AMD Fury X Postmortem: What Went Wrong?

[YBS1]

Agreed, the 290 series was fantastic hardware. Had the complete package been ready to roll day one the battle could have went differently, same could mostly be said about the 7900 series.

 

[JDG1980]

If they were going with +20% shaders over the 290X they should have just dropped HBM completely. GB for GB at this time HBM is probably twice as expensive as GDDR5 and you have the cost of the interposer ($1 per 100 mm^2 is about $10) though you can cut down the on the PCB.

One problem with the Fury X is that they had to add the expense of a CLC and high-quality fan (the Gentle Typhoon can't be that cheap, even in quantity). Also, since the board has a very high power limit (over 430W) and was apparently designed for overclocking (even though no voltage control utility is yet available - go figure), it had to be built with the same beefy power-control components found in Hawaii and other high-TDP cards.

If we assume that AMD needed something as a test bed for HBM, and that it had to be a shipping product because AMD can't afford to waste any money, then something just a little bigger than Hawaii starts to look like it would have been a better option. AMD could have created a chip with 3200 shaders (50 GCN 1.2 compute units) and the same 64 ROPs that Hawaii and Fiji got. Using HBM would mean that Hawaii's big 512-bit GDDR5 memory controllers could be cut out, which would save quite a bit of die space. Even with updated UVD/VCE blocks, you could probably keep the die size under 450 mm^2, which would be far cheaper to manufacture than Fury. The interposer would be correspondingly smaller, too. A chip like this would be designed to compete on price, performance, and perf/watt with GM204. Even a R9 390X which is just an overclocked Hawaii card is pretty close to a GTX 980 in pure performance; its drawbacks are its outdated feature set and enormous power consumption. HBM cuts way back on power usage, and with AMD trying to compete with the GTX 980 instead of the GTX 980 Ti and Titan X, there would be no need to overclock the balls off the chip just to eke out a bit of extra performance. You'd wind up with something closer to what the Fury Nano looks like it might be, except it would be purpose-built (rather than detuned) and could be manufactured much cheaper. One 8-pin connector, a much more modest power stage, and a small PCB with a less expensive cooler. It could debut at a price point of $499 and still make some money. At that price point, the 4GB RAM limit also wouldn't be a problem.

Fury should have never made it off the drawing board. 4 GB Vram prevents it from being sold in any large numbers in the professional market (at least they castrated DP). AMD should have realized that the costs of fury were too high with HBM and an interposer and the break even cost was prohibitive. After the disaster that was Tonga AMD should have been more careful about releasing new cards.

I think both Tonga and Fiji were tech demos that wound up being released because AMD is financially strapped and had to squeeze whatever money they could out of every product they have. I'm sure that Nvidia has worked on HBM tech demos as well and they may even have taped out, but they didn't bother to ship them because it wouldn't have been worth it and it might have damaged their brand image.

It's interesting to note that according to a leaked slide Maddie posted here a while back, both Tonga and Fiji look to have been produced on "28HPM" - a TSMC mobile process. Maybe they were originally designed to be on 20nm TSMC and had to be backported when it turned out that was a nonstarter, and it was easier/cheaper to port to 28HPM than to GloFo or a more performance-focused TSMC process. Or maybe AMD realized they would have to be working with a mobile-focused process on 14nm FinFET and decided to get a head start learning how to handle it.

 

[utahraptor]

I think both Tonga and Fiji were tech demos that wound up being released because AMD is financially strapped and had to squeeze whatever money they could out of every product they have. I'm sure that Nvidia has worked on HBM tech demos as well and they may even have taped out, but they didn't bother to ship them because it wouldn't have been worth it and it might have damaged their brand image.

You may be right (HBM 2)...

http://www.kitguru.net/components/g...ly-taped-out-on-track-for-2016-launch-rumour/

 

[x3sphere]

https://youtu.be/8hnuj1OZAJs?t=91
https://youtu.be/8hnuj1OZAJs?t=136

I was just watching some gameplay footage with this card at 4K in GTA V. Look at what happens at the 1:39 and 2:20 mark, massive stutter. It's like the card is running out of VRAM because the usage goes down a lot when this happens, speculation here but seems like it's hitting swap memory.

 

[RampantAndroid]

https://youtu.be/8hnuj1OZAJs?t=91
https://youtu.be/8hnuj1OZAJs?t=136

I was just watching some gameplay footage with this card at 4K in GTA V. Look at what happens at the 1:39 and 2:20 mark, massive stutter. It's like the card is running out of VRAM because the usage goes down a lot when this happens, speculation here but seems like it's hitting swap memory.

Interesting - it drops a GB when it stutters....but it never hits 4GB (on his 980ti video, it's above 4GB of VRAM in use.)

 

[5150Joker]

https://youtu.be/8hnuj1OZAJs?t=91
https://youtu.be/8hnuj1OZAJs?t=136

I was just watching some gameplay footage with this card at 4K in GTA V. Look at what happens at the 1:39 and 2:20 mark, massive stutter. It's like the card is running out of VRAM because the usage goes down a lot when this happens, speculation here but seems like it's hitting swap memory.

But but but HBM is like cache and 4 GB isn't limiting at 4k! Basically AMD will have to work the same kind of driver bs NVIDIA did with 970 to manage the limited VRAM. Except 970 is a mid tier card and this is AMD's best.

 

[RampantAndroid]

But but but HBM is like cache and 4 GB isn't limiting at 4k! Basically AMD will have to work the same kind of driver bs NVIDIA did with 970 to manage the limited VRAM. Except 970 is a mid tier card and this is AMD's best.

Again, I don't know if this is the 4GB limit being hit. At times, it stutters below 3GB usage. Maybe MSI Afterburner is reading the usage wrong....or maybe it's some other issue.

 

[5150Joker]

Again, I don't know if this is the 4GB limit being hit. At times, it stutters below 3GB usage. Maybe MSI Afterburner is reading the usage wrong....or maybe it's some other issue.

NVIDIA juggles the vram usage via drivers in the 970 to avoid it hitting the 4 GB wall by keeping it under 3.5 GB even when it goes >4 GB in other cards w/the capacity. I suspect the same thing is being done for the Fury X except it's not working so good.

 

[.vodka]

HBM is a completely different paradigm compared to GDDR5. I suppose there is a lot of untapped potential in the driver side going forward because of that. nV will have to face this with Pascal, too.

Remember that GDDR5 appeared with the 4870 7 years ago... it's had a lot of time for polishing on both hardware and software sides of the coin.


Do the other 4GB cards (290/390x, 980) have the same stuttering problem as Fury in these tests? If they don't, that rules out a capacity problem.

 

[utahraptor]

This is slightly off topic, but years ago when I had my Pentium 4 2.6c rig and a ATI 9800 Pro (128 mb I think) when playing World of Warcraft if I spun the camera around real fast it would suddenly stutter to hell. Would that have been a VRAM issue? I have always wondered.

 

[Azix]

But but but HBM is like cache and 4 GB isn't limiting at 4k! Basically AMD will have to work the same kind of driver bs NVIDIA did with 970 to manage the limited VRAM. Except 970 is a mid tier card and this is AMD's best.

I'd guess drivers like the person doing the video said.

I do think AMD should be focusing on lower res, unless they have something planned beyond 4k. 4K will always be a losing battle unless something changes. Maybe 14nm and beyond. When they do take all this tech to 14/16nm maybe.

 

[Spanners]

NVIDIA juggles the vram usage via drivers in the 970 to avoid it hitting the 4 GB wall by keeping it under 3.5 GB even when it goes >4 GB in other cards w/the capacity. I suspect the same thing is being done for the Fury X except it's not working so good.

That's not a comparable situation so I doubt that's the case. The GTX 970 looks to the game engine/OS like a 4GB card but in reality isn't (in terms of full bandwidth memory). Hence the need for specific 970 optimisations to avoid/de-prioritise that lower speed portion. Fury X already reports the correct VRAM amount why would they need to "juggle" to avoid hitting 4GB? The game engine is already aware the card is 4GB and isn't going to try and allocate more memory than is present on the card. No need for driver kludges.

 

[utahraptor]

You all know GTA V warns you with a big red bar when you are using settings that will lead to more VRAM being used then is available right? (I ignore the warning because geforce experience says to and it seems to work fine)

 

[RampantAndroid]

You all know GTA V warns you with a big red bar when you are using settings that will lead to more VRAM being used then is available right? (I ignore the warning because geforce experience says to and it seems to work fine)

Sure, but the usage in that video doesn't approach the max of the card?

 

[x3sphere]

That's not a comparable situation so I doubt that's the case. The GTX 970 looks to the game engine/OS like a 4GB card but in reality isn't (in terms of full bandwidth memory). Hence the need for specific 970 optimisations to avoid/de-prioritise that lower speed portion. Fury X already reports the correct VRAM amount why would they need to "juggle" to avoid hitting 4GB? The game engine is already aware the card is 4GB and isn't going to try and allocate more memory than is present on the card. No need for driver kludges.

I think what he is getting at is AMD said they would be doing driver optimizations to avoid 4GB being a limitation at 4K. This may consist of intelligently swapping data out/in BEFORE the 4GB limit is hit to avoid hitting system RAM entirely. But here in GTA V, we see it may not be working well.

Something is going on with VRAM in that video because the usage drops big time when it hits those stutter points, almost as if the cache gets wiped out and then the game starts filling it again.

 

[RampantAndroid]

Forgive me if someone linked this already:

https://youtu.be/iEwLtqbBw90?t=7m32s

Another review with a noisy pump. :-/

 

[utahraptor]

Forgive me if someone linked this already:

https://youtu.be/iEwLtqbBw90?t=7m32s

Another review with a noisy pump. :-/

Yesterday when I saw the first coil whine / pump noise post I had a physical cringe. Felt like I had a dodged a bullet in a way. I was so excited for the launch that those two factors did not even weigh on my mind. After seeing the performance I would never consider buying it, but after hearing that I just felt sick to my stomach for AMD. Is it too soon to bring out... erm, #pumpgate?

 

[Silverforce11]

Yesterday when I saw the first coil whine / pump noise post I had a physical cringe. Felt like I had a dodged a bullet in a way. I was so excited for the launch that those two factors did not even weight on my mind. After seeing the performance I would never consider buying it, but after hearing that I just felt sick to my stomach for AMD. Is it too soon to bring out... erm, #pumpgate?

#pumpgate would apply if they did not allow buyers to RMA it. Just like coil whine happens, AIO do have faulty samples that are noisy.

 

[RampantAndroid]

#pumpgate would apply if they did not allow buyers to RMA it. Just like coil whine happens, AIO do have faulty samples that are noisy.

It's just a matter of how common the problem is...

 

[Silverforce11]

Looks to be a Fury X issue because the 4GB R290X has no stutters.

Again,

4GB GPUs, like 980 & R290X have no issues and in CF/SLI, beat Titan X which has 12GB vram. You can see the fps chart yourself, no massive stutters.

This combined with poor performance at lower resolution (not CPU overhead, because R295X2 scaling is still great, no CPU bottlenecks) suggest Fury X has rushed drivers, unoptimized.

Combined with the lack of vcore modding tools for the press means AMD were just not ready but had to launch anyway due to delays and the 980Ti putting pressure on them.

 

[x3sphere]

Combined with the lack of vcore modding tools for the press means AMD were just not ready but had to launch anyway due to delays and the 980Ti putting pressure on them.

I agree. I'm hoping for a much better showing when the regular Fury releases next month.

 

[Hitman928]

Easy pump noise fix?

https://www.youtube.com/watch?v=0VUeTd6aPN8

 

[utahraptor]

Easy pump noise fix?

https://www.youtube.com/watch?v=0VUeTd6aPN8

I am starting to think the closed loop coolers are just a scam. I myself purchased the corsair h100 for my 3770k and was extremely disturbed to find that at medium or high the pump was louder than an air cooler by a large margin and only low was tolerable. Further more the performance when overclocking was terribad at any speed. I never got to experience the joy of a 2600K. :\

 

[RampantAndroid]

I am starting to think the closed loop coolers are just a scam. I myself purchased the corsair h100 for my 3770k and was extremely disturbed to find that at medium or high the pump was louder than an air cooler by a large margin and only low was tolerable. Further more the performance when overclocking was terribad at any speed. I never got to experience the joy of a 2600K. :\

For CPUs, the AIOs have NEVER been as good as real water cooling. They've only ever been as good as a top end heatsink...with the main difference being that the failure point on a heatsink is the fans, and if they die the heatsink still works. The added failure points on the AIO cooling are the pumps, tubing and connections. If they fail, enjoy buying new parts.

It's odd that pressure would do it. Are the mounts loose?

 

[Stormflux]

After watching The TechReport podcast, which was quite interesting:
https://www.youtube.com/watch?v=28CECF_Cieo

I'm real curious to see how the FuryX's huge shader array can be used to it's advantage. They discussed it around the 37 minute mark.

The big sticking point was divergent technologies, NVidia specializing in Geometry output, and AMD specializing in Compute and Shader throughput. The former being best for right now, and the latter being better for tomorrow, allegedly.

VR being a huge upcoming market and a lot of VR companies utilizing LiquidVR and it's compute centric approach to problems. This might become THE card for VR.

Reviews doing Apples to Apples comparisons using the exact same settings possibly aren't playing to the strengths of each card. MSAA vs. SMAA for example.

Now this does get fuzzy in how to directly compare the cards, but I just thought it was quite an interesting point in regards to the disappointing showings of the FuryX.