AMD reveals its CPU & GPU roadmap for next five years

Rav3n said:
AMD and efficient//// :confused: I guess efficiency started with Bulldozer...:eek:
Click to expand...

If you're talking strictly about CPUs, yeah no argument there.

When it comes to GPUs however, 4870, 5870, and 6970 all tied or beat their Tesla and Fermi counterparts in the performance per watt race. Wasn't until 7970 that it took a back seat against 680, but still wasn't a day and night difference -- more like within 10%. Per/watt charts on TechPowerUp says it all.
 
Last edited:
According to the article the 300w APU is intended for the HPC market, so that's definitely not a consumer product. It seems like an APU would make sense in the HPC market if they have unified memory working.

Without unified memory I think an HPC application would just be better off with a discrete card. To me this seems to say that AMD does have confidence achieving a unified memory solution.
 
jwcalla said:
According to the article the 300w APU is intended for the HPC market, so that's definitely not a consumer product. It seems like an APU would make sense in the HPC market if they have unified memory working.

Without unified memory I think an HPC application would just be better off with a discrete card. To me this seems to say that AMD does have confidence achieving a unified memory solution.
Click to expand...

It is also one of the reasons they purchased SeaMicro.
 
n=1 said:
When it comes to GPUs however, 4870, 5870, and 6970 all tied or beat their Tesla and Fermi counterparts in the performance per watt race. Wasn't until 7970 that it took a back seat against 680, but still wasn't a day and night difference -- more like within 10%. Per/watt charts on TechPowerUp says it all.
Click to expand...

It seems like a small difference when you look at simple benchmarks.

I was as hopeful as you were that AMD would prove Tahiti would be highly efficienct, but AMD was never able to bring the full or partial underclocked Tahiti card to market on the mobile side. It could not provide anywhere near full performance without topping the 100w barrier, so AMD fell back on relabeling an underclocked full Pitcairn as the HD 7970M, and then the R9 M290X. Meanwhile Nvidia found a way to sell the partial-enabled GTX 670 at 100w with lower clocks (680M), and one year later managed a core rev that fit the fill GTX 680 at reduced clock (780M) so you can guess who won that performance crown for 3 years.

I also have to question AMD's management during this period, as the impressively efficient Bonaire chip never saw light in any mobile parts. And of course, AMD's mobile savior Tonga was late and underpowered (was obliterated by the GTX 970M, which launched months earlier). What shining management, and sheer efficiency they have shown us!
 
Last edited:
defaultluser said:
It seems like a small difference when you look at simple benchmarks.

I was as hopeful as you were that AMD would prove Tahiti would be highly efficienct, but AMD was never able to bring the full or partial underclocked Tahiti card to market on the mobile side. It could not provide anywhere near full performance without topping the 100w barrier, so AMD fell back on relabeling an underclocked full Pitcairn as the HD 7970M, and then the R9 M290X. Meanwhile Nvidia found a way to sell the partial-enabled GTX 670 at 100w with lower clocks (680M), and one year later managed a core rev that fit the fill GTX 680 at reduced clock (780M) so you can guess who won that performance crown for 3 years.

I also have to question AMD's management during this period, as the impressively efficient Bonaire chip never saw light in any mobile parts. And of course, AMD's mobile savior Tonga was late and underpowered (was obliterated by the GTX 970M, which launched months earlier). What shining management, and sheer efficiency they have shown us!
Click to expand...

You have just summed up the Rory Read dog and pony show. He's gone now, though.
 
PRIME1 said:
A 300W APU. I think they have officially given up.
Click to expand...

Quite the opposite. Once AMD starts releasing relatively competitive high performance APU's with powerful built in graphics, it's over for nvidia. Mantle and more importantly dx12/vulkan is laying the groundwork for better utilization of leveraging different gpus in a system, even if they are of differing strengths. So once this more powerful apu launches with much higher built in memory bandwidth with ddr4 or something more exotic, everyone who goes with the amd apu has a built in second gpu in addition to their primary gpu.

If it's possible you could couple that with an nvidia gpu then it would not matter, but if nvidia makes it so that only other nvidia gpus are allowed to work together to improve overall performance, it's game over.

By default every single amd apu + discreet amd gpu user will have a built in gpu advantage over an equivalently priced/specced intel + nvidia combo, and it might not even be close.

nvidia can flail about creating titan omegas or god knows what for thousands of dollars, but it will be a case of the emperor has no clothes. They will wither, they will wane, and they will die.


EDIT:

http://www.extremetech.com/computin...-between-wide-io-hbm-and-hybrid-memory-cube/2

this suggests amd is eyeing hbm for a future apu, so I think we may have a higher performing bandwidth unchained apu future, with 8-16 cores (I hope).
 
Last edited:
Kor said:
HBM will be unreasonably expensive in 2.5D form for some time to come, it's not cost effective for the APU market. Hell running alternate and more expensive memory types has proven to be a bad idea historically. Just ask Intel how well RDRAM worked out for the Willamette Pentium 4.
Click to expand...

We are still in the first half of 2015, the rumored release of said high performance APU is 2017, that is two years away, more than enough time for prices to come down, especially if they start selling hbm memory for gpus this year with amd and next year with both amd and nvidia.
 
defaultluser said:
It seems like a small difference when you look at simple benchmarks.

I was as hopeful as you were that AMD would prove Tahiti would be highly efficienct, but AMD was never able to bring the full or partial underclocked Tahiti card to market on the mobile side. It could not provide anywhere near full performance without topping the 100w barrier, so AMD fell back on relabeling an underclocked full Pitcairn as the HD 7970M, and then the R9 M290X. Meanwhile Nvidia found a way to sell the partial-enabled GTX 670 at 100w with lower clocks (680M), and one year later managed a core rev that fit the fill GTX 680 at reduced clock (780M) so you can guess who won that performance crown for 3 years.

I also have to question AMD's management during this period, as the impressively efficient Bonaire chip never saw light in any mobile parts. And of course, AMD's mobile savior Tonga was late and underpowered (was obliterated by the GTX 970M, which launched months earlier). What shining management, and sheer efficiency they have shown us!
Click to expand...

I was simply pointing out that contrary to what some may think or assert, AMD was actually ahead in the efficiency race for 3 generations, and it wasn't until Kepler that nVidia made major gains in efficiency.

On the mobile side, the 7970M was typically within 10-15% of the 680M, so it wasn't too big of a gap. But I should point out it actually got rebranded TWICE -- first as the 8970M but at least with a 50MHz bump on core clock, then the M290X was just a 100% rebrand of 8970M without changing a goddamn thing. And the Tonga based M295X is just a sad joke with its 125W TDP.

Now to be fair, the 780M's official TDP might've been specified as 100W, but in practice to maintain full boost the TDP cap had to be raised to 110W (and KBT shows the default TDP as 110W in the vbios as well), so it ran pretty toasty. I know because I own a 780M SLI laptop. Then nVidia release the short lived 880M, which was a total abomination and complete failure. It ran ass hot and a was a throttling mess, unless you increased the TDP cap to 125W to maintain a constant 954 base clock.

Bottom line is nothing good ever comes from shoving a 125W part into a notebook, as both companies have demonstrated with their fails.
 
PRIME1 said:
A 300W APU. I think they have officially given up.
Click to expand...

A 300W APU with HBM/V2 could be fucking amazing even if they were using their current GPU tech with a die-shrink but it looks like it's going to be new tech.

We've yet to see how Zen is going to perform but if it's better than steamroller that could one badass chip. Not to mention it will more than likely be targeted at the non-consumer segments.

A planned 300W APU from AMD doesn't by any stretch of means spell doom or that they've given. I've seen way too many comments like this since this news came out.
 
Kor said:
HBM will be unreasonably expensive in 2.5D form for some time to come, it's not cost effective for the APU market. Hell running alternate and more expensive memory types has proven to be a bad idea historically. Just ask Intel how well RDRAM worked out for the Willamette Pentium 4.
Click to expand...

One, cost has little to do with the original argument. Original argument was that an APU of that power would be greatly bottlenecked by the bandwidth of current ram, albeit true, this APU is in the distant future where the new found tech HBM, with a vast bandwidth, will likely alleviate that particular problem.

Two, Cost is irrelevant and not generally the biggest concern in the Server Market which the afore mentioned APU is targeted.

add: And then consider the Titan: Why buy it? Its cost is greatly exaggerated compared to the cost of other GPUs. Many who do purchase Titans claim it is the best performer. So now Consider HBM...
 
Last edited:
tybert7 said:
If it's possible you could couple that with an nvidia gpu then it would not matter, but if nvidia makes it so that only other nvidia gpus are allowed to work together to improve overall performance, it's game over.
Click to expand...

We have been hearing about how "it's game over" for nvidia for years now, but they continue to find ways to get into new markets and still turn profits quarter after quarter, despite the fact that, yes, Intel (not AMD) is carving out the largest segment of the GPU market.

Part of the reason nvidia has been able to succeed is because its CEO and management are not a bunch of drooling fools.

nvidia ... will wither, they will wane, and they will die.
Click to expand...

I... got nothin'.
 
Problem with a high end APU is no matter how awesome they make the IGP, it's still going to be hampered by the memory sub system which is quite a bit slower than GDDR5 found on video cards. They could compensate for that buy having a large "eSRAM" similar to the XBOX One, but I'd imagine that would add to the cost of the processor significantly.
 
Yup, Intel has the right approach: Since Ivy Bridge after they figured out the L3 cache sharing thing with Sandy, concentrate on shipping the minimum die size for the desired performance level. Their goal was to design a scalable architecture, and only produce as much silicon as you need. There are typically four core variants:

2c gt2
2c gt3
4c gt2
4c gt3

And they can typically forecast demand and yield well enough to keep the channel supplied with the optimal mix of parts. GT3 was canceled for Ivy Bridge, but was there if the OEMs showed interest. And they offer a separate dram cache for those rare OEMs that desire it.

And see, this right here is why AMD was destined to fail in the high-performance mainstream APU market: they don't have the volume to afford multiple scaled GPU performance levels, or offer expensive add-ons like DRAM cache. That means that AMD only ships one BIG die. And whether all GPU compute units are enabled or not, it still costs top-dollar to make. When you spend the kind of die space AMD does on the GPU, you really need to sell ALL those suckers for a premium.

See here, where Trinity 4c (32nm) is larger than Sandy Bridge 4c with GT2 (32nm):

http://www.anandtech.com/show/7003/the-haswell-review-intel-core-i74770k-i54560k-tested/5

This is also why Brazos/Jaguar has been making tons of profit for them, until the netbook market dried up. The right balance of GPU and CPU performance. Trinity would have been just fine with 192-256 cores, but they went overboard, so go figure.
 
Last edited:
Man.. dyed in the wool AMD user here. Since Thunderbird in fact.

I feel an Intel build coming on. Never thought I'd say that.
 
Rav3n said:
I disagree. The enthusiast market (>2% of all PC sales) thought it was a terrible idea. Most PC's are pre built and 98% of them contain integrated GPU's.

Wed hardware enthusiasts are the minority. Intel knew exactly where the bread and butter was and is.
Click to expand...
You're right. However I still wish Intel would make a few SKU's without or minimal only iGPUs. The die area devoted to graphics in current CPUs is significant, so without that they could be smaller and thus cheaper or more powerful. The "K" versions would be a good candidate for that.
 
What is the chance of having HBM onboard as cache for these motherboards?
 
zero, they run too hot.

3D HBM is currently only usable in lower power SOC on-die.

They could use 2.5D with an interposer, but that would make it rather large.
 
The Mac said:
zero, they run too hot.

3D HBM is currently only usable in lower power SOC on-die.

They could use 2.5D with an interposer, but that would make it rather large.
Click to expand...

Im not talking about "on die" as in- on the chip.
Im talking about motherboard makers using High bandwidth memory on the motherboard.

AMD K6-III Used memory (2 megs) on the motherboard.
Today that could be fast HBM unattached to the CPU/APU like they do on graphics cards.

It seems that DDR4 memory would cripple any APU compared to HBM used on the video card. So why not place HBM on the motherboard?
 
HBM requires a massive density of traces.

This is the reason for interposers, they are manufactured separately from the main PCB.

What you are suggesting with HBM on the Motherboard would be OUTRAGEOUSLY expensive to manufacture, and im not even sure it would be possible in any durable form appropriate for the retail market.
 
griff30 said:
Im not talking about "on die" as in- on the chip.
Im talking about motherboard makers using High bandwidth memory on the motherboard.

AMD K6-III Used memory (2 megs) on the motherboard.
Today that could be fast HBM unattached to the CPU/APU like they do on graphics cards.

It seems that DDR4 memory would cripple any APU compared to HBM used on the video card. So why not place HBM on the motherboard?
Click to expand...

You do know why they shipped with this, right?

It was because existing Super 7 motherboards had all shipped with off-chip L2 cache. The K6-III was just reusing that as L3 cache. All Super 7 motherboards shipped with this cache by default because the vast majority of users were using K6-2 processors with no L2 cache, and without this the memory subsystem of the Super 7 motherboard would have been even more of a performance constraint with each faster-clocked chip released.

On-board cache became ubiquitous because it was cheapo to add when (1) built in quantity and high-density, and (2) non-upgradable. Five years previous, no motherboard in existence shipped with high-speed built-in cache, because it was too expensive. HBM is more at the "too expensive to give away for free" point in it's lifetime.

Also, I thought the purpose of HBM was to put the chip as close to the die as possible, to reduce latency and to also cut power consumption. Not sure if you can get those same benefits at the socket-level, or PCIe level.
 
I'm excited about the 300watt APU. The PS4 uses an APU that's what, 150 watts under full load? So AMD is giving us a chip probably significantly faster than next gen consoles for ~$150. I would slap that bad lad in an itx case with an AiO cooler faster than a redheaded step child.
 
Giving an APU like this a sufficient TDP means AMD is finally going to get their APUs taken seriously.

So far they have packed in a pretty impressive amount of GPU performance for their given wattage, but it's still pretty meager.

Really, it should amount to an APU that people have been hoping for since Llano. However the RAM bottleneck may still an issue for gamers, but since they seem to be targeting clusters with it, I guess that wont matter so much for that particular application. At least you can expect the GPU to have good compute performance.
 
Reading the last few pages shows me how seriously misinformed some people are about the AMD APU's in the latest road map. Wow!
 
The Mac said:
HBM requires a massive density of traces.

This is the reason for interposers, they are manufactured separately from the main PCB.

What you are suggesting with HBM on the Motherboard would be OUTRAGEOUSLY expensive to manufacture, and im not even sure it would be possible in any durable form appropriate for the retail market.
Click to expand...

Exactly HBM gets its main benefits from keeping the trace length as short as possible. The longer the trace the more power it uses, which goes up exponentially when pushing data around.

If someone wanted to do something like that, I would think HMC would be the better option except for cost.
 
LordEC911 said:
Exactly HBM gets its main benefits from keeping the trace length as short as possible. The longer the trace the more power it uses, which goes up exponentially when pushing data around.

If someone wanted to do something like that, I would think HMC would be the better option except for cost.
Click to expand...

Which may be a reason to use... wait for it... optical connectors for memory !!!!

https://www.youtube.com/watch?v=QPQ1AheNro8#t=9m20s


then maybe large clusters of memory could be added where both the system and gpu could access the same memory set without the large distance penalties.... or not, I have no idea what I'm talking about here, just hoping it's real.
 
tybert7 said:
Which may be a reason to use... wait for it... optical connectors for memory !!!!

https://www.youtube.com/watch?v=QPQ1AheNro8#t=9m20s


then maybe large clusters of memory could be added where both the system and gpu could access the same memory set without the large distance penalties.... or not, I have no idea what I'm talking about here, just hoping it's real.
Click to expand...

It is real.
They have been talking about this for a good 20 years at least.

The issue is shrinking it down to the micro-scales needed to actually integrate it into semiconductors.
 
Last edited:
I assume it's easier to make optical connections for components. Why isn't that mainstream yet?
 
Only problem here is here is you guys are using current HBM tech to rationalize a future APU part that is a year or more away.
 
griff30 said:
AMD K6-III Used memory (2 megs) on the motherboard.
Today that could be fast HBM unattached to the CPU/APU like they do on graphics cards.
Click to expand...

That was a poor man's solution to a growing problem that would of cost AMD millions to fix (adding a lot of cache to their CPU). It was by no means ideal to do. Intel even introduced it as a cheap way to add more cache to their Celeron's.

The way they worked they already had a direct pipeline to the CPU. If we were to do this with our current CPU's it would drag them down drastically. Also HBM isn't a method you can apply anywhere to the computer subsystem and BAM you have a billion gigabits of bandwidth, HBM exists solely due to its location.
 
You must log in or register to reply here.
Back
Top