News AMD Explains Why 110-Degree Operating Temps Are 'in Spec' for RX 5700

[Admin]

AMD explained why junction temperatures above 100°C are considered "in spec" for the Radeon RX 5700 and Radeon RX 5700 XT that debuted in July.

AMD Explains Why 110-Degree Operating Temps Are 'in Spec' for RX 5700 : Read more

 

[justin.m.beauvais]

I figured it was related to the way they manage thermals on Radeon VII. I saw a bunch of launch day stuff where people were SUPER concerned about the thermals on Navi, and I was shaking my head. Obviously they had forgotten about VII. I'm pretty surprised that it took AMD this long to clear this up.

 

[Giroro]

So a temperature scale based on boiling water makes more sense to measure the weather than a more finely graduated scale that is more closely aligned to human living conditions.... and also measuring multiple parts of meat makes it cook faster.... And also, AMD really wants to ruin the fun of overclocking by doing it better out of the box.

Got it.

 

[salgado18]

So a temperature scale based on boiling water makes more sense to measure the weather than a more finely graduated scale that is more closely aligned to human living conditions....

But it's not being used to measure weather, it's being used to measure the temperature of a chip (which could interact with water btw). Also the majority of the world (especially the scientific community) uses it.

and also measuring multiple parts of meat makes it cook faster....

Not faster, better. You can let it cook until one spot hits critical temperature, and stop before burning. It's written right there.

And also, AMD really wants to ruin the fun of overclocking by doing it better out of the box.

So they should hold back performance so that only geeks can get the most out of it? It's like those smart suspensions and brakes that balance everything in ways a human can't with a pedal: it's not fun, but very efficient.

 

[TheSecondPower]

To be fair the Fahrenheit scale does have a couple strong advantages.

Interesting that they've designed the 5700 to intentionally reach these temperatures. I hope it still comes with a long life span.

 

[Loadedaxe]

I don't want something on my desk that heats up to 230°.

Drop the thermals and increase performance AMD.

 

[closs.sebastien]

amd cards... for ever cooking eggs... and electrical power-eaters = wasting energy for nothing...
oh, if your heating is broken at home, just go buy a radeon...

 

[jimmysmitty]

So they should hold back performance so that only geeks can get the most out of it? It's like those smart suspensions and brakes that balance everything in ways a human can't with a pedal: it's not fun, but very efficient.

It takes the fun out of being an enthusiast, which was and is some of their biggest supporters.

That said to me it shows they are hitting a wall and wont have room to "improve" existing products much. I don't think AMD will have a fully competitive GPU until they finally ditch anything GCN related.

 

[TJ Hooker]

I don't want something on my desk that heats up to 230°.

Drop the thermals and increase performance AMD.

If it's designed to operate at that temperature, who cares? It's not going to get nearly that hot on any external surface, so there's no danger from it if that's what you're worried about.

The GPU may not even be getting any hotter than previous generations. It could just be that the the temperature sensors are closer to the GPU hot spots than they were in the past, so they're better able to measure the true hottest temperature within the chip.

 

[lxtbell2]

(That's 230 degrees Fahrenheit for those of us forced to live with the clearly inferior temperature scale.) "

WOW. Another stupid millennial who hasn't a clue of anything. Inferior? How? The fact that Fahrenheit is based on the point where sea water freezes is some how Inferior? Just an education and leave your stupidity for your articles. Oh, you did.

Wow not sure who needs more education. Heat your sea water ice cube to 1F or even 10F and see if it melts. Now look up again what Fahrenheit used to define 0F. (Hint: it's a salt solution but nothing like sea water)
Anyone who received formal education in physics knows which one to use.

 

[vaughn2k]

Basic die (which is all semiconductor- majority of them) can withstand HTOL 125degC for 1000hrs, non stop.
Even Automotive has 150degC, same 1000hrs HTOL.
Don't bother.

 

[cryoburner]

The GPU may not even be getting any hotter than previous generations. It could just be that the the temperature sensors are closer to the GPU hot spots than they were in the past, so they're better able to measure the true hottest temperature within the chip.

Exactly, just because there are sensors reporting higher temperatures doesn't necessarily mean the chip is running hotter. Since there are significantly more sensors, it's more likely that there will be a sensor located at the hottest location within the chip, whereas in a GPU with fewer sensors, the reported temperature might not necessarily be from the hottest part, and more guesswork may be involved with knowing how close parts of the chip are actually getting to their limits.

That being said, blower-style coolers like those used on the reference 5700 XT are not what most would consider ideal, but that doesn't really matter much unless one needed one of these cards within their first month or so of release. Most people buy partner cards anyway, which typically have better coolers, and we're starting to see some of those cards now. The Sapphire 5700 XT Pulse that Tom's reviewed the other day kept its temperatures in the 70s during their tests, similar to Nvidia's competing options, while remaining quieter than the reference card.

amd cards... for ever cooking eggs... and electrical power-eaters = wasting energy for nothing...
oh, if your heating is broken at home, just go buy a radeon...

Actually, these RX 5000-series cards offer similar efficiency to Nvidia's current offerings, drawing a similar amount of power and outputting a similar amount of heat for a given level of performance, at least at reference clocks. Core temperatures do not directly correlate to heat output, and again, it's the blower-style reference cooler that's resulting in these seemingly high core temperatures. Furthermore, the entire point of this type of cooler is that it pushes heat directly out of the case rather than dumping it inside, so while the core temperatures might be higher, case temperatures will be lower.

I suspect Nvidia may once again pull ahead with efficiency when they move to a smaller process node for their next generation of cards, but seeing as they just launched their SUPER refresh, that probably won't be for about another year or so.

That said to me it shows they are hitting a wall and wont have room to "improve" existing products much. I don't think AMD will have a fully competitive GPU until they finally ditch anything GCN related.

I certainly don't see any indication that AMD's graphics cards have "hit a wall", and they apparently have higher-end cards in the works, though they haven't announced when those might be out, and I wouldn't expect them for a number of months. But even the 5700 and 5700 XT are already quite competitive, offering better performance in most current games than similarly-priced models from Nvidia. And even $400 cards are getting into kind of a niche market, with the vast majority of people buying less expensive hardware than that.

My guess is that they will probably fill in their lineup at the more popular price points over the coming months before bringing in anything higher end, which might not be until around the new year. As for where they could go with the higher-end, I could see them taking things in a more Ryzen-like direction, putting multiple modest-sized chips on a card, and having them connected at a low-level to avoid scaling or compatibility issues that multi-GPU setups might normally bring.

 

[jimmysmitty]

Exactly, just because there are sensors reporting higher temperatures doesn't necessarily mean the chip is running hotter. Since there are significantly more sensors, it's more likely that there will be a sensor located at the hottest location within the chip, whereas in a GPU with fewer sensors, the reported temperature might not necessarily be from the hottest part, and more guesswork may be involved with knowing how close parts of the chip are actually getting to their limits.

That being said, blower-style coolers like those used on the reference 5700 XT are not what most would consider ideal, but that doesn't really matter much unless one needed one of these cards within their first month or so of release. Most people buy partner cards anyway, which typically have better coolers, and we're starting to see some of those cards now. The Sapphire 5700 XT Pulse that Tom's reviewed the other day kept its temperatures in the 70s during their tests, similar to Nvidia's competing options, while remaining quieter than the reference card.


Actually, these RX 5000-series cards offer similar efficiency to Nvidia's current offerings, drawing a similar amount of power and outputting a similar amount of heat for a given level of performance, at least at reference clocks. Core temperatures do not directly correlate to heat output, and again, it's the blower-style reference cooler that's resulting in these seemingly high core temperatures. Furthermore, the entire point of this type of cooler is that it pushes heat directly out of the case rather than dumping it inside, so while the core temperatures might be higher, case temperatures will be lower.

I suspect Nvidia may once again pull ahead with efficiency when they move to a smaller process node for their next generation of cards, but seeing as they just launched their SUPER refresh, that probably won't be for about another year or so.


I certainly don't see any indication that AMD's graphics cards have "hit a wall", and they apparently have higher-end cards in the works, though they haven't announced when those might be out, and I wouldn't expect them for a number of months. But even the 5700 and 5700 XT are already quite competitive, offering better performance in most current games than similarly-priced models from Nvidia. And even $400 cards are getting into kind of a niche market, with the vast majority of people buying less expensive hardware than that.

My guess is that they will probably fill in their lineup at the more popular price points over the coming months before bringing in anything higher end, which might not be until around the new year. As for where they could go with the higher-end, I could see them taking things in a more Ryzen-like direction, putting multiple modest-sized chips on a card, and having them connected at a low-level to avoid scaling or compatibility issues that multi-GPU setups might normally bring.

Their power is slightly above and their temps even with the better cooler is above the stock RTX cooler.

As for the wall, if AMD is pushing these out with little to no OCing headroom its because they have a wall with the uArch. Much like Ryzen 3000 where the CPUs are pretty much performing as good as they can and can't clock much higher, if at all, than their stock configurations.

While Navi is a newer design it still has a lot of GCN in it and GCN, while good when it first launched, is showing its age. They really need a newer uArch that can overcome some of the limitations of GCN.

 

[Giroro]

But it's not being used to measure weather, it's being used to measure the temperature of a chip (which could interact with water btw). Also the majority of the world (especially the scientific community) uses it.

Most people look at the weather a lot more often than they look at any other temperature. That's why phones are preloaded with a weather widget on the homescreen instead of a CPU monitor.
If you can represent that temperature range with no decimals (meaning fewer written characters) and higher precision, then I think that is a tangible benefit... For example 72 F uses 2 characters to represent 2 significant figures, but 22.2 C uses 4 characters to represent the 3 sig figs needed to represent (almost) the same 'room temperature'.

Or, another way of saying that is that since 1 degree C is roughly 2 degrees F, then Celsius is about half as precise (or less if there's a decimal), for a fixed number of written characters. That doesn't matter very much if you're writing out analog measurements on paper or working with high-precision scientific instrumentation.. but it can be significant when you only have a display width of 2-3 characters on a common digital thermometer.

Besides, people don't even measure the temperature when they're boiling and freezing water since they can just look at it, so even Celsius is using arbitrary and not particularly useful stopping points. At least with Fahrenheit has the fringe usefulness where less than about 0 or greater than 100 is a temperature where unprotected exposure has a chance of killing you.

So what I'm really saying here is both systems are arbitrary and a little bit stupid - and we never would have had this divide in the first place if Celsius would have seen the benefit of deciding that water froze at 0 but boiled at 200.

 

[Urzu1000]

The concept is pretty simple and the explanation clear. I think some people here skimmed the article a little too quickly. This is a good thing and has nothing to do with how "hot or cool" these cards run compared to X, Y, or Z. That's a totally different conversation topic.

Presumably other cards would be able to operate at those temps if they had the same multi-sensor configuration, but with a single sensor, they have to be conservative to avoid damaging the card. This is a forward thinking move, but I'm curious why something so obvious hasn't been done before.

 

[TerryLaze]

The concept is pretty simple and the explanation clear. I think some people here skimmed the article a little too quickly. This is a good thing and has nothing to do with how "hot or cool" these cards run compared to X, Y, or Z. That's a totally different conversation topic.

Presumably other cards would be able to operate at those temps if they had the same multi-sensor configuration, but with a single sensor, they have to be conservative to avoid damaging the card. This is a forward thinking move, but I'm curious why something so obvious hasn't been done before.

It's not about the GPU chip or any of the components,a huge problem with all past cards that had high temps is the solder melting breaking the card and forcing you to reflow it in hopes to fix it.

 

[lxtbell2]

Most people look at the weather a lot more often than they look at any other temperature. That's why phones are preloaded with a weather widget on the homescreen instead of a CPU monitor.
If you can represent that temperature range with no decimals (meaning fewer written characters) and higher precision, then I think that is a tangible benefit... For example 72 F uses 2 characters to represent 2 significant figures, but 22.2 C uses 4 characters to represent the 3 sig figs needed to represent (almost) the same 'room temperature'.

Or, another way of saying that is that since 1 degree C is roughly 2 degrees F, then Celsius is about half as precise (or less if there's a decimal), for a fixed number of written characters. That doesn't matter very much if you're writing out analog measurements on paper or working with high-precision scientific instrumentation.. but it can be significant when you only have a display width of 2-3 characters on a common digital thermometer.

Besides, people don't even measure the temperature when they're boiling and freezing water since they can just look at it, so even Celsius is using arbitrary and not particularly useful stopping points. At least with Fahrenheit has the fringe usefulness where less than about 0 or greater than 100 is a temperature where unprotected exposure has a chance of killing you.

So what I'm really saying here is both systems are arbitrary and a little bit stupid - and we never would have had this divide in the first place if Celsius would have seen the benefit of deciding that water froze at 0 but boiled at 200.

If you're talking about precision at fixed character width - Fahrenheit has twice so below 100F, but 1/5 so between 100F and 100C. And that's most things I call "hot", computer components included.

 

[Pence128]

(That's 230 degrees Fahrenheit for those of us forced to live with the clearly inferior temperature scale.) "

WOW. Another stupid millennial who hasn't a clue of anything. Inferior? How? The fact that Fahrenheit is based on the point where sea water freezes is some how Inferior? Just an education and leave your stupidity for your articles. Oh, you did.

Farenheit is defined such that freezing water is 32°F and boiling water is 212°F. Seawater freezes at around 28.4°F.

 

[Olle P]

... to me it shows they are hitting a wall and wont have room to "improve" existing products much. I don't think AMD will have a fully competitive GPU until they finally ditch anything GCN related.

I think you draw the wrong conclusion.
The existing products are running at their limit, yes, but that doesn't imply that there can't be another more refined product with more compute units and faster RAM available relatively soon.

WOW. Another stupid millennial who hasn't a clue of anything. Inferior? How? The fact that Fahrenheit is based on the point where sea water freezes is some how Inferior? ...

That's not what makes it inferior, the scaling is!

Kelvin is the reference standard, starting from absolute zero.
To convert from centigrade to kelvin just add 273.15, which is very simple.
To convert from farenheit to kelvin you must first add 459.67, then multiply by 5 and divide by 9. Not that simple to do on the fly...

 

[Giroro]

If you're talking about precision at fixed character width - Fahrenheit has twice so below 100F, but 1/5 so between 100F and 100C. And that's most things I call "hot", computer components included.

Yes, that is what I'm talking about. The range of temperatures in which humans and agriculture can survive, that is where precision matters to most people.
I don't think Fahrenheit would still be around if Celsius were more precise in that range.

 

[King_V]

I don't want something on my desk that heats up to 230°.

Hope you're not using a current Intel higher end CPU then.

amd cards... for ever cooking eggs... and electrical power-eaters = wasting energy for nothing...
oh, if your heating is broken at home, just go buy a radeon...

Fun fact regarding the Vega 56 reference card. From the original Vega 56 review:

Using the secondary BIOS with a power limit reduced by 25% gets us 159.4W and 32.7 FPS. Compared to the stock settings, just 71.6% of the power consumption serves up 89% of the gaming performance.

The GPU hierarchy chart says that the GTX 1070 has an overall score of 69.9, and a TDP of 150.
The Vega 56's overall score is 76.7,

Reduce that by 8.9% (6.8), and its overall score comes to . . 69.9 - and at a TDP of 159.4W.

So, is it really wasting energy?


Now, they pushed it in order to outdo the 1070, but, the increase in energy consumption is disproportionate to the gains in performance. That's not really unusual.


Further - from the 5700/5700XT review:

If we take each card’s average frame rate across our benchmark suite and divide by the typical gaming power consumption we measured, the Radeon RX 5700 XT provides 0.43 FPS/watt. Radeon RX 5700 achieves 0.47 FPS/watt, as does the GeForce RTX 2060 Super. Nvidia’s vanilla 2060 is only slightly more efficient at 0.48 FPS/watt. So relative to the competition, AMD’s Radeon RX 5700-series fares well in efficiency.

The 5700XT is minimally faster than the 2070 Super. Rated at 95.8 vs the 2070 Super's 94.1. So, performs about 1.8% faster.
The 5700XT is rated at 225W vs the 2070 Super's 215W. That's 4.7% more power draw for that slight extra performance.

Less efficient? Sure. By much? No.


Care to make those ridiculous "space heater" and "inefficient" statements again?

 

[TJ Hooker]

That's not what makes it inferior, the scaling is!

Kelvin is the reference standard, starting from absolute zero.
To convert from centigrade to kelvin just add 273.15, which is very simple.
To convert from farenheit to kelvin you must first add 459.67, then multiply by 5 and divide by 9. Not that simple to do on the fly...

Well, Fahrenheit also has scale that starts at absolute zero: Rankine. Rankine is to Fahrenheit what Kelvin is to Celsius.

 

[Pence128]

Yes, that is what I'm talking about. The range of temperatures in which humans and agriculture can survive, that is where precision matters to most people.
I don't think Fahrenheit would still be around if Celsius were more precise in that range.

You need three digits to measure that range to 1°F. I can measure it to 0.1°C with three digits.

 

[Anton Hunter]

"We noted in our review of the AMD Radeon RX 5700 and Radeon RX 5700 XT that junction temperatures peaked above 100 degrees Celsius under load. (That's 230 degrees Fahrenheit for those of us forced to live with the clearly inferior temperature scale.)"

Um...since when does 100 degrees C = 230 degrees F? Anyone who has gotten past the eighth grade (in the USA education system) should know that 100C = 212F . So, then, the guy writing this piece is a real bonified tech? As-well-as a writer? Starting to look a little sketchy to me. My-oh-my, how the education system has taken a nose-dive under the watchful eyes of liberal politicians. Oh, BTW, 230 degrees C = 110 degrees F...just so you know! LOL Don't take it to heart - we all blunder from time-to-time. ^_^

 

[TerryLaze]

"We noted in our review of the AMD Radeon RX 5700 and Radeon RX 5700 XT that junction temperatures peaked above 100 degrees Celsius under load. (That's 230 degrees Fahrenheit for those of us forced to live with the clearly inferior temperature scale.)"

Um...since when does 100 degrees C = 230 degrees F? Anyone who has gotten past the eighth grade (in the USA education system) should know that 100C = 212F . So, then, the guy writing this piece is a real bonified tech? As-well-as a writer? Starting to look a little sketchy to me. My-oh-my, how the education system has taken a nose-dive under the watchful eyes of liberal politicians. Oh, BTW, 230 degrees C = 110 degrees F...just so you know! LOL Don't take it to heart - we all blunder from time-to-time. ^_^

Don't throw that many stones,what education system did you go through that didn't taught you what above means?!?!
" junction temperatures peaked above 100 degrees Celsius under load."