AMD CEO Lisa Su Interview: Confident In Next Graphics Launch, Zen's Success Is Key

[junkeymonkey]

ya , its not like I left amd chips and boards for the rirst time in over 14 years but now here I sit with my old 7000 card looking at this .. first what did the 200 cards really offer me over my work in well 7000 card that would make me spend to replace it ?? little to nothing .. so here we are with another refresh once again whats my incentive ? nothing .. my poor old 7000 is still holding its own no issues .. so far amd is loosing me as for as a sale .. and seeing the ''new fluffy '' card is going to be priced for more then I'm more than likely willing to pay they loose out once more on that...

amd I guess don't look at guys like me who been long term dedicated fan boys that now see the need to jump ship cause there offerings just fall short like this zen thing how long has that been talked up ??

maybe ''next year''

 

[Onus]

As boring as this strategy may be to enthusiasts, I believe I said a year or two ago (probably more than once, as repetitive as I can be), that AMD was wise to focus on the high-volume segments, which are adequately served by an APU. I hope this buys AMD the wiggle room they need to develop into other areas too, bringing back competition to Intel and nVidia.
I don't like rebrands, but as far as the new stuff goes, I like where AMD is focusing its efforts, and hope they bear fruit soon enough to matter.

 

[ubercake]

This seems like some really great tech. I really want to be excited about all of this information, but the hype that AMD has neglected to live up to in the past (especially regarding CPUs) is always something I think about.

I'll wait and see what the Tom's reviews uncover about the new AMD tech before deciding on whether or not to use it again.

 

[jimmysmitty]

As boring as this strategy may be to enthusiasts, I believe I said a year or two ago (probably more than once, as repetitive as I can be), that AMD was wise to focus on the high-volume segments, which are adequately served by an APU. I hope this buys AMD the wiggle room they need to develop into other areas too, bringing back competition to Intel and nVidia.
I don't like rebrands, but as far as the new stuff goes, I like where AMD is focusing its efforts, and hope they bear fruit soon enough to matter.

I think one of the issues was in corporate. They canned Dirk Meyers who actually steered them in the proper direction and was trying to focus on markets that could make them the money needed to focus on other markets that were not as profitable which I agree was the proper course, especially after the disaster that was K10.

But the past is the past. Can't change it unless you are a terminator.......

 

[Onus]

Ubercake is right though. Having been burned myself thinking / hoping Bulldozer would at least partially live up to its hype, as far as new stuff from AMD goes, I'll believe it when I see it.
Whether from Lisa Su's more cautious approach, to the absence of a lot of hoopla from AMD, I am optimistic though, that their marketing guys have been called to heel, warned that another fiasco could doom the company. This time, they absolutely must be able to walk the walk.

 

[InvalidError]

Fiji XT, however, has a pretty big change to it with HBM and it isn't as easy as slapping it on with an interposer. They now have extra heat added to the GPU to dissipate as well as other changes to compensate for.

Shortening the memory bus from several centimeters to a few millimeters should allow the use of greatly simplified and much lower power IO circuitry at both ends of the interposer traces. HBM likely also enables AMD to simplify a bunch of things in their memory controller and scheduler. At the end of the day, I would not be surprised if HBM causes a net reduction in total package power compared to off-package DRAM of comparable bandwidth.

If you have a 512bits wide interface at 15pf per trace operating at 1.5V and 6Gbps/pin, driving the memory bus alone can use nearly 30W. Almost eliminating trace and package capacitance, along with dropping the voltage to about 1V should shave 20-25W off of that despite bumping the interface width to 4096bits.

Since a 4096 shader core on 28nm would end up pushing a gigantic 600sqmm and an extra 100W of TDP, I am hoping AMD will have that made on 20nm. I am surprised the target process does not appear to have been leaked yet.

 

[junkeymonkey]

''I'll wait and see what the Tom's reviews uncover about the new AMD tech before deciding on whether or not to use it again''

like with NVidia - I'm glad I did not jump quickly on that 900 band wagon .. saved me good money holding off on that deal [970]

 

[jimmysmitty]

Ubercake is right though. Having been burned myself thinking / hoping Bulldozer would at least partially live up to its hype, as far as new stuff from AMD goes, I'll believe it when I see it.
Whether from Lisa Su's more cautious approach, to the absence of a lot of hoopla from AMD, I am optimistic though, that their marketing guys have been called to heel, warned that another fiasco could doom the company. This time, they absolutely must be able to walk the walk.

The biggest issue in marketing with BD was how they did it. They cherry picked every benchmark. In some they compared it to the older Nehalem uArch where it was better, in others they compared it to Sandy Bridge where it was better but lost to Nahalem. They never compared it to Sandy Bridge-E.

That was part of the downfall, basically cherry picked benchmarking. Then add on top that performance per clock was down compared to K10.5 and the power usage was through the roof, especially for an AMD CPU.

I am hopeful but reserved. Intel needs a kick in the butt in the CPU market. I love my CPUs but man Intel has been slacking recently.

 

[jimmysmitty]

Fiji XT, however, has a pretty big change to it with HBM and it isn't as easy as slapping it on with an interposer. They now have extra heat added to the GPU to dissipate as well as other changes to compensate for.

Shortening the memory bus from several centimeters to a few millimeters should allow the use of greatly simplified and much lower power IO circuitry at both ends of the interposer traces. HBM likely also enables AMD to simplify a bunch of things in their memory controller and scheduler. At the end of the day, I would not be surprised if HBM causes a net reduction in total package power compared to off-package DRAM of comparable bandwidth.

If you have a 512bits wide interface at 15pf per trace operating at 1.5V and 6Gbps/pin, driving the memory bus alone can use nearly 30W. Almost eliminating trace and package capacitance, along with dropping the voltage to about 1V should shave 20-25W off of that despite bumping the interface width to 4096bits.

Since a 4096 shader core on 28nm would end up pushing a gigantic 600sqmm and an extra 100W of TDP, I am hoping AMD will have that made on 20nm. I am surprised the target process does not appear to have been leaked yet.

I have seen rumors for both processes so far. I am hoping 20nm but no idea yet. I assume it would have to be if we consider how power hungry and hot Hawaii XT is and Fiji being based on it I doubt it has a ton of efficiency improvements in the core design itself.

7 days to go to find out.

 

[Eggz]

When it comes to current parts, I do Intel/Nvidia for performance and AMD for value. But when it comes to investing for payout, I think AMD shows a lot of promise right now. They have pretty much an all-time low stock price. Stock's spiked only three times in 37 years, which has corresponded to three significant first-to-market events, and it's possible that the fourth is just around the corner. Fingers crossed!

 

[f688xt6]

I was excited about FX when it debuted. When it didn't live up to expectations, I was disappointed. I really wanted it to work. In this case, I do hope that AMD manages to pull something off, but in a realistic view I don't see it happening. They need to make such a major leap forward without increasing cost in order to win back any significant portion of the market. I hope it happens, but I'm not going to hold my breath. And besides, Intel just released some compelling APU competitors as well.

 

[seraneth]

is this some more of the AMD 'next year '' funny thing with that is next year never comes around ?? look at there new cards how hard can it be to produce refresh items ?? so long a wait and delays .. AMD needs to put there money where there mouth is at some point in time .. all talk and no action don't sell [opinion]

Mod edits applied. Please leave off personal attacks. Thanks. Later gen FX and recent R9 AMD products rock the market.

 

[InvalidError]

And besides, Intel just released some compelling APU competitors as well.

Compelling is relative - few people who are interested in IGPs are willing to pay $270 for a CPU unless they have specific applications in mind where they have limited alternatives. For Broadwell to be truly compelling for budget gamers, Intel would need to lower the GT3e entry-level ticket price somewhere below the $200 mark.

At $270, Broadwell is not particularly compelling since you can get much better graphics performance out of a $120 i3 + $150 750Ti. Sure, Broadwell's IGP may get you by until you can toss a more powerful discrete GPU in but then the premium you paid for specifically for GT3e goes to waste.

 

[JackNaylorPE]

I dunno what the hullabaloo is all about with what peeps like to call "rebranding". No different than Intel's tick-tock. Hardware performance and yields improve over time, there's no better example of this than RAM where in the early days of a new platform, hi speed yields are low. 133 was the standard when DDR3 broke and grew to 1600, while hi speed 2400 stuff warranted a huge price increase.

Now we have 2400 often available within $2 - $5 of 1600. When a new GFX platform is launched, yields are low and the vendors target a certain speed where yields are enough to maintain a positive return. Over the next year, production line improvements mean that a new bar can be set.

nVidia pouts 30+% of it's income into R & D (dunno what AMD does) .... imagine how that number would have to climb to put out a brand new design every year. It's not feasible for nVidia / AMD just as it's not feasible for Intel.

 

[junkeymonkey]

'' Compelling is relative - few people who are interested in IGPs are willing to pay $270 for a CPU unless they have specific applications in mind where they have limited alternatives. For Broadwell to be truly compelling for budget gamers, Intel would need to lower the GT3e entry-level ticket price somewhere below the $200 mark.''

I wonder how well a xeon would sell if there were ''K'' models like the i7 and i5's ?? like I did not load the intel graphics drivers on this . you get something you really don't want or need taking up die space on the chip , but I guess in case of a emergency and you need to display I assume it could come in handy ?? then with all my old builds its not like I got ''back up ''cards boxed up to chunk in if needed to test

 

[redgarl]

For my part, I will past judgement when we actually know what is going to be released.

 

[jimmysmitty]

I dunno what the hullabaloo is all about with what peeps like to call "rebranding". No different than Intel's tick-tock. Hardware performance and yields improve over time, there's no better example of this than RAM where in the early days of a new platform, hi speed yields are low. 133 was the standard when DDR3 broke and grew to 1600, while hi speed 2400 stuff warranted a huge price increase.

Now we have 2400 often available within $2 - $5 of 1600. When a new GFX platform is launched, yields are low and the vendors target a certain speed where yields are enough to maintain a positive return. Over the next year, production line improvements mean that a new bar can be set.

nVidia pouts 30+% of it's income into R & D (dunno what AMD does) .... imagine how that number would have to climb to put out a brand new design every year. It's not feasible for nVidia / AMD just as it's not feasible for Intel.

I wouldn't call Intels tick-tock rebranding. Rebranding is taking the same product and renaming it. Intels tick-tock at least changes something. One is a die shrink (with some improvements sometimes) and the other is a new uArch design with more feature implementations.

Taking a Hawaii XT 8GB and calling it a R9 390X is not different at all, especially since you can actually buy and overclock a R9 290X to the same levels as is supposed to be the R9 390X.

I have nothing against it, it happens all the time in the GPU segment. NVidia is just a guilty of it.

I honestly do not know why AMD isn't making the R9 390x the Fiji XT, the R9 390 the Fiji Pro and then the R9 380x/380 the 290X/290 just as they did with Tahiti with the R9 series. But hey, I don't work for them so my input is useless.

 

[freeskier93]

Fiji XT, however, has a pretty big change to it with HBM and it isn't as easy as slapping it on with an interposer. They now have extra heat added to the GPU to dissipate as well as other changes to compensate for.

Shortening the memory bus from several centimeters to a few millimeters should allow the use of greatly simplified and much lower power IO circuitry at both ends of the interposer traces. HBM likely also enables AMD to simplify a bunch of things in their memory controller and scheduler. At the end of the day, I would not be surprised if HBM causes a net reduction in total package power compared to off-package DRAM of comparable bandwidth.

If you have a 512bits wide interface at 15pf per trace operating at 1.5V and 6Gbps/pin, driving the memory bus alone can use nearly 30W. Almost eliminating trace and package capacitance, along with dropping the voltage to about 1V should shave 20-25W off of that despite bumping the interface width to 4096bits.

Since a 4096 shader core on 28nm would end up pushing a gigantic 600sqmm and an extra 100W of TDP, I am hoping AMD will have that made on 20nm. I am surprised the target process does not appear to have been leaked yet.

Anandtech covers power consumption pretty well in their HMB article. "By AMD’s estimate 15-20% of Radeon R9 290X’s (250W TDP) power consumption is for memory" (Source). Anandtech estimates roughly a 50% reduction in memory power by switching from GDDR5 to HBM (Source).

 

[jimmysmitty]

Fiji XT, however, has a pretty big change to it with HBM and it isn't as easy as slapping it on with an interposer. They now have extra heat added to the GPU to dissipate as well as other changes to compensate for.

Shortening the memory bus from several centimeters to a few millimeters should allow the use of greatly simplified and much lower power IO circuitry at both ends of the interposer traces. HBM likely also enables AMD to simplify a bunch of things in their memory controller and scheduler. At the end of the day, I would not be surprised if HBM causes a net reduction in total package power compared to off-package DRAM of comparable bandwidth.

If you have a 512bits wide interface at 15pf per trace operating at 1.5V and 6Gbps/pin, driving the memory bus alone can use nearly 30W. Almost eliminating trace and package capacitance, along with dropping the voltage to about 1V should shave 20-25W off of that despite bumping the interface width to 4096bits.

Since a 4096 shader core on 28nm would end up pushing a gigantic 600sqmm and an extra 100W of TDP, I am hoping AMD will have that made on 20nm. I am surprised the target process does not appear to have been leaked yet.

Anandtech covers power consumption pretty well in their HMB article. "By AMD’s estimate 15-20% of Radeon R9 290X’s (250W TDP) power consumption is for memory" (Source). Anandtech estimates roughly a 50% reduction in memory power by switching from GDDR5 to HBM (Source).

So about 50w. That would be nice is it was Hawaii XT with HBM.

However, as Invalid pointed out, the core itself is going to be taking more power. It is going from 2816 SPUs to 4096 SPUs and if it is on 28nm that means the die size will also increase meaning it will suck way more power. If it is on 20nm then it should, in theory, drop power consumption as the die size should also shrink.

That is one of the biggest things we are waiting for the 16th for.

 

[JackNaylorPE]

Taking a Hawaii XT 8GB and calling it a R9 390X is not different at all, especially since you can actually buy and overclock a R9 290X to the same levels as is supposed to be the R9 390X.

I have nothing against it, it happens all the time in the GPU segment. NVidia is just a guilty of it.

The 390X however is a different platform w/ HBM, everything below that is the "tweaked" version of the 2xx series. And I have never given credence to the statement that you can OC this to that because it ignores the simple fact that "that" card can be OC'd to.

Again, I just don't see how the word "guilty" fits into what is standard procedure for most products.... iPhone 6 vs 6s, Ford Mustang, Monitors get a letter tacked onto the end cause this year they tacked on a headphone port, etc. And even if ya don't agree with what is a a very sound strategy and customer benefit (imagine cost and bug impacts with a once a year release cycle), by this point in time, it certainly should not be a surprise. If ya getting 10 - 15% more than last year at about the same price .... "Where's the beef ?"

Why do they do this ? Cause business product launches are run by the marketing departments.... the same folks who redefined a GB to include only 1000 MB, case it mage things sound bigger. Same reason there was no Windows 9 and no Word 3 thru 5 ... it sounds newer.....especially when your competitor is using a newer / higher number.

Why did AMD choose x90 for their top card ? ..... cause nVidia was using x80 is my guess. I think what we are seeing now has more to do with marketing and sales than technological limitatiions.... why is the 960 such a dud compared with previous generations ? Two 560 Tis were $100 cheaper and 40% faster than the 580 .... two 650 Ti Boosts were $100 cheaper and faster than the 680....Would nVidia purposely cut the performance of the card in that market niche so as to stop cannibalizing x80 sales ? .... I think so.

 

[Onus]

As far as I'm concerned, power usage going up is a non-starter. The cards I have now are already more than I need; get the power consumption way down if you want any [more] of my money.

 

[xchoice]

the only thing i care about amd is Zen.........

 

[jimmysmitty]

Taking a Hawaii XT 8GB and calling it a R9 390X is not different at all, especially since you can actually buy and overclock a R9 290X to the same levels as is supposed to be the R9 390X.

I have nothing against it, it happens all the time in the GPU segment. NVidia is just a guilty of it.

The 390X however is a different platform w/ HBM, everything below that is the "tweaked" version of the 2xx series. And I have never given credence to the statement that you can OC this to that because it ignores the simple fact that "that" card can be OC'd to.

Again, I just don't see how the word "guilty" fits into what is standard procedure for most products.... iPhone 6 vs 6s, Ford Mustang, Monitors get a letter tacked onto the end cause this year they tacked on a headphone port, etc. And even if ya don't agree with what is a a very sound strategy and customer benefit (imagine cost and bug impacts with a once a year release cycle), by this point in time, it certainly should not be a surprise. If ya getting 10 - 15% more than last year at about the same price .... "Where's the beef ?"

Why do they do this ? Cause business product launches are run by the marketing departments.... the same folks who redefined a GB to include only 1000 MB, case it mage things sound bigger. Same reason there was no Windows 9 and no Word 3 thru 5 ... it sounds newer.....especially when your competitor is using a newer / higher number.

Why did AMD choose x90 for their top card ? ..... cause nVidia was using x80 is my guess. I think what we are seeing now has more to do with marketing and sales than technological limitatiions.... why is the 960 such a dud compared with previous generations ? Two 560 Tis were $100 cheaper and 40% faster than the 580 .... two 650 Ti Boosts were $100 cheaper and faster than the 680....Would nVidia purposely cut the performance of the card in that market niche so as to stop cannibalizing x80 sales ? .... I think so.

You must have missed the slip up from XFX that confirms the R9 390X is just a rebranded Hawaii XT. The Fiji XT core will probably be the rumored Fury X naming, pitting it against the 980Ti and Titan class cards.

As I said, I am not against it just that rebranding is different from Intels tick-tock.

And I guess we will have to wait and see what kind of overclocks the R9 390X will get compared to the R9 290X. The thing is that when they rebranded the HD7970 to the R9 280X it was getting the same overclocks, about 1150-1200MHz core depending on the card and cooling and about 1600-1700MHz memory depending on the RAM ICs. It was almost a straight rebrand except the R9 280X came with a lower stock clock speed than the HD7970GHz.

Unless they somehow changed the core, I doubt we will see overclocked R9 390Xs much higher than current R9 290Xs.

 

[InvalidError]

Why do they do this ? Cause business product launches are run by the marketing departments.... the same folks who redefined a GB to include only 1000 MB, case it mage things sound bigger.

They never "redefined the GB" since the k/M/G/T/etc. prefixes are defined in the SI measurement system. They merely chose to adopt the correct metric definition. For people who want to use the "classic" computer science version which follows powers of 1024, the EIA defined the ki/Mi/Gi/Ti/etc. prefixes for those to disambiguate between proper metric prefixes and CompSci prefixes.

 

[ammaross]

The next line of graphics cards look great, can't wait for the 300 series. I am excited to see its performance and power usage. As for Zen, it seems they are getting smarter by moving toward higher IPC but I still don't think even a 40% IPC increase will reach what Skylake brings to the table. But if AMD pulls off high clock rates again and lots of cores it can for sure bring more competition to the table and drop Intel of the CPU dawg.

According to Steam Hardware survey Intel comes in 75% of all PCs that have a Steam account. It will be impossible to drop Intel in the next 4-5 years, AMDs chips would have to be THAT much better than Intels and substantially cheaper too as Intel can cut a few bucks and convince major PC makers to continue using it's chips.

I wish this lady nothing but good luck as AMD is the reason nVidia and Intel innovate, but she has her job cut out for her.

Intel is in 75% of gaming PCs due to their measurable performance increase due to IPC. Who wouldn't want a 10% boost to FPS just because you used an i5 instead of similar-priced high-end AMD? It's been that way for at least a couple generations now. This is likely the inadequacy that Zen's IPC fixes are trying to address.

Unfortunately, the 300 series isn't looking all that great though. Rebadged 200-series which were (mostly) rebadged 7000-series cards....*sigh*. A 7970 becomes an R9-280X becomes a R7-370 (or whatever one it falls into now). Seems if you want true new architecture, you have to get the 390X (or 290/X of that generation). The thing that always holds true though is FPS per dollar: even if the chip itself is 3 years old, its price is in line with nVidia's performance per dollar curve, so it doesn't matter.