AMD - Bulldozer Chitchat

[scaryjim]

Yes, and this is the bit I don't get. Run as a single core it should have better issue width, improved branch prediction, more cache. The FPU is largely the same, but with better avx support (and the fma instruction that currently nothing uses).

Yet at a faster clock speed it is beaten as a single core by the Phenom II. They must have a problem here, lets hope they fix it fast.

As far as I can tell, much longer pipeline therefore lower IPC, and they haven't been able to up the clock speeds enough to gain parity. Reading around, it looks like it can only process AVX FP ops in parallel (1x256bit, or 2x128bit), which means that when you feed in an x87 / SSE2/3 instruction half the FPU sits idle. Seems like a real waste of resources to me - don't know if that's a silicon issue, a scheduling issue, or "other" though...

[watercooled]

When we're talking about power consumption vs SB though, don't forget BD has >2x the transistor count so it's pretty competitive, especially at idle.

[djpc]

Do we know when this are due for actual retail availability, I have yet to see a website with any in stock, and most seem to say around to 25th to the 28th. I have preordered a FX-8150 from scan already, but if its likely that im going to wait till the 28th and even then risk not getting one, i might cancel and just get an i7.

Before anyone flames me for ordering bulldozer, the main reason is for the platform as a whole, all the intel boards of a similar spec seem to cost £50 higher if not more. Also coming from an e8400 any modern processor will be a huge upgrade.

[Noxvayl]

Got to agree with you djpc, I am not liking the motherboard choice for Sandy Bridge in comparison to what you can get for Bulldozer. Unfortunately I find the benefits don't impact me enough to out weigh the reduced performance between FX-8150 compared to i5 2500K so I am sure my decision won't change.

[badass]

When we're talking about power consumption vs SB though, don't forget BD has >2x the transistor count so it's pretty competitive, especially at idle.

Relevant metrics for a user to measure a CPU competitiveness by:

Performance in workloads the user will use the CPU for.
Power consumption at idle and at the loads the CPU will be used for.
Price of the CPU and platform.

Completely irrelevant metrics.
Clock speed
transistor counts
IPC.

As a general rule, SB is faster for most users than BD, uses less power at idle and load. It's only relevant disadvantage is that both the CPU and the platform are more expensive.
For intel, the fact that they are getting much more performance from a smaller die with less transistors means their gross margin is much higher as well. Good for investors in Intel. Bad for AMD.

Bulldozer is a pretty rubbish architecture for current software on current OS's. Maybe it will be much better in the future.
I suspected all along that Bulldozer is another R600. Lets just hope it evolves onto the next Evergreen/Northern Islands.

After Sandy bridge being a small dissapointment performance wise, I was hoping that Bulldozer would deliver. It hasn't so it looks like I'll wait some more for Ivy Bridge/Piledriver.
Current CPU's do not offer enough of a performance increase over my 4 Year old Q6600 and I am not buying anything less than twice as fast.

[Biscuit]

Do we know when this are due for actual retail availability, I have yet to see a website with any in stock, and most seem to say around to 25th to the 28th. I have preordered a FX-8150 from scan already, but if its likely that im going to wait till the 28th and even then risk not getting one, i might cancel and just get an i7.

Before anyone flames me for ordering bulldozer, the main reason is for the platform as a whole, all the intel boards of a similar spec seem to cost £50 higher if not more. Also coming from an e8400 any modern processor will be a huge upgrade.

Got agree with you djpc, I am not liking the motherboard choice for Sandy Bridge in comparison to what you can get for Bulldozer. Unfortunately I find the benefits don't impact me enough to out weigh the reduced performance between FX-8150 compared to i5 2500K so I am sure my decision won't change.

I agree on the motherboard front, i have been looking for a viable alternative to the 990FXA-UD5 and there is isnt really much for the same price. Annoying really.

Relevant metrics for a user to measure a CPU competitiveness by:

Performance in workloads the user will use the CPU for.
Power consumption at idle and at the loads the CPU will be used for.
Price of the CPU and platform.

Completely irrelevant metrics.
Clock speed
transistor counts
IPC.

Im glad someone else said that as its exactly what i was thinking.

[djpc]

I agree on the motherboard front, i have been looking for a viable alternative to the 990FXA-UD5 and there is isnt really much for the same price. Annoying really.

Personally ive gone with the C5F, like the red and black color scheme and its a sweet board. The only real z68 alternative is the asrock fata1ty board, and its only £30 more. Only problem is i have never had an asrock before, and the asus board I have is a solid as rock.

The main reason I like the 990fx is the two full lanes of PCI-e connectivity, I run two monitors hooked up to individual cards, and dual box a game called eve. this ofc means that the nf200 chips are useless, and the data being sent is not identical like in xfire. I am being silly rly though as I doubt a HD4890 will saturate a x8 link, but some of these boards steal some of the PCI-e for USB 3 ect which could be an issue.

[watercooled]

Relevant metrics for a user to measure a CPU competitiveness by:

Performance in workloads the user will use the CPU for.
Power consumption at idle and at the loads the CPU will be used for.
Price of the CPU and platform.

Completely irrelevant metrics.
Clock speed
transistor counts
IPC.

As a general rule, SB is faster for most users than BD, uses less power at idle and load. It's only relevant disadvantage is that both the CPU and the platform are more expensive.
For intel, the fact that they are getting much more performance from a smaller die with less transistors means their gross margin is much higher as well. Good for investors in Intel. Bad for AMD.

Bulldozer is a pretty rubbish architecture for current software on current OS's. Maybe it will be much better in the future.
I suspected all along that Bulldozer is another R600. Lets just hope it evolves onto the next Evergreen/Northern Islands.

After Sandy bridge being a small dissapointment performance wise, I was hoping that Bulldozer would deliver. It hasn't so it looks like I'll wait some more for Ivy Bridge/Piledriver.
Current CPU's do not offer enough of a performance increase over my 4 Year old Q6600 and I am not buying anything less than twice as fast.

Oh I agree and I'm forever telling people to ignore clockspeed and such, I'm talking purely from a technical standpoint - running that many transistors at that clockspeed at that TDP is pretty impressive IMO. The idle power is lower than Thuban, and there's really not much in it BD vs SB comparing core only, but Intel chips draw extra for the uncore silicon so might actually use more. As I said earlier, I think the current chipset is limiting how low the power consumption can go. As for load power though, again from a technical standpoint it's good, but performance/power - well I'm just hoping it's down to yield and will improve soon...

I see it being as you describe, a late bloomer - the potential is there, it's down to software development to take advantage of it now.

It's quite rare to see a massive jump in CPU power between releases, it's more incremental; you're more likely to benefit after about 3 gens have past, unless you must have THE latest hardware and the 10% performance increase it brings. The difference is more worthwhile if you go from a low power to a performance CPU though, e.g. I went from an E7200 to 1055T - a massive jump but now I'm using that chip it will be a while before it's possible to make a similar jump.

[CAT-THE-FIFTH]

A few retailers this week have started to sell the 95W version of the Phenom II X6 1055T again for around £115.

[Urbanite]

hmm this is all making my next build a real hard decision.

I've got £800 to play with, do I get a FX-8150 or a i5 2500K?

[scaryjim]

... Current CPU's do not offer enough of a performance increase over my 4 Year old Q6600 and I am not buying anything less than twice as fast ...

This has been pretty much the default position of everyone who bought a Q6600 for the last two years. It was such a good chip at the time (particularly if you got a good G0 and overclocked it) that it's been hard to see any reason to upgrade from it since - same thing with the 8800GTX/GTS 512., They were so good at the time it's taken a long time for anything worth upgrading to to come up...

In fact, for the kind of workload I put my main rig under I decided a Q6600 was *overkill*, and I've recently downgraded to a Pentium dual core (which I intend to overclock the bejesus out of ).

... It's quite rare to see a massive jump in CPU power between releases, it's more incremental; ...

Let's be honest, this is really what's hurting AMD at the minute. Core 2 was a massive jump in performance - so much so that the cheapest Core 2 on release was faster than the fastest Pentium 4 in almost every situation. It was a game changer, and ever since it was released AMD have been running to stand still. From being consistently ahead of Intel, they suddenly - and I think unexpectedly - found themselves barely competitive with Intel's mid-range and value processors. There's two obvious knock-on effects of that. Firstly it left them with a huge performance gap to bridge and no quick and easy answer: it took until Phenom II launched for them to be competitive with Core 2, let alone anything Intel had been working on in the mean time. But, perhaps more importantly, it also meant they had to adjust their pricing to fit their new position in the performance hierarchy. AMD haven't launched a processor at more than ~ £250 for as long as I can remember, when their top-of-the-range Athlon FXes used to ship with the same $1000 price tag Intel still sees fit to apply to their flagship processors. That means compaction of the entire range, and much lower margins on the top end parts, along with a reduction in revenues and therefore R&D budget (amongst others). That must have had a detrimental effect on the development of Bulldozer, leading to a late, and apparently slightly flawed, chip, stacking up not only against Intel's further developments of a game-changing architecture, but also against a highly revised and optimised version of the K8 architecture in Phenom II which is surprisingly capable.

I can't help wondering if they couldn't have made a high clocked, 8 core Phenom II with the move to 32nm - they've done amazing things with the architecture on a 45nm node, after all...

[CAT-THE-FIFTH]

This has been pretty much the default position of everyone who bought a Q6600 for the last two years. It was such a good chip at the time (particularly if you got a good G0 and overclocked it) that it's been hard to see any reason to upgrade from it since - same thing with the 8800GTX/GTS 512., They were so good at the time it's taken a long time for anything worth upgrading to to come up...

In fact, for the kind of workload I put my main rig under I decided a Q6600 was *overkill*, and I've recently downgraded to a Pentium dual core (which I intend to overclock the bejesus out of ).



Let's be honest, this is really what's hurting AMD at the minute. Core 2 was a massive jump in performance - so much so that the cheapest Core 2 on release was faster than the fastest Pentium 4 in almost every situation. It was a game changer, and ever since it was released AMD have been running to stand still. From being consistently ahead of Intel, they suddenly - and I think unexpectedly - found themselves barely competitive with Intel's mid-range and value processors. There's two obvious knock-on effects of that. Firstly it left them with a huge performance gap to bridge and no quick and easy answer: it took until Phenom II launched for them to be competitive with Core 2, let alone anything Intel had been working on in the mean time. But, perhaps more importantly, it also meant they had to adjust their pricing to fit their new position in the performance hierarchy. AMD haven't launched a processor at more than ~ £250 for as long as I can remember, when their top-of-the-range Athlon FXes used to ship with the same $1000 price tag Intel still sees fit to apply to their flagship processors. That means compaction of the entire range, and much lower margins on the top end parts, along with a reduction in revenues and therefore R&D budget (amongst others). That must have had a detrimental effect on the development of Bulldozer, leading to a late, and apparently slightly flawed, chip, stacking up not only against Intel's further developments of a game-changing architecture, but also against a highly revised and optimised version of the K8 architecture in Phenom II which is surprisingly capable.

I can't help wondering if they couldn't have made a high clocked, 8 core Phenom II with the move to 32nm - they've done amazing things with the architecture on a 45nm node, after all...

TBH,even if they stuck two additional cores to a Phenom II X6 and increased the L3 cache to 8MB, to maintain 1MB/core and clocked at 3.3GHZ like the Phenom II X6 1100T,it would have done a very decent job. I assume they would need to improve the memory controller but faster RAM is not as expensive as it used to be too!



Such a Phenom II X8 at 3.3GHZ would be as fast as a Core i7 2600K at 4GHZ in a highly multi-threaded application such as HandBrake! 4 core Turbo say to 3.7GHZ to 3.8GHZ would have still made the CPU around Phenom II X4 975 level in lightly threaded applications.

[watercooled]

Let's be honest, this is really what's hurting AMD at the minute. Core 2 was a massive jump in performance - so much so that the cheapest Core 2 on release was faster than the fastest Pentium 4 in almost every situation. It was a game changer, and ever since it was released AMD have been running to stand still. From being consistently ahead of Intel, they suddenly - and I think unexpectedly - found themselves barely competitive with Intel's mid-range and value processors. There's two obvious knock-on effects of that. Firstly it left them with a huge performance gap to bridge and no quick and easy answer: it took until Phenom II launched for them to be competitive with Core 2, let alone anything Intel had been working on in the mean time. But, perhaps more importantly, it also meant they had to adjust their pricing to fit their new position in the performance hierarchy. AMD haven't launched a processor at more than ~ £250 for as long as I can remember, when their top-of-the-range Athlon FXes used to ship with the same $1000 price tag Intel still sees fit to apply to their flagship processors. That means compaction of the entire range, and much lower margins on the top end parts, along with a reduction in revenues and therefore R&D budget (amongst others). That must have had a detrimental effect on the development of Bulldozer, leading to a late, and apparently slightly flawed, chip, stacking up not only against Intel's further developments of a game-changing architecture, but also against a highly revised and optimised version of the K8 architecture in Phenom II which is surprisingly capable.

I can't help wondering if they couldn't have made a high clocked, 8 core Phenom II with the move to 32nm - they've done amazing things with the architecture on a 45nm node, after all...

I agree the Core 2 was a massive jump, and I still speak highly the architecture to this day - AMD got complacent and probably eased off development. Of course, the dirty tactics Intel used whilst clawing their way back didn't do much for competition and made me lose a lot of respect for them.

I was thinking the same thing about a Phenom II X8 myself, it would make it easier to move an existing architecture to a new process rather than starting both from scratch, surely? But again, I think AMD should be given a bit more credit - surely they wouldn't release BD if they didn't see a clear path to success with it. I think the comparison to HD3000 with its unified shaders is a good one.

[scaryjim]

I think bulldozer is a great idea tbh. On paper it should perform a lot better than it does: I think it's gone one step too far ahead in that it seems optimised for parallel int / AVX / AES workloads and a particular thread scheduling paradigm that simply isn't in place yet in the vast majority of software. So I don't think we're seeing the best of bulldozer. On the other hand, it took years for game developers to start doing anything worthwhile with DX10 after the DX10 graphics cards came out, so there's a big question over whether this was a good time for AMD to push ahead with such a radical architecture.

The biggest worry is the FPU performance: the chip's got 4 big fat 256bit FPUs on it, running at in excess of 3.5GHz. They can alledgedly process 2 128bit FP operations simultaneously. There's no reason for it to be slow in single/lightly threaded, floating point based, workloads. But it is. Something's not quite right there...

I really can't think of any good reason for them to not push K10.5 to 32nm as a test platform, tbh. They've done it in graphics before (the 4770 was a 40nm test card - they didn't really need to release a new card when it came out, but it let them sort out the 40nm process for the 5000 series), so I don't see why they didn't do it with their CPUs and try to iron out any process issues with what was a completely untested process. As far as I can tell they're now paying the price for twice jumping new architectures (I'm counting Llano due to the integrated graphics core) to a new a process at the same time. Only time will tell whether the real issue with BD is the architecture or the process it's fabricated on...

[Terbinator]

I agree. A new architecture on (and?) cadence t once was a bad idea, more so like you say, the 4770 proved to be a good litmus test for 5xxx production - why not repeat it again?

[shaithis]

Source?

Linky