Ryzen Threadripper 1950X vs Core i9-7960X | Hacker News Reader

Ryzen Threadripper 1950X vs Core i9-7960X

pcworld.com

49 pointsrbanffy25 comments

That pricing comment is almost laughable and makes me wonder how much Intel paid for this review.

Framing the pricing in the context of a $9,000 PC is a blatant attempt to dilute the price discrepancy. In the context of a common ~$2,500-$3,000 rig, $700 matters a lot. I'd be curious to see how many people buy these "dream" $9,000 PCs. I'd wager it's not many.

I don't see this as Intel biased, in the "The bang for buck stops here" the author very clearly points out that the AMD CPU is vastly more cost effective for likely scenarios.

Though saying it's a Mobo and a decent SSD somewhat understates what those builders would likely spend $700 on.

> I don't see this as Intel biased,

The article was written to sell the idea that Intel's offering beats AMD's in all relevant features considering all aspects deemed relevant by the writer, and then he proceeds to bend himself backwards to try to come up with any excuse to ignore the 700$ price difference between Intel's i9 rig and AMD's Ryzen Threadripper and how that price difference buys only marginal performance gains.

I would also add that it's also questionable how the author decides for no reason at all to setup Intel's rig with two SSDs in RAID0 while leaving AMD's rig with two separate drives.

All in all, it looks like a far-fetched and somewhat desperate attempt at finding a particular offering by Intel that can show some performance gains over the AMD Ryzen Threadripper, and in the process commit so many methodological errors to get to those marginal gains that a good portion of the text is dedicated to come up with excuses to try to sell the idea that none of that matters.

Sort of offtopic, but what is with those charts? Is there some rhyme or reason to the ordering that I was missing? Sometimes the AMD was above the Intel and sometimes vice versa. Sometimes the better result was above the worse and sometimes vice versa. I had to squint at the "longer/shorter bars indicate higher performance" captions on every chart to make heads or tails of it...

Understandably different benchmarks have different measurements of success (scores vs. time spent in seconds, etc) but I agree this could be displayed better some way.

Perhaps simply outlining/indicating by colour or a visual marker for the better performer on the chart.

You can always normalize it to a "the longer bar is better" measurement. For example instead of "Kernel Compile Time" you graph "Kernel Compiles / hour".

Yes, the charts were useless. In several cases the chart results contradicted the commentary. They should highlighted the better performer.

Something interesting a lot of reviews don't consider is that Threadripper, since it's a multi-chip module with each chip having its own memory controller, has a non-unified memory architecture (NUMA), where half the memory will have higher latency because it has to go to the other chip. Some software is written with this in mind, some isn't. Consequently, Threadripper has two modes: NUMA mode (OS sees two NUMA zones and different latencies) and UMA mode (single NUMA zone, single average latency). It may be worth trying both when benchmarking to see what difference it makes. I think most reviewers aside from AnandTech have overlooked this since NUMA hasn't been a concern on the desktop before.

> has a non-unified memory architecture (NUMA)

Small, pedantic, point of fact, NUMA stands for Non-Uniform Memory Access. You could say Non-Uniform Memory Architecture in this context, obviously, as well.

…oh, huh, okay. Point taken.

Definitely agree there is 0 attention paid directly to memory bandwidth tests. Also, most of these are floating point tests as well. As someone who does a lot of systems work on a workstation, these figures are just as important if not more important.

I hate to be the one to say it but this looks very one-sided to me:

is 1-20% of a performance benefit really a "It gives you great performance at light-duty applications and generally can't be touched by the Threadripper 1950X in heavy-duty applications, either." ?

tl;dr As you'd expect, AMD's chip is a no-brainer on bang/buck.

I wish more reviewers would compare chips at various price levels, because I think that's how most people think about their purchase.

If you want to build a $1500 gaming rig, you're going to look at the whole system cost, and you'll designate a certain price for the CPU. You'll want to buy "the best CPU for that money".

Similarly, if a medium development company has $10,000 to upgrade some of its systems, does it buy 6 Intel CPUs or 10 Ryzen CPUs with that money?

Comparing product brands like "Intel's top of the line" vs "AMD's top of the line" is not very useful, especially if there's a 2x delta price difference between them, while the performance delta is nowhere that big.

This should be even more the case for data center purchases, where total cost of ownership is king.

For data centers the cost is dominated by electricity, not CPU price, so they look for performance per Watt. This was not addressed in the article.

> so they look for performance per Watt. This was not addressed in the article.

The did test power consumption though: https://images.idgesg.net/images/article/2017/09/fnw_showdow...

That plot shows the power consumption during one rather synthetic benchmark. What is necessary for meaningful conclusion about price/watt is to have power consumption during a particular benchmark so the results can be scaled. This is not addressed.

They actually did address it tangentially in the article. The highest the AMD CPU used wattage wise was under 275W. They showed Intel using over 50W more at peak and claimed they’d heard of it pulling over 500W (total, and in a scenario that almost no DC would likely do). So yes, AMD has a significant advantage OpEx and CapEx wise.

to be fair though, these are not server processors, they belong in the office not in a data center.

? It's sitting there one page 2.

Threadripper is segmented as a workstation processor - Epyc is the server line.

Real-world energy efficiency is extremely complex, because it's hugely dependent on workload. If you're the kind of customer that is highly sensitive to efficiency, you really need to perform your own evaluation.

https://www.cpubenchmark.net/cpu_value_available.html

It's only the Passmark benchmark, and there's some noise from older CPUs, but it's still a useful tool.

Tom's also does a decent job of including price as a first-class metric.

Absolutely agreed though. I skip over any comparison review site that doesn't index for price: my goal for computer personal hardware is always to get the most (whatever) per dollar, never to get the best regardless of cost.

The actual TL;DR is that the Core i9 beat the Threadripper ist almost every test.

The bang-for-the-buck argument is only briefly mentioned towards the end, and is followed up by:

One last thing to consider is that the price of just the CPU isn't everything. When it's watered down by the total cost of the system, it may indeed be worth it for some. For example, the Falcon Northwest Talon systems that we used were both dream configurations with $2,400 in GPUs, maybe $1,800 in RAM and $1,200 in SSDs. That doesn't count the motherboard, hard drive, cooling, case, power supply or custom hand-paint jobs. When you're building a box in this range, the price difference in the CPU isn't that much.

Not by a margin that warrants the price difference. Tbh most are negligible and power consumption is worse as well.

I'll take both as TL;DRs.

I'm sort of unsure if that argument makes any sense.

Say you're at a restaurant, and they're selling a salad at $600. It's a really good salad, but not $600-salad.

But hey, you're at the restaurant with the realtor you're buying a house from. Combined, the salad and house are $865,300. Would that change your choice of meal?

More realistically, people often make the mistake to see money in relative terms where it should be treated in absolute terms. Say you're buing a nice ballpoint pen for $20. On the way out the store, you see a sign that the pen was actually on offer for just $10. Chances are, you'll invest the two minutes to get the rebate.

Now consider the same situation, but you're buying an iPhone. Again, you see a sign advertising the iPhone for just $675, not the $685 you paid. Most people are less likely to invest the two minutes here, even though the payoffs are equal.

Reply on news.ycombinator.com