For the first time in a decade, AMD’s CPU portfolio is positioned to not only serve one or two segments of the workstation market, but its entire breadth. Based on its Zen CPU architecture the company has rolled out four brands to power virtually any type of workstation model a professional customer may be interested in — from the single socket mainstream to the ultra-high end, no-compromise dual sockets. Most recently, we got a chance to see an example of the latter, with Velocity Micro launching the ProMagix HD360A an impressive dual-socket, 64-core beast of a workstation built around AMD’s server-focused Epyc brand processors.
AMD’s Up and Down History in the Market for Workstation CPUs
For all intents and purposes, the company’s CPUs have been absent in the space for the better part of the last decade. AMD did have some success prior to that, first and foremost in the form of the Opteron-branded Hammer, a processor that quite literally disrupted high-performance computing markets for servers and workstations. Launched in 2003, Hammer introduced two key innovations — 64-bit instruction set extensions and direct-attach memory — both of which allowed Opteron to leapfrog Intel, especially in dual-socket configurations common in workstations and servers. On the back of Hammer, AMD began taking meaningful share from Intel, peaking in the years from 2006 to 2008.
Ironically, however, the excellence of Hammer eventually worked against it, by spurring on its chief rival. Intel begrudgingly followed AMD’s lead with both 64-bit extensions and direct-attach memory, helping bring its CPUs back to a leadership position. Meanwhile, AMD stumbled a bit on Hammer follow-ons, particularly with Barcelona in 2008 and subsequent core updates. The combination doomed Opteron in workstations, and AMD’s hard-won sockets at Sun, Fujitsu, IBM and HP all gradually disappeared, leaving AMD with virtually no CPU presence in workstations after 2010.
But now the company has Zen, a designed from-scratch x86 architecture not coincidentally spearheaded by Jim Keller, the same technical force behind Hammer of a decade prior (and now ironically on board at Intel). The Zen architecture looked credible in its goal to bring AMD CPUs back to a competitive position in performance-oriented markets. Ever since the first Zen disclosures, I’ve been left wondering and writing about when AMD might use it to make a serious assault on the workstation market again.
Well, that time has arrived as Zen products bearing a few different brands — Ryzen 7, Ryzen Threadripper, Ryzen Pro, and Epyc — are all positioned to potentially begin taking back workstation CPU share, from entry-level single-socket (1S) all the way up to max-configured dual-socket (2S) monsters. On the dominant deskbound side of the market, AMD’s Zen-based lines now serve all four of the de facto standard product tiers: Entry 1S, Premium 1S, Entry 2S and Premium 2S.
AMD’s Zen-based family of CPU products can now legitimately serve the breadth of the deskside workstation market |
The first of those CPUs to make some workstation noise was Threadripper, adopted by several vendors including Puget Systems, Velocity Micro, NextComputing, Maingear, and Boxx, showcased in the Apexx 4 model 6301, a machine I previously reviewed that fits in the Premium 1S category. A few Ryzen/Ryzen Pro machines have also emerged. Finally, thanks to this ProMagix HD360A from Velocity Micro, AMD has their first dual-socket Zen-based workstation in the market. And thanks to Velocity, we got a chance to review it, both stand-alone and viewed in context of its lower-end sibling CPUs.
The first Epyc workstation … and a dual-socket, max-core’d one at that
Targeting the absolute thinnest of the workstation market ultra-high-end thin air, the Velocity Micro ProMagix HD360 is certainly not for everyone. I’ll get into more about who exactly it is for, but rest assured the vast majority of the workstation market won’t have an interest in this product. And that’s no knock, but rather a reflection of the purpose of this machine, which simply put is to provide the highest possible performance for massively-threaded, compute-intensive workloads that can be had in a client-side solution. For the right applications, it will deliver well beyond mainstream performance, and at a price tag well beyond mainstream price points. Specifically, as configured, this $24,999 ProMagix 360A (base price $4,499) will look to serve the top 0.1% of the workstation market, based on historical volume vs. price curves.
The volume vs. price curve for deskside workstations has been remarkably consistent over time |
The exterior
Where’s the front grill? With the airflow orientation of this machine, there’s no need for one |
Much more on the airflow choices ahead in a discussion on ergonomics and noise levels, but in terms of looks and layout, the choice to vent hot air out the top opened the door to giving the machine solid and clean lines viewed from the front. Even the ubiquitous few front ports (always a couple USB and audio I/O ports) were moved to the top of the box. I’m not sure that’s ideal, as it may invite dirt, dust or even a coffee spill, but honestly that’s more of a user preference. It does mean that sliding it under a desk requires a bit more attention to clearance, not just for those ports and cables attached to them but much more so to provide clearance for exhaust air.
The outside of the Velocity Micro ProMagix HD360A immediately catches the eye, from two perspectives: the aesthetics and layout. Now a workstation’s looks are both subjective in appeal and importance, but these eyes found this machine particularly striking, with its solid, brushed aluminum case of a color I can best describe as gun-metal (but not exactly). Taking a cue from high-end gaming oriented PCs, the ProMagix HD360A’s windowed side panel shows off the machine’s clean and potent internals. The solid front pleases as well, with the very notable absence of a conventional air-intake grill.
And that leads to the layout. Rather than a traditional airflow from front to back, this machine employs the Velocity Micro’s home-grown GX5 chassis (and previously employed in the company’s Raptor gaming PCs) to draw fresh air in from below and vents out the top and sides. Note the machine has a slight open pedestal that allows air intake from the bottom with a very sensible and easy-to-maintain dust filter snapped on.
The ProMagix HD360A’s underside (rear vent for the PSU to the left and primary front air intake with dust cover to the right) |
Much more on the airflow choices ahead in a discussion on ergonomics and noise levels, but in terms of looks and layout, the choice to vent hot air out the top opened the door to giving the machine solid and clean lines viewed from the front. Even the ubiquitous few front ports (always a couple USB and audio I/O ports) were moved to the top of the box. I’m not sure that’s ideal, as it may invite dirt, dust or even a coffee spill, but honestly that’s more of a user preference. It does mean that sliding it under a desk requires a bit more attention to clearance, not just for those ports and cables attached to them but much more so to provide clearance for exhaust air.
A missing feature common to rivals’ premium dual-socket (2S) workstations is an integrated handle to pick up and move these bulkier models. Whether that’s a negligible or considerable issue depends on the user and/or an IT administrator who may need to relocate the machine often.
On the top: not the usual spot for the ubiquitous front-side USB ports and analog audio mini-jacks |
The Interior
Velocity Micro touts the meticulous, by-hand assembly of components and wiring, and for good reason. Removing the windowed side panel (as simple process) reveals an immaculate interior, with minimal cabling visible, hand-trimmed and neatly tucked away. In addition, five toolless storage drive bays and all components appearing very accessible for service and swapping.
The immaculate interior of the 64-core, 128 GB ProMagix HD360A |
Given this product’s pricing puts it at the very top end of the market, it’s worth pointing out that the chassis does not support rear retention on the graphics add-in board (AIB), a feature to secure (in particular) bulkier, dual-slot AIBs like the Quadro GP100. It’s quite possible Velocity Micro consciously chose against the retention, as it would interfere with the GX5 chassis’ airflow for cooling (more on that ahead). I personally don't find that a significant issue with the machine in use in a fixed location as the PCIe bracket holds it quite securely. The rear of the board will flex modestly under pressure, and that's why Velocity Micro ships the machine with a styrofoam block inside supporting the AIB during shipment.
The Key Components: Motherboard, CPUs, AIBs, Memory and Storage
The maximum configuration of the ProMagix HD360A versus a popular Tier 1 OEM (e.g. Dell) Premium 2S workstation |
Velocity Micro clearly wasn’t going to create a dual Epyc workstation with 64 cores and skimp on supporting component types and SKUs. On the contrary, the company allows for custom component selection that — at its maximum level — meets or exceeds that available by just about any workstation on the market. Below is a rundown of the maximum capabilities for the ProMagix HD360A of the four top metrics — CPU, AIB, memory and storage — as compared to a top-tier supplier’s Premium 2S workstation model.
With two notable exceptions, the capabilities of the ProMagix HD360A meet or exceed those of the Dell Precision 7920 (which is also quite comparable to HP’s Z8 and Lenovo’s ThinkStation 920 Premium 2S workstations). One exception is AIB capacity. Full tower 2S workstations tend to all have the option of at least four, single-width, moderate-wattage (e.g. a Quadro P2000 or P4000) and at least two double-width, high-wattage (i.e. 235+ W) AIBs, like the Quadro P6000 or GP100. The Dell Precision 7920 can actually handle three of the latter, with the appropriate PSU chosen. The ProMagix HD360A has only two PCIe x16 slots for AIBs, both of which can be populated with high-wattage AIBs. The second exception is the presence of just one M.2 slot for PCIe NVMe SSD storage. Most deskside workstations, even mainstream models, tend to support at least two. Both these exceptions are based on the machine’s motherboard, the Super Micro H11DSi (https://www.supermicro.com/Aplus/motherboard/EPYC7000/H11DSi.cfm).
Why such limitations on a motherboard designed for the ultimate possible workstation configurations? Well, simply because that’s really not the platform it’s primarily designed for. While it can certainly serve as a dual-socket workstation motherboard, its intent is first and foremost to act as a dual-socket server motherboard. And servers don’t typically have a lot for high-performance graphics and still err toward massive amounts of drive bay storage versus smaller-footprint M.2. In fact, was it not for the emerging applications in GPU-compute driven HPC and server-hosted virtual workstations, the PCIe x16 slot count might be just one, or even zero (a server console could rely on the motherboard’s integrated GPU controller). The choice of the H11DSi is no criticism of either Velocity Micro or AMD. It’s just a result of the vast majority of interest in Epyc (especially 2S Epyc) focused in these early on servers.
The constraints are a bit ironic, however, given that Velocity Micro touts “an unprecedented 128 PCIe Gen3 lanes to meet large AIB and NVMe needs.” Yes, the dual EPYC CPU chips can drive more Gen3 lanes than any other platform currently, including dual Xeon Scalable designs. But those capabilities are handcuffed by the Super Micro motherboard, limiting AIB and NVMe capabilities below that of many entry class 1S workstations.
When it came to choosing an AIB for our evaluation machine, Velocity Micro picked the Quadro GP100, Nvidia’s top end on the price spectrum. But it’s worth pointing out that the GP100 is
The rear: base I/O, slots and PSU |
not the top-end of the Quadro line when it comes to visual performance. That would be the P6000, which is actually less expensive but delivers more FLOPS and (with the rare exception that a high-quality graphics animation may be bottlenecked by memory bandwidth) higher 3D graphics throughput, albeit a slightly higher power consumption And that’s a telling choice for this machine, which is much more likely to be employed for high-performance, massively-parallel computation — things like complex scientific or financial simulation and modeling that employs a huge burden in both single and double precision floating point math. Because where the GP100 might tradeoff a little relative to the P6000 in 3D visual performance, it makes up for in other areas, most notably in double-precision floating point math and high-bandwidth, low-latency memory via its HBM2 memory subsystem. That is, the GP100, while obviously quite capable in interactive 3D graphics, is tailored more for those who also need a high-level of compute performance out of their AIB.
I/O ports are plentiful enough, with four in the rear USB (two 3.0 and two 2.0) and two in front (2.0), along with aforementioned analog mic and line-in. Three LAN ports reside in the rear, two conventional RJ45 Gigabit Ethernet ports and one IPMI port (intended primarily for server applications). A PS/2 port is missing (yes, believe it or not, they still come standard on most workstations and valued by many for old enterprise applications) and it’s shy a couple of USB ports compared to the typical machine, but most should find the I/O offerings acceptable.
Delivering the power to drive all these components is an 80 Plus Gold efficient, 1000 W EVGA SuperNova (PSU), mounted on the rear floor of the chassis with its own internal fan. The PSU’s “Zero RPM” fan mode means no additional noise at lower loads. Even when active, it doesn’t seem the PSU’s fan is responsible for much of the system’s total decibel output (more on noise levels ahead). While this 1000 W PSU is not modular and tool-less, access appears quite simple, and Velocity offers an upgrade to 1300 W modular PSU, if that’s of importance (or required for power, as I’d imagine if a second 235+ W AIB were added).
The Performance
To gauge performance levels of this beast of a workstation from Velocity Micro, we employed the latest 2.1 version of SPEC's workstation-focused benchmark, SPECwpc. While no benchmark is perfect, SPECwpc does the best job I’ve seen of stressing all workstation components in a “whole system” environment that users actually experience. It’s both broad and deep, and aggregates sub-tests into workload groups representative of the most common workstation verticals: Media and Entertainment, Product Development, Life Sciences, Financial Services, Energy and General Operations. It even borrows the same viewsets that its graphics-focused sister test, Viewperf, uses to measure 3D graphics performance.
Apples to apples comparison? No, but that doesn’t mean it’s not a useful exercise.
The old adage of be careful not to compare apples and oranges is, on the surface, a sensible one. The last thing you want to do is judge one product inferior based on an unfair, straight-up comparison to another that had different very different goals and constraints in its design. But in the case of a machine like the ProMagix 360A one, it's not easy to find a very similar model to make a fair apples/apples comparison. In fact, no other workstation can hit the 64 core count, though one built around a dual Xeon Scalable Platinum with 28 cores would come close.
But the lack of a perfect match for the ProMagix HD360A shouldn’t stop some experimentation and analysis in comparison with other platforms. In fact, one can argue that a true apples-apples comparison would be boring anyway, as the same make-up of IHV-supplied internal components is most likely going to deliver virtually the same performance. So in a way, when it comes to reviewing workstations, comparing appropriate apples and oranges can actually yield the more insightful observations, particularly for potential buyers struggling with the pros and cons of different types of machines. And that’s the opportunity we have here, to compare the different workstation products whose designs evolved from different priorities and tradeoffs among the competing constraints that face all designers of high-performance workstations: size, materials cost, labor cost, power limits, and thermal dissipation limits, to name some of the most glaring.
As luck would have it, we had an ideal opportunity to test and assess this massively core’d 2S Epyc workstation in the context of two other machines — also benchmarked on the same SPECwpc 2.1 test suite — which represent different, yet valid takes on the what types of CPUs should be driving performance-oriented workstations. When the 16C Threadripper emerged on the workstation scene in the Boxx Apexx 4 6301 last fall, I had the opportunity to compare it to another recent Boxx release, the Apexx S3 with few fewer cores, the 6C “Coffee Lake” CPU, but whose cores were overclocked to nose-bleed levels.
Tossing into that mix the Velocity Micro ProMagix HD360A Epyc machine, we’ve got a unique opportunity to evaluate the relative merits of very different workstation CPUs. In one corner we’ve got a Core i7 with a mainstream number of cores (6, though 4 is even more mainstream) running at a bleeding-edge frequency, in the other we’ve got a 16C Threadripper running at mainstream frequency but populating nearly 3X the number of cores. And now we add the absolute max-possible core count available, but at the most modest frequency.
Configuration specifications for our ProMagix HD360A “apple” and two interesting “oranges” to compare |
The widely varying system specifications for our three test workstations (normalized to Boxx Apexx S3 with 6C Intel “Coffee Lake” Core i7 CPU |
Given the relative strengths and tradeoffs for the three workstations’ CPUs, I’d expected to see the 6C Coffee Lake machine perform better on single-to-few threaded workloads than the 2x32C Epyc machine perform the best on the higher thread count tests. And the 16C Threadripper would own the middle ground. The degree of those disparities would likely be the more interesting discovery in testing.
That spread in ability is absolutely OK, and doesn’t mean one is better than the other, but rather that the three CPUs are equipped to operate better on different types of workloads. It’s like comparing a diesel engine that delivers huge torque at low RPM with a sports car engine that drives RPM high in order to achieve its horsepower: two different tools optimized for different applications.
It’s also worth emphasizing here that the faster-core versus more-cores decision represents a true engineering tradeoff, all else equal. That means the more you have of one, the less you’ll tend to have of the other. Ultimately, chip thermal, power and electrical constraints will limit designers to how much they can push on one or the other design points. Populate a few cores and it’s far easier to drive the frequency up, but start piling on cores and (again, all else equal) the frequency will need to come down. It’s impossible, or at least very difficult and costly, to break that tradeoff and offer lots more cores at the same frequency.
A true tradeoff: cores vs. core count |
And that leads what I find the most illuminating exercise in this review: to compare the relative merits of a workstation based on the several-core, high-frequency Core i7 with the many-core, modest-frequency Threadripper and the massively-core’d 2S Epyc with the lowest frequency, 2.2 GHz. First up, let’s take a look at the single-thread SPECwpc test scores for the three systems, viewed when normalized to the 6C Coffee Lake Apexx S3 scores (i.e. the Apexx S3 is a “1” always). As expected, and to varying degree, the Apexx S3 reigns supreme, albeit to varying degrees.
Relative performance for minimally-threaded, compute-focused (i.e. no Viewperf viewsets) SPECwpc tests — normalized to Apexx S3 / 4.8 GHz 6C Coffee Lake |
Next up are the heavily-threaded SPECwpc tests, which take advantage of virtually all “logical” processing cores available in the underlying hardware. Logical cores include both the physical cores, as well as the same number of virtual cores enabled by technology like Intel HyperThreading that allow two threads to timeshare one physical core. Accordingly, the 6C Coffee Lake has up to 12 threads allocated, the 16C Threadripper up to 32 threads, and the 2S 32C Epyc up to 128 threads. And again, just as expected, the 16C Threadripper system handily outpaces the 6C Coffee Lake, while the 2 × 2C Epyc crushes both. Still, of course, the mileage from additional cores varies, though the 16C Threadripper averages 1.8X the performance of the 6C Coffee Lake, while the 2 × 32C Epyc averages around 3.3X the throughput.
The spread in performance at each thread count can likely be attributed to differences in microarchitecture, for example the aforementioned differences in cache size and memory access. And of course, even multi-threaded tests will scale with clock frequency as well. The outlier srmp test presents a bit of a quandary, as it’s heavily threaded and logical cores appear heavily utilized during the test, meaning they don’t appear memory-starved.
Relative performance for heavily-threaded SPECwpc tests (normalized to Apexx S3 / 4.8 GHz 6C Coffee Lake) |
Finally, one SPECwpc test in particular, Handbrake, is interesting in the sense that it presents a range of threading during its execution. The primary Handbrake code may be single-threaded, but portions of code it leverages — the codec(s) — can harness up to 32, with 16 concurrent threads predominant (and consuming capabilities of 16 cores). And sensibly, in the case of Handbrake, it’s the 16C Threadripper that reigns supreme, as Epyc’s more plentiful cores are not effectively exploited.
Relative performance for moderately-threaded Handbrake test (normalized to Apexx S3 / 4.8 GHz 6C Coffee Lake) |
Cooling and Noise
To achieve the ProMagix HD360A’s level of multi-threaded performance, engineers must make tradeoffs and users must accept some compromises. These Epyc parts chew up a lot of watts, 180 each for these particular parts, the physical reality of running 32 cores (and similar to Intel’s 28-core Xeon Scalable parts). Supplying the watts is the easier of the engineering issues to address, as managing the resulting thermal output from those parts is daunting. Keeping temperatures across the silicon inside consistent and within specifications requires a lot of cooling, and cooling typically involves a copious amount of well-thought-out airflow routing.
The trouble is, regardless of how well thought out, copious cfm (cubic feet per minute) means fans and turbulence, and that in turn invariably means some level of noise. Of course, it’s obvious and irrelevant that a machine like this will be noisier (particularly under load) than a mainstream quad-core tower workstation at modest frequencies. It will be, and higher decibels — to some degree — will be a compromise that buyers demanding this level of performance will understand. The questions are to what degree it’s louder and to what extent the user is willing to accept the incremental noise.
As addressed previously, with the GX5 chassis’ unconventional approach, air flows in actively via an intake fan on the bottom, then exhausts out the top actively via two vents and fans, and (according to Velocity Micro’s illustration) passively out the rear as well.
The dual Epyc 7601 CPUs, with heatsinks and adjacent to dual ceiling exhaust fans/ports |
While not delivering the level of multi-thread performance as the ProMagix HD360A, both the liquid-cooled Boxx Apexx S3 (with overclocked 4.8 GHz 6C Coffee Lake Core i7) and the more conventional air-cooled Boxx were no performance slouches either. And both produced more noise than a typical workstation. (For the record, liquid-cooling should really be redefined as “liquid-assisted” cooling, as airflow and fans are typically required to cool the liquid in the radiator at the enclosure’s vents, somehow, so fan noise and turbulence still impact, though to a lesser degree.)
The ProMagix HD360A and GX5 chassis’ airflow scheme |
Despite all the extra watts being consumed in the pursuit of overclocked performance, the Apexx S3 was not conspicuously noisy under nominal loading. Under heavy load (portions of SPECwpc), however, fan speeds kicked up significantly, driving perceptible noise up significantly, from a generally tolerable 47.2 db to around 50.7 db (approximating distance from under desk to ears). That’s a level I would find annoying, were it to last for lengthy periods. The Apexx 4 with Threadripper also got a little loud under heavy loading (same portions of SPECwpc), but on the order of 49.2 db peak, a little than the S3. That makes sense as the S3 has to push cfm probably on the order of the Apexx 4 but through a significantly smaller volume, leading to both higher fan speed and air turbulence.
However, assessing the air and fan induced noise of the ProMagix HD360A hit another level entirely. My reasonably-well-calibrated and standardized testing showed 52 db idle and 55 db under load, levels I personally would only be able to tolerate for lengthy periods with noise-cancelling headphones. I did not get a chance to assess a ProMagix HD360A with a liquid-cooling option, as it was not available at the time of review but may substantially drop noise levels.
The verdict: AMD’s Zen now capable of serving the ultra-high end, no-compromise target of 2S Epyc and the Velocity Micro ProMagix HD360A
If this is the machine for you, you don’t have to ask …
Users anywhere near the mainstream don’t have the combination of the workloads and urgency to take advantage of what the ProMagix HD360A can deliver. And they’re certainly not going to be interested in paying ultra-high end price points for it. And that’s fine, this machine is not meant for anywhere near the bulk of the marketplace, which today remains satisfied by mainstream quad-core CPUs (that account for about 75% of workstation shipments) and sub-$2,000 hardware configurations.
No, this machine is purpose built to serve the top 0.1% of the workstation market, the highest demand corner that struggles very specifically with massively-parallel, compute-intensive workloads. And view shortening execution times for those workloads is highly or critically desirable, perhaps directly improving the business’s bottom line. Those users know who they are — they don’t take a whole lot of convincing or a whole lot of education that they’re going to need specialized and pricy hardware to achieve their goals. They know there are very few products capable enough to address their demands, and fortunately, they're also aware of some of the compromises they’ll need to accept in order to get that type of performance.
A fully-loaded 2S Epyc Velocity Micro ProMagix HD360A sits firmly at the very top end of today’s workstation market |
Compromises to make to get the ProMagix HD360A’s unprecedented level of massively-threaded workstation performance
There’s no free lunch when it comes to anything, especially top-end workstation performance. What the ProMagix HD360A can deliver — the highest level of performance for massively-threaded, compute and memory intensive workloads — comes with some compromises. The obvious one is price. Even the base price of $4,499 is more than 125% higher than the average selling price of today’s workstations. With this configuration we’re talking a hefty $25K, as noted before in the highest level of the workstation’s high end.
But you can’t compare this machine with a machine anywhere near the average selling price, and if you look at it relatively speaking, you can make the argument the $25K is actually a pretty good deal. Configuring a closely comparable Dual Xeon Scalable Platinum 8180 Premium 2S workstation with the same AIB, memory size and storage size and type yielded a price substantially higher — roughly $8K more — than the ProMagix HD360A. (Note that dropping from 2.5 GHz of the 8180 to 2.1 GHz of the 8160 would shave off that premium, but then you’re left with fewer cores at a lower frequency than the Epyc 7601.) So yes, you’re going pay more
than 10X the average selling price of a workstation today. But you’re getting performance for massively-threaded compute-intensive workloads that the average machine can’t touch … and one close configuration available in other premium 2S workstations could run $8K more. So price is a major con if you’re not the perfect candidate for this machine, but if you are that candidate, you may actually see the price a solid pro for the ProMagix HD360A.
The motherboard not ideal, but probably not a significant issue to most
The fact that the Super Micro motherboard is sub-optimal for high-performance client-facing workstations isn’t ideal, but it’s far from a showstopper as well. It’s the old chicken-and-egg dilemma — AMD needs the volume to justify the board and the board to stimulate the volume. But the fact that that the ProMagix relies on a server-class motherboard is much less of an issue, when again considering the target workloads and usage models. In a way, this workstation is running server-class workloads on a client. So a thing like housing four AIBs probably isn’t a high priority for a user whose primary need is massive client-side computation. It definitely does make the claim of Epyc’s “unprecedented 128 PCIe Gen3 lanes to meet large AIB and NVMe needs” awkward, however, given that entry-level workstations built around workstation-focused motherboards can support far more slots for AIBs and M.2 SSD storage.
This user’s most significant compromise: noise
So yes, it’s perfectly acceptable — even required — to make accommodations and compromises to achieve the level of performance the dual EPYC ProMagix 360A can deliver on massively parallel, compute-intensive workloads. Price is an obvious one, but again, given the hardware specs of this machine and somewhat comparable alternatives, you could just as easily argue it’s a plus. A few shortcomings on the enclosure, chassis and the motherboard are really user and application dependent. Many (most?) won’t care about the lack of a handle, AIB retention, modular PSUs, external hot-swap drives, and the lesser capacity of AIBs. One versus two M.2 slots may be an issue limiting max PCIe NVMe storage (depending on needs and specific M.2 cards offered), but one can always add a conventional PCIe add-in card SSD.
Those are all minor tradeoffs in my opinion. But there’s another tradeoff that isn’t: noise. Right off the bat, let me emphasize I’m aware sensitivity to noise is a subjective one. Some need to work in complete quiet, while others can tolerate all kinds of racket and still focus and remain productive. I personally work with music and speakers on loud enough for my family to ask me to turn it down with my home office door shut. However, it’s worth emphasizing that noise levels in the age IHV-dominated workstation builds, noise levels are not only not a secondary issue for most, more often it’s a front-and-center decision criterion. In fact, I know of more than a few enterprise-class buyers whose first task in evaluating machines for purchase is to throw it in an acoustic chamber.
A great candidate for workstation remoting and liquid-cooling (option to come)
Personally, I could not work with a machine this loud under my desk without headphones, regardless of its performance level. Its 55 db at load and 52 at idle is beyond what I’ve measured, and well beyond what I could tolerate for anything but a short time. And that’s with music speakers on fairly loud in an attempt to drown it out.
Those who need this level of performance — and who would just the noise levels similarly — do have options: noise dampening or cancelling headphones, for one. But there are two other interesting option for a machine like this: one, upgrading to a liquid-cooling option that Velocity has in the works (but not available at the time of review), remoting the workstation. Liquid-cooling still requires air cooling at the perimeter (to transfer heat from liquid in the radiator to the exterior), but airflow can be dramatically reduced and should result in substantially lower noise levels. Another option could be remoting, placing the tower in a separate (and appropriately cooled/ventilated) room or as a 4U rackmount in the datacenter. That would effectively move the noise elsewhere and let the user harness the machine’s unprecedented level of 64 core performance. Velocity Micro doesn’t make a 4U enclosure as a standard offering but would build to suit.
A Fact of Life to Get This Level of Multi-thread Performance … or not
The discussion of cooling, airflow and noise all leads to a salient question: would every workstation tower with components inside of this level of power consumption going to hit that noise level if relying on air cooling alone? Or would another chassis design and airflow scheme drop the decibels below my subjective threshold? Those are more difficult questions to answer.
Say I were to load up a Dell Precision 7920 (or HP Z8 or ThinkStation P920) with two of those 28C Xeon Scalable parts (Platinum 8180), each with TDPs of 205 W versus the Epyc 7601’s 180 W (and otherwise similarly configured). Would I end up with decibel readings as high or higher? That I don’t know, but the ProMagix HD360A’s airflow scheme, particularly driving a lot of cfm out two relatively small top grated, openings, has me wondering if a more conventional but custom-engineered cooling scheme might perform better. I hope to ascertain more with an opportunity to review another dual-socket 56+ core, 128+ GB tower workstation sometime soon.
In the meantime, for those few of you struggling with lengthy execution times for highly threaded, compute intensive applications, you owe it to yourself to check out the ProMagix HD360A. Evaluate the degree of both the performance headroom this machine offers, as well as the noise levels, and I’d certainly encourage checking out their soon-to-be-released liquid-cooling option to judge how much that can dampen the decibels. The importance of both is subjective, and the only one whose opinion ultimately matters is the buyer himself.