is now just as expensive as server. So it's not like giving you Epyc capabilities would mean AMD foregoing any money.
AMD is just drawing arbitrary lines between these product segments because they can.
They should've done that when they changed the socket for the latest generation. Just admit that the existence of Socket TR4 was a mistake and release the new Threadrippers on Socket SP3.
I suspect there are even some people interested in dual socket workstations with 2TB of memory that nonetheless have a 4.5GHz boost clock, because workstations often have mixed workloads. And there is no point in doing market segmentation when all the prices are the same anyway, it's just incompatibility for no reason.
Cooling would become a big issue, though. Air cooling a 280W 3970X in a big tower with sufficient airflow to maintain ambient temperature internally is difficult. I cannot maintain max turbo on all cores.
Putting this in 1 or 2 Us in a server would likely be impossible without water cooling (requiring piping to external liquid coolers (whether radiators or chillers)). There's only so much that increasing air velocity can do in cooling.
I don't know if this is the primary thing holding back Epyc specs, but I know it will be an issue.
huh, where ? from what I see here the 3900X is listed at 450€ and the 3950X at 900€ on average in the mainstream websites...
If you've gotten locked into needing that, you have no other options aside from not using hardware that you can currently slot.
AMD entirely eliminated the lower end HEDT parts and raised prices for TR3.
Intel has a nasty shortage, too.
Compare that to x99, which had an entry level msrp of just $350, but often went for exact same price as 7700k, significantly under $300.
HEDT is in a bad value proposition right now, but if you're stuck needing it for various reasons, your options are limited, hence price increases.
That's exactly the area where I argue there is some fuzziness.
For PCIE lanes, it's true that AMD only has 24 lanes (4 of which dedicated for the chipset) but it has the equivalent of 48 PCIE gen3 lanes in terms of raw bandwidth. Now, it's still not as flexible as having that many actual PCIE gen3 lanes, but high-end x570 motherboards do support some level of bifurcation. You also need compatible devices, but we already compatible storage. There are a couple GPUs out now, but given the upcoming launches of RDNA2 and Ampere GPUs, I expect there will be plenty more compatible GPUs. X570 is somewhat limited now, but as more gen4 devices are released that chipset will only become more versatile.
For RAM, there you have another fuzzy spot. I mean sure, HEDT usually has 4 channels, but what does that mean in terms of memory capacity and bandwidth? Only within the last year or so have 32GB UDIMMs been available (iirc) so HEDT was limited to 128GB until very recently. And if you really need more, you can use an Optane storage device as a swap partition (It's been done before, it's fine). As for bandwidth, 2 channels is a bit of a limitation, but Ryzen does support faster memory, which can help bridge the gap a little bit to quad channel Intel parts running at lower RAM speeds.
I'm not saying that Ryzen + X570 is HEDT, but it's so darn close that I can understand why AMD chose not to produce low-end TR 3000 parts. Between Ryzen 9, TR 2000, Epyc 7002P, and limited use of TSMCs fabs, I think the move they made was smart.
> HEDT is in a bad value proposition right now, but if you're stuck needing it for various reasons, your options are limited, hence price increases.
I don't think it's a bad value prop overall. We have a $4000 CPU that's beating out two $10,000 CPUs (dual Xeon Platinums), which is a phenomenal improvement in value. I think we are in this weird space where, when trying to buy the system capabilities you need (in terms of cores, clock speeds, memory capacity, and expansion slots) you sometimes have to pay a bit extra for capabilities that you don't need. That's unfortunate and I hope next gen fixes that. However, it doesn't seem like AMD or Intel will have new HEDT parts for at least a year.
I ssh in from my XPS 13 running Fedora and use X forwarding to do work on it. It's so much faster than trying to do everything on an ultrabook. Not to mention quite a bit cheaper than a high-end laptop.
The VSCode remote extension installs extensions onto a VSCode remote client installed on the remote machine. So the C++ extensions spawns rg processes on the remote which is a lot faster.
I was under the impression that sshfs is based on FUSE, which works in userspace.
It sounds like an amazing build, but definitely not needed for most tasks. The XPS 13 seems a little wasteful in the setup. You could practically do it with a chromebook.
-U series (somehow, magically the same designation between AMD and Intel) will typically run ~15W, though this is configurable by the OEM (sometimes in BIOS?).
-H series (also, magically the same designation) will typically run ~45W, and this is also configurable by the OEM.
The increased power envelope of -H series typically allows higher base clocks. Turbo can sometimes be close between a 15W and a 45W chip, but a -H series will be able to maintain higher frequencies for longer, thanks to a superior thermal solution.
I think some may discount Ryzen 9 because it's technically not "HEDT", but the market boundaries have changed over time.
HEDT used to include all CPUs more $400 dollars, all of which shared the same socket (for a given generation). Mainstream was on a separate socket and all below $400.
However, Broadwell-E and Skylake-X moved the top end of the HEDT from $1K to $2K, which meant that HEDT spanned a huge segment.
What AMD did with Ryzen 9 and X570 was to take the middle segment of that HEDT market and make it it's own segment.
It did carve out a pricing segment there, but the really cool thing is that it's so compatible, backward and forward with other CPUs in the Ryzen family. There is huge upgrade potential.
It would be entirely reasonable to build a system around an R7, with the expectation that its workload will grow, and throw in a discounted 3950X in a year when Zen3 hits.
[1] Well, has them for a reasonable price. Newegg as third-party sellers offering a few of them for 60%+ markups, but I feel better paying $500 for a $500 board than I do paying $275 for a $160 board.
If only 12-16 cores with 64GB of RAM are enough to be useful, then the Ryzen 9 gets A LOT of work done for an unprecedented low price. But if the work being done can make use of 128GB quad-channel RAM and 24-32-64 cores, then the $350 extra for a Threadripper motherboard is nothing. Doesn't take many billable hours to pay for it :)
I'd not ignore Ryzen 7 and Ryzen 9. If you can work with 128GB ram (and especially if you just want a regular strong system with the usual 16GB) 3700X, 3900X and 3950X are really strong processor and apart from the last one a lot cheaper. Most developers don't even need to go that high, not even on a workstation, one based on the Ryzen 5 3600 would not be weak.
AMD Ryzen 5 3600 currently rocks 103 CPU-Mark/$ whereas Threadripper 3960X has 40 CPU-Mark/$. You get a lot of power out of Threadrippers but you loose a lot of value.
I wasted 6h of research trying to find a small machine with R1606G (or any of the V1000's) with dual 10GbE NICs, and failed. :(
Even Sapphire who promised such a machine retracted their pages and now only their dual 1GbE NIC offering are being sold[0]. The dual 10GbE NIC variants only exist in two PDFs[1][2] now.
I went through AMD's official list[3] of embedded offerings based on the latest R1000/V1000 platforms but exactly zero of the options have the promised dual 10GbE NICs.
Can somebody point me in the right direction? I need a strong and power-efficient embedded NAS with 1 or 2 10GbE NICs that covers several other homelab/server needs and I feel the R1606G is the best compromise in terms of money, power, wattage expense, form factor and required cooling (ideally I'd keep that thing in my bedroom or living room and want to never hear its fans but hey, it can sit in a closet as well).
I am willing to look elsewhere, it's just that the latest R1000/V1000 offerings seem to be so good!
[0] https://www.sapphiretech.com/en/commercial/ipc-fp5v-1ge
[1] https://www.sapphiretech.com/-/media/sites/sapphire/product/...
[2] https://www.sapphiretech.com/-/media/sites/sapphire/product/...
[3] https://www.amd.com/en/products/embedded-minipc-solutions
[0] https://www.ebay.com/itm/Mellanox-ConnectX-3-QDR-QSFP-Infini...
https://www.servethehome.com/hpe-proliant-microserver-gen10-...
Depending on how many hdd's you're trying to use, it might be the go. Low power (energy use wise), but decent cpu, already assembled, etc.
It's not as small as a NUC, but the older generation one I have here is about 1ft x 1ft x 1ft (easily fits on a shelf, under desk, etc), and is pretty much silent.
Do you use yours as a NAS + media server? That's what I plan to use mine for.
It runs FreeNAS (works "out of the box"), which does have media server capabilities and many people apparently use them. I just have no real need for media server stuff personally. :)
If you put a Noctua L9a on it, anything will be silent (although you may need to set fans manually on a server board, or use the included "low noise adapter" to drop the fan speed). That plus a 3600 (or maybe a 3100/3300X if you can source one yet) would be a good little NAS build. Has IPMI, which is nice if you will be running headless.
Or if you don't care about IPMI (if it's not going to be running headless) then you can use pretty much whatever mITX/mATX board you want.
Trouble is, I can't find it anywhere except in the US Amazon, which will set me back like 30-40% premium on top of the $450 that the motherboard costs. :(
If I am going for that kind of money then I'll probably just shell out $1000 for some of the extremely good Xeon D mobos.
Wow
Still sitting there at that price.
https://buy.hpe.com/us/en/servers/proliant-microserver/proli...
> From $109.99
Sounds like the opportunity there, is in finding a place that does price matching. :)
I am thinking about going for the Epyc 7502p with 32 cores but having a hard time knowing if the Threadripper counterpart with 32 cores would be better given its higher frequency.
Guess it is very hard to answer but which one would you go for? Will the threadripper perform much better?
On HN I still don't know if it's an easy way to get upvotes, or the best way to get down-voted hell.
(I haven't checked if they've different cache sizes ;>)
The 7502p is better in situations where you need more memory (either bandwidth or capacity) or you need to to consume less power (such as in a server deployment or a fanless case).
There are almost no cooling solutions that can passively dissipate the heat from a 180W TDP CPU. I am not aware of any commercially available at present. Fanless is not really in the cards for any of these CPUs.
Thanks for sharing!
Ninja edit: Ooof $700-$1400 for the chassis and custom fab for the CPU block. If I'm going for silence, I'd probably just spring for the radiator case instead: https://www.youtube.com/watch?v=dfbcD248n4Y&t=361s
Ninja edit 2: Looks like there's a retailer (rather than going through the manufacturer to Taobao) in the US selling it for $1090. Unsure where to price out a custom CPU heat spreader to integrate with this (per the Anandtech article, that was a custom component Turemetal made for one customer). Either way this is awesome to see available.
No mention of POWER9, for example (my own personal favourite), which I’m running on my workstation (https://www.raptorcs.com/TALOSII/)
There’s several other ISAs out there (ARM, obviously, but desktop-, let alone workstation-class chips are hard at least for now a dicy proposition).
EDIT: Surprised by all the downvotes. Is mentioning anything outside the domain of x86 considered controversial nowadays?!
Well yes, it probably is. That's another reason why everyone is ignoring it.
What I was referring to as possibly dying was POWER on the desktop generally, and Raptor/Talos specifically.
I also don't like that some languages have become a niche within a niche with similar consequences.
I like the idea of a POWER9 desktop machine, but I don't see the pricepoint ever getting to the point where I'd seriously consider one.
Nobody will consider an alternative platform until it becomes affordable.
Intel's main selling point is AVX512 support, which is a niche but it does provide benefits in some workloads. AMD's "more cores" approach is probably superior in most applications however. (It gets complicated: AVX512 uses more power, causing frequency drops. But SIMD-parallelism is innately more power-efficient than more cores... it depends on how much your computer uses AVX512, how sensitive it is to frequency slowdowns... etc. etc)
Well, now this is the case, yes. Wasn't when I bought the iMac Pro. It's a very respectable machine and it has the same drawbacks as the Threadripper -- only 4 channels of memory (with 10c/20t CPU).
> Intel's main selling point is AVX512 support, which is a niche but it does provide benefits in some workloads. AMD's "more cores" approach is probably superior in most applications however. (It gets complicated: AVX512 uses more power, causing frequency drops. But SIMD-parallelism is innately more power-efficient than more cores... it depends on how much your computer uses AVX512, how sensitive it is to frequency slowdowns... etc. etc)
Heard about that, it's quite embarrassing for Intel. As for SIMD itself, it's a sad story. Most, if not 99%, of the compilers out there don't utilise it. Makes me want to contribute to a compiler but I can't seem to find such a job, especially having never done it before.
I had high hopes for RISC-V but I already heard the usual BS about "extensions that vendors may choose to implement or not" and I kind of lost hope. :(
POWER9 is very interesting, especially with it's open licensing. It's even in the relative price and performance category as Intel Xeons, and I'm hopeful that POWER10 on Samsung 7nm will also be competitive.
But POWER9 doesn't have the software compatibility that x86 does, nor is it remotely close in price or performance to AMD. At best, POWER9 might receive a special mention as a niche product (along with the Arm Ampere workstation[0]).
Also, RaptorCS has had problems with their support process. I know they've been working on it, but that doesn't mean they've worked through all of the teething issues.
[0]https://www.anandtech.com/show/15737/arm-development-for-the...
A little older (~2017), but the whole system at a lower price than a 3970X, for similar performance on my embarrassingly parallel workloads, at similar TDP.
Not saying your machine isn’t nice, but the CPUs were probably not worth it. Each of the 30000 series Threadrippers will be about 50% faster in single core benchmarks than the 2673v4 and have better multi-core results.
Overall, I am a convert to workstations with ECC. Worth it the peace of mind that I won't corrupt my files and that it won't randomly crash. Had my share of those over years.
https://www.anandtech.com/bench/product/2125?vs=2519
The 3900X has 50% more cores and Blender is embarrassingly parallel, so it's more than 50% faster, as expected.
That's the real question for whether it's worth it. For single threads the 3900X will be 10-20% faster. For well-threaded applications it's >60% faster.
Keep in mind that the 2700X still has a resale value of over $200, so if you're selling one and buying the other it's not a very expensive upgrade. But the value proposition is a lot better if you can use the extra cores. (Of course, if you can't and you just want the single-thread performance improvement, there's always the 3700X.)
Their graphs are completely fictional, probably as a result of bad data scraping code since it looks like they're just pulling results (inaccurately) from other sites.
If the 2700X is not on the source page that page either got changed or it's a combined benchmark that was presented in multiple articles. In which benchmark do you see that?
https://www.pc-kombo.com/us/benchmark/apps/cpu/compare?ids%5...
This is the Anandtech benchmark that it links to from the first Blender graph (you can click on the graph):
https://www.anandtech.com/bench/CPU-2019/2247
The numbers on that page for the 3900X and 2700X bear no resemblance to the ones in the graph that links to it.
Compare to their first 7zip graph which links here and contains the actual numbers for the 3900X and 2700X, because they evidently scraped that one correctly:
https://www.anandtech.com/bench/CPU-2019/2240
This is the page they link for the second Blender graph:
https://www.computerbase.de/2019-11/amd-ryzen-threadripper-3...
The numbers in their graph don't match what that page has for the 3900X and the 2700X doesn't appear on that page at all.
I understand that the number in that specific benchmarks is confusing right now. Again, that's from the conversion done for the "lower is better" benchmarks. That's why it's not the same number, just the same distance. I hope a better solution can be used soon.
For both the Anandtech and Computerbase Blender results, the 2700X takes ~60% longer in reality, which is ~66% more renders per second (which is the same percentage more per day), but the difference in the length of the bars on pc-kombo.com is <10% for Computerbase and ~1% for Anandtech. Something has gone wrong there.
...
What people are going to care about is the difference as a percentage. :)
> Look how small the bars are at the anandtech page, https://www.anandtech.com/bench/CPU-2019/2247 also does not show you the one is 66% faster - same problem, just inverted.
Actually the bar for the 2700X on the Anandtech page is 66% longer than the bar for the 3900X, though it's admittedly harder to eyeball when they're both that short because of the much slower processors at the bottom of the page. You can imagine how much easier that page would be to see if they didn't include those two Celerons at the bottom. (You can also imagine how that would affect the results using the math you were using; if adding or removing a CPU changes not just the length of the bar but the actual numbers for an entirely different CPU, something's fishy.)
Doing 1/X for timed benchmarks helps that for the modern processors though, because it makes it so the bars that are short and hard to eyeball are the slowest processors rather than the fastest ones.