Yes, basically all AMD motherboards that take Zen/Zen+/Zen2 architecture cpus support and implement ECC. This includes many of the original Zen mobos which did not have it enabled in the BIOS but subsequently made the feature available after a BIOS update.

The main trade-off is that virtually no ECC sticks are certified for overclocking out of the box, meaning that you have to overclock them yourself if you want to match non-ECC DDR4 that you can buy OCd out of the box. This isn't difficult to do, but is not everyone's cup of tea. And, generally speaking, most people wouldn't notice the improvement in performance anyway.

All Ryzen and Threadripper systems take unbuffered DDR4 (ECC or non-ECC). All EPYC systems take registered DDR4 (ECC or non-ECC). Not sure if you can stick unbuffered sticks into an EPYC but you definitely cannot stick registered sticks into a ryzen or TR mobo.

-Matt

You will wish you had used dark mode your whole life, instead of starting just last week :-). Truthfully, though, white backgrounds push a lot of light directly into your eyes and face, and having sat in front of CRTs ever since I was around 12 years old, I learned fairly quickly that I could pull much longer stints in front of displays that weren't blaring bright white pixels at me (well, bright green pixels way back then) 24x7.

I thank that insight now that I'm over 50.

My personal preference... a dark (but not black) background with modest (but not excessive) contrast and minimal bleeding:

xterm*background: #100010000000
xterm*foreground: #7FFFDFFFDFFF

-Matt

I take it you have never lived in the mountains ? During certain parts of the year on clear nights the day-v-night temperature delta can be as much as 80F. A 50F delta is more typical, though. But its still a lot.

-Matt

The duck curve was a worry a decade ago, but only because researchers thought there might be problems stabilizing the grid without base load. It turned out to be a non-problem... the (for example) CA grid is actually more stable now with less base load than at any time in the past.

The reason is that DC inverters used with solar and battery systems can react to changes in voltage and frequency on the grid in mere milliseconds, even microseconds, whereas traditional base-load sources actually take on the order of hours and even NG peaking plants take 30 minutes if they are cold and 5 minutes if they are hot.

The problem now is how to deal with the situation where renewables generation is able to take 100% of the load at certain times during the day. In California, this happens in the spring and fall (lots of sun, very little air conditioning needed). At the moment CA is forced to load-shed the solar a little during these periods in order to allow less-agile generation sources to continue operating.

-Matt

It actually doesn't, because the ambient temperature of the air will keep it 'warm' relative to the near absolute zero of space that it is radiating into.

But again, we are not talking very much power generation here. Only a few watts per meter squared at night, at best.

-Matt

At best it can take away a few watts per meter squared and during the day you have roughly 1000W/m^2 hitting the panel. So as a cooling mechanism for the panel in the daytime its a big fail.

-Matt

The only way to do this is to radiate at frequencies that punch through the atmosphere, which is precisely what these radiative panels do. With careful selection of the frequency, the panel is basically 'seeing' the near absolute zero temperature of space and can thus radiate into it.

However, the efficiency of this mechanism is quite low. We're talking, at best, a few watts per meter squared (verses one to two orders of magnitude more energy when operating normally as a solar panel during the day).

So in terms of power generation at night, not so much. But it isn't a total loss. The radiative mechanism works 24x7 so on a 24x7 basis it can add around 10-15% more energy production to normal operation. And at night even though the power generation is very low, its high enough to provide voltage and frequency services to the grid.

-Matt

I don't know the answer to that (which compiler they are using), but it almost doesn't matter. Modern CPUs do such a good job optimizing the instruction stream that is handed to them that vendor-specific micro-optimizations don't really do a whole lot these days. Over the years we have removed most of them because they just don't do anything any more.

I can think of only one micro-optimization that is vendor-specific, and that's Intel's optimization of stosb and movsb which is designed to support generic memcpy() and memset() operations. But it won't matter on a machine with this little memory.

All other optimizations are relatively agnostic and apply to both CPU vendors. Just common sense, really. Doing things like avoiding certain pushq/popq sequences (which has to manipulate %rsp) in favor of pre-adjusting %rsp and doing normal movq's, aligning code and certain branch targets, collapsing call/ret sequences, linker-based inlining, and so forth.

-Matt

Virtually guaranteed to be the case. You could probably make any linux distro match it just with a little cache-tuning for low-memory. Lots of bits and pieces of the Phoronix test suite are going to give skewed results based on available ram due to filesystem caching and other effects.

-Matt

32GB EUDIMMs are now widely available in retail channels (last year and in prior years they were mostly only available in commercial channels). You can buy them on Amazon.

64GB and 128GB EUDIMMs are only available in commercial channels insofar as I can tell. And expensive as hell... and have to be run at 2133 due to the line load. Not really the best fit for a threadripper, honestly. If you need that sort of capacity you would be buying an EPYC and not a TR.

-Matt

There are actually not very many workloads that will blow out the available ram bandwidth on a Zen 2 threadripper. The huge CPU caches make up for a lot. The 3990X will have 256MB of L3 cache and since there is only one NUMA zone it can distribute the memory load evenly across all 8 slots.

Given the performance we see with its smaller brother (the AM4 consumer 3950X with 16-cores and 64MB of L3), it is clearly not going to have any problems for most workloads.

-Matt

Ram capacity for the retail AMD chips like the 3900X or 3950X (really any Zen 2 AM4 chip), and both Zen+ and Zen 2 threadripper chips, is limited by the fact that there is only support for unbuffered DIMMs.

People trying to compare Xeons against TRs based on wanting to stuff a terrabyte or more of ram wouldn't be buying a TR for that, they would be buying EPYCs, so its a stupid comparison. Its also stupid because Intel charges thousands of dollars more for Xeons with large physical address spaces (even with recent price cuts, Intel gouges buyers just for wanting more addressable ram even if they don't need the cores).

I have personally stuffed 128GB into a 3900X (AM4 socket) using 4 x 32GB ECC UDIMMs, and 256GB into a 2990WX (threadripper socket) using 8x of the same type of memory. I can run the memory at up to around 2666.

Insofar as I know, one can use 64GB DIMMs in both situations (256GB on AM4 and 512GB on TR), and I think 128GB DIMMs can be used on the TR. But since they are unbuffered, they would have to at low frequencies (1066 MHz for a 2133 MHz data rate). But the biggest DIMMs I personally own are 32GB each so I can't test higher capacities.

AM4 and TR Motherboard vendors do not generally validate for high-capacity memory, which is why they list lower capacities, but I they all support high-capacity memory just fine.

Very few people need that much memory even on a threadripper. We need it for bulk compiles... around 2GB per cpu thread, so we need around 128GB of ram with a 32-core/64-thread threadripper and 256GB of ram with a 64-core/128-thread threadripper (the 3990X releases on February 7th). Most other (likely) workloads do not need that amount of memory though, particularly when one can get NVMe storage devices with 5GByte/sec bandwidths.

The EPYC chips support 2TB per cpu socket (4TB total for dual-socket EPYCs), using registered DIMMs.

--

The bigger deal with the threadrippers is the massive PCIe bandwidth. Not only do you get 128 PCIe lanes (actually more when you include the chipset), but the Zen 2 I/O hub built into the cpu chip has over 400 GBytes/sec of peer-to-peer bandwidth. Intel chips clock in at more around 100 GBytes/sec (or less).

DRAM bandwidth is roughly the same for both vendors, but AMD cleans Intel's clock out on peer-to-peer PCIe bandwidth and this is quickly going to become important in the commercial space.

-Matt

It doesn't matter, really. Anything with content. So e.g. youtube, google news, a newspaper or two... stuff like that. Nothing fancy.

-Matt

The 'laptop crap' is a much bigger market than desktop APU. That said, I'm sure AMD will release a 4-series APU this year some time.

AMD is constrained by TSMC's wafer production volumes. AMD couldn't release all the chips at the same time even if they wanted to, TSMC just doesn't have the capacity to handle the rush.

Remember how long it took for the 3800X and 3900X to settle down to their MSRPs? It took at least 4 months, even longer. And the 3950X is only just now becoming widely available at +$40 over its MSRP.

AMD has to wait a few months for current demand spikes to ease before they can release an APU, so look for summer. TSMC is expanding its capacity and AMD will be able to take a chunk of Apple's TSMC capacity at 7nm as Apple moves to 5nm, but that's the basic constraint.

-Matt

Web clients and servers negotiate the format. Most CDNs can distribute video in numerous formats and resolutions (some even translate on the fly so they don't have to store them). Anything with hardware video acceleration in the last 10 years is likely to be supported.

-Matt