Things AMD Needs to Fix

[Julio Franco]

This is the second part of our "Needs to Fix" series and it's now AMD's turn. As the underdog, AMD has far more reason to play nice and you could argue they've been forced into doing many of the things we want Intel to do because of their smaller market share. We don't believe AMD is a saint, it's still a big company trying to accomplish what most businesses should: make money.

Read the full article here.

https://www.techspot.com/article/1652-things-amd-needs-to-fix/

 

[Puiu]

Next up is Nvidia. I really like these articles.

 

[Potato Judge]

AMD should try to fix their position in the GPU market, they have been fixed on 2nd place far too long.

 

[dirtyferret]

How about stop focussing on the more cores for your money modus operandi and focus on IPC. Ryzens success has little to do with cores, the phenom II x 6 came out a decade ago and no one cared, but with a vast improvement of IPC from the quasi multi core FX series.

Once you remove the niche of gamers and HEDT multi taskers you are left with the average consumer who runs programs that rely on high IPC rather then six cores.

FYI, I have a six core CPU so I fall into the niche group of gamers and multi taskers.

 

[Theinsanegamer]

The VEGA/polaris fiasco is something I called back in 2015. AMD split their meger resources on 2 different GPU architectures, just like 3DFX, and just like 3DFX, the GPUs that resulted were of poor quality.

Now, the 480 vs 1060 argument isnt as black and white. The 8GB 480 was better in most games then a 6GB 1060, and 4GB 480 VS 3GB 1060 was a no brainer. since launch, the 480 has only improved.

The problem is that AMD did nothing. Vega should have been put on the back burner, and larger Polaris chips should have been made, like a 3072 and 4096 core part. They would have been more competitive then VEGA 56/64 thanks to a lower price (and the power consumption argument is a moot point seeing as vega is a power hog anyway), and more importantly, they wouldnt have let nvidia rake in the profits cornering the high end market for more then a year and a half.

Thankfully, with Raja gone as well as their head of marketing (why people miss that clown is beyond me) AMD seems to have become more serious, I only hope Navi delivers. We really dont need a third generation of "Nvidia dominated everything above $200".

How about stop focussing on the more cores for your money modus operandi and focus on IPC. Ryzens success has little to do with cores, the phenom II x 6 came out a decade ago and no one cared, but with a vast improvement of IPC from the quasi multi core FX series.

Once you remove the niche of gamers and HEDT multi taskers you are left with the average consumer who runs programs that rely on high IPC rather then six cores.

FYI, I have a six core CPU so I fall into the niche group of gamers and multi taskers.

Agreed. As much as enthusiasts love to scream "WE NEED MOAR CORES FOR EVERYBODY TO GET RID OF DUAL/QUAD CORE CPUS FOREVER" the reality is 99% of users are just fine with a high end dual core. A modern dual core i3 can play just about any game out there, and now they are basically i5s, which are overkill for most users.

AMD is selling right now, but intel has come out with six core i7s with coffee lake, and they are putting a lot of pressure on AMD. AMD's core design will be a huge advantage in the server space, assuming they can keep up yearly improvements to things like cache and memory controller, but IPC should be a top priority for future CPU cores, especially with keller at intel now.

 

[GreenNova343]

OK, some of these are very valid complaints:

The TDP nonsense: Yes, a very valid point. However, this becomes a major concern primarily for the enthusiest/overclocker sector, & a minor-to-mid-range concern for the DIY sector. For both groups, though, the only reason it's a factor is because of having to estimate how much "growth" they need in their PSU wattage...but if they're adding in a 20-40% margin, they're probably going to be OK anyway.

Improve the memory controller: Valid complaint, although for the majority of users it's not going to be that big of an issue. Again, DIY consumers & the enthusiest overclockers are more affected by this than anybody else.

Improve Radeon GPU competitiveness & Stop rebadging: Yes, these are major issues. For the first one, AMD needs to maintain competitiveness with nVidia, if for no other reason than to give consumers some options. If nVidia was the only game in town, we'd either see much slower development in GPUs, a LOT more rebadging & repackaging of the same GPUs by nVidia, much higher GPU prices, or a combination of all 3. For the second one, it's the primary reason why I haven't really looked for a new GPU for a couple of years now. My current one (R9 380) is more than sufficient for my 1600x900/60Hz monitor, & there isn't a lot of performance improvement I can get from the GeForce 10 lineup that justifies paying for it, so AMD's practice of rebadging the GPUs hasn't helped me at all.

FreeSync: Definitely some valid points here. Again, though, the majority of consumers probably aren't buying FreeSync monitors (or nVidia's alternative), so perhaps not as big of an issue.

But some of these are either not issues at all, or the blame belongs somewhere else than on AMD.

Chipset naming nonsense: AMD's B350 chipset was named that because it fit into their Ryzen line (A320, B350, Z370). Blaming AMD for confusion between their B350 chipset & Intel's B360 chipset when the only reason for the confusion is because Intel chose B360 after AMD already had their B350 chipset is the real nonsense.

Make BIOS flashback a standard feature: Sorry, but again this isn't an AMD issue...unless you're also going to call out Intel for that -- remember, you couldn't install a Kaby Lake CPU into a Skylake motherboard unless you had a Skylake CPU because they didn't have BIOS flashback either. The CPU manufacturer may be responsible for providing the technical specs to the motherboard manufacturers so that they can develop the necessary code for the BIOS updates, but the actual production of the BIOS code updates and the method of delivery of said updates rests squarely on the motherboard manufacturers. Remember, if you want an update for your MSI Z370 Tomahawk, you don't find it on Intel's website, you find it on MSI's website...& you'll also notice that even for a manufacturer the BIOS varies from board to board (I.e. most recent BIOS for the MSI Z370 Tomahawk is 7B47v14, but for the Z370 Gaming Pro Carbon AC it's 7B45vA5, while for the Z370 Godlike Gaming it's 7A98vA5), let alone from manufacturer to manufacturer or from chipset to chipset.

 

[dj2017]

I think AMD doesn't need to fix anything with naming. People go like sheep and buy hardware, so copying Intel's naming is essential. People who buy B360 for INTEL, to put on an AMD processor, need to stop playing it genius and start asking people who KNOW. Should AMD change the AM4 platform name, because some people confuse it with AM3? I think not.

No one expects from everyone to know everything. That "sheep" in the beginning wasn't me trying to be rude, just a pure description of reality. I will go like a brainless m0r0n and buy for example a fridge, having zero interest learning about technologies behind fridges. The same, people who have little to no interest learning about PCs, will just go and buy a 3, 5 or 7 something they heard. A 4, 6 or 8 will be a huge red flag for them and loss for AMD(probably for the buyers also, if they are not deep in AAA gaming).

 

[Lounds]

"The flagship part was the RX 480 and it struggled to compete with the GeForce GTX 1060"

This simply isn't true, in benchmarks yes, but in actual recorded gameplay side by side the RX 480 was way ahead.

 

[msroadkill612]

If 16 core TR is 64 lanes, and 32 core epyc is 128 lanes, are we not missing a few on 24 lane am4?

Even another 4 lanes would make a big difference.

 

[Vulcanproject]

Only having one realistic choice for high end graphics is the biggest nonsense. It has never been that way for such an extended period of time since the consumer era of 3D acceleration began over 20 years ago now.

All hail Intel, hope they manage to knock out something decent and shake up the market in a couple years.

 

[wiyosaya]

Make BIOS flashback a standard feature: Sorry, but again this isn't an AMD issue...unless you're also going to call out Intel for that -- remember, you couldn't install a Kaby Lake CPU into a Skylake motherboard unless you had a Skylake CPU because they didn't have BIOS flashback either. The CPU manufacturer may be responsible for providing the technical specs to the motherboard manufacturers so that they can develop the necessary code for the BIOS updates, but the actual production of the BIOS code updates and the method of delivery of said updates rests squarely on the motherboard manufacturers. Remember, if you want an update for your MSI Z370 Tomahawk, you don't find it on Intel's website, you find it on MSI's website...& you'll also notice that even for a manufacturer the BIOS varies from board to board (I.e. most recent BIOS for the MSI Z370 Tomahawk is 7B47v14, but for the Z370 Gaming Pro Carbon AC it's 7B45vA5, while for the Z370 Godlike Gaming it's 7A98vA5), let alone from manufacturer to manufacturer or from chipset to chipset.

Your entire post was excellent, IMO. And I agree with this point in particular, however, TechSpot does say that this is not AMD's falut.

In addition, when TechSpot published that article on next gen CPUs not working with previous gen MBs, I posted a comment relating my experience in the past. To somewhat reiterate it, I was in a situation where I had a previous gen MB and was getting a next gen CPU to run on the same board. I contacted the MB maker, and they said the new CPU would boot in the board, and it did.

To me, the issue is simpler. The MB makers should simply implement a generic fallback mode as they did on mine. As I see it, all the MB needs to do is when it boots and sees a CPU in the MB that it does not recognize is fall back to a generic mode that allows the MB to boot. Once the board has booted, then all that needs to be done is flash the bios for full support of the CPU. Maybe there are issues preventing this in the current generations of CPUs, but I doubt it.

 

[Theinsanegamer]

If 16 core TR is 64 lanes, and 32 core epyc is 128 lanes, are we not missing a few on 24 lane am4?

Even another 4 lanes would make a big difference.

One could argue that, yes, AM4 should have 32 lanes, not 24. On the other hand, it is consumer oriented, why push more then you need? If you need 16 for the GPU and 4 for the GPU, that leaves you with 4 left, 2 for the southbridge and 2 for some PCIEx1 cards.

The average users doesnt need more then 24, arguably. I believe they also limited it to 24 so that you didnt have a different number per CPU. Intel tried that with their HDET platform, and it has been a confusing mess. AMD was avoiding that.

 

[pencea]

How about stop focussing on the more cores for your money modus operandi and focus on IPC.

Bulldozer to first gen Ryzen was over 50% faster clock for clock. Second gen Ryzen did improved the IPC by 3% which may not be much but it’s still better than Intel who’s been slacking off.

6700k to 7700k to 8700k had no IPC improvements whatsoever when clock identically.

With Ryzen 2 you can be sure AMD will continue to increase the IPC.

 

[MaikuTech]

Thats something I don't feel too hopeful for from amd, unless their current lead engineers implement a better ipc pipeline rate.
They will always fall behind on intels feet, if you want to see a great example of what I am saying.
Go find a dolphin emulator, find a metroid prime 1 - 2, star fox adventures, mario kart double dash iso's.
Those games alone need a really good ipc rate with a good video card at 4gb ddr5 for it to play steadily with no stuttering.
Compare intel current cpu's to amd's cpu's during playing, see which one stutters the most.

 

[msroadkill612]

One could argue that, yes, AM4 should have 32 lanes, not 24. On the other hand, it is consumer oriented, why push more then you need? If you need 16 for the GPU and 4 for the GPU, that leaves you with 4 left, 2 for the southbridge and 2 for some PCIEx1 cards.

The average users doesnt need more then 24, arguably. I believe they also limited it to 24 so that you didnt have a different number per CPU. Intel tried that with their HDET platform, and it has been a confusing mess. AMD was avoiding that.

"The average users doesnt need more then 24" - Nonsense.

The average user uses 24. DGPU+nvme+chipset=24.

What little expansion path exists is a complicated juggling act.

It was sorta ok til nvme came along. To be limited to one full 4 lane NVME is hardly what we are accustomed to in storage devices, and nor is it ok to waste half the speed of an expensive nvme drive to save lanes.

Increasingly popular 10Gb lan ports are another full lane each.

If you had a boring dgpu anyway, running it at 8 lanes would make the am4 a pretty good platform.

pcie4 would address my grumbles mostly. Our current resources should get by on half the lanes.

 

[hay fizzy]

In the UK it is almost impossible to buy a Ryzen laptop at a reasonable price. Dell on there website sell about 100 laptops only 6 are AMD and only 1 is Ryzen. a lot of PC manufacturers list a few AMD laptops including 1 expensive Ryzen. this is tokenism no actual commitment to the AMD laptops below £600

 

[msroadkill612]

Same old same old, but this time the tail wags the dog.

The Apu is just too inarguably ideal for mobile up to where prices get silly and power and weight, not so mobile.

To match the apu requires a dear and thirsty dgpu, which is another league in price and power.

Anecdotally, similarly reticent big US retailers, are gradually switching to not only stocking them but actively promoting them.

Retailers are generally incentivised one way or another to sell old stock first & fulfill contractual obligations (which intel have been big on) & there is a boatload of old intel stock to be shifted.

People mostly dont have time to educate themselves about an incredibly complex tool they buy twice a decade, so it takes time for the perfect knowledge assumed by economists to ~apply in a competitive market.

Commonly, the word filters down from an enthusiast/pro in their social/family network when a purchase is imminent. It takes time but its inexorable. They can only delay this process, not defeat it.

The piece de resistance is more dramatic.

The whole (intel) laptop industry is in crisis. Whole product roadmaps are committed to intel 10nm. A complete no show, probably to be abandoned.

All they can do for some time, is tart up existing models to palm off on consumers.

Expect news v soon of a 20Bn$+ longstanding company going belly up for exactly this reason. They bet the company on intel 10nm, & have no plan B.

https://segmentnext.com/2018/06/28/intel-10nm-delay-resulting/

 

[Athlonite]

If 16 core TR is 64 lanes, and 32 core epyc is 128 lanes, are we not missing a few on 24 lane am4?

Even another 4 lanes would make a big difference.

One could argue that, yes, AM4 should have 32 lanes, not 24. On the other hand, it is consumer oriented, why push more then you need? If you need 16 for the GPU and 4 for the GPU, that leaves you with 4 left, 2 for the southbridge and 2 for some PCIEx1 cards.

The average users doesnt need more then 24, arguably. I believe they also limited it to 24 so that you didnt have a different number per CPU. Intel tried that with their HDET platform, and it has been a confusing mess. AMD was avoiding that.

You really do need 32 lanes you have the NIC and Audio and WIFI (if onboard) not to mention USB3.1 or C 6+ SATA ports M.2 NVME x4 M.2 SATA (why should I loose a sata port when running that) PCIe 1x PCIe 16x this sharing of PCIe lanes between bits is really not needed and is just a lazy way of getting more shiz running

 

[msroadkill612]

You really do need 32 lanes you have the NIC and Audio and WIFI (if onboard) not to mention USB3.1 or C 6+ SATA ports M.2 NVME x4 M.2 SATA (why should I loose a sata port when running that) PCIe 1x PCIe 16x this sharing of PCIe lanes between bits is really not needed and is just a lazy way of getting more shiz running

24 lanes I guess really is 32 lanes if you tally up the io ports, and it was reasonable pre nvme for consumers, but post nvme, it is wanting.

amd may not need it as the competition is worse, but consumers do, if only for a little future proofing.

It will be a major bugaboo for current desktops in the near future.

More lanes would certainly tempt me.

 

[HardReset]

24 lanes I guess really is 32 lanes if you tally up the io ports, and it was reasonable pre nvme for consumers, but post nvme, it is wanting.

It's not. Ryzen has 32 lanes but 8 lanes are disabled. That makes 24 lanes.

 

[msroadkill612]

It's not. Ryzen has 32 lanes but 8 lanes are disabled. That makes 24 lanes.

I believe u, but do u have a link?

 

[HardReset]

Don't know what you accept as a "credible" source but when you search net for amd zeppelin pcie, you can easily find that Zeppelin core has 32 PCI Express lanes. Also Threadripper (search Threadripper PCIe lanes), that has two Zeppelin cores has 64 PCI Express lanes (2*32). Epyc (search Epyc PCIe lanes) is basically 4*Zeppelin and has 128 lanes (4*32). Mathematics agree with those.

So it's very evident that single Zeppelin core has 32 lanes and Ryzen has 24. So 8 lanes disappeared somewhere.

Or if Wikichip is OK source, then this is very easy: https://en.wikichip.org/wiki/amd/microarchitectures/zen#Multiprocessors

"Each Zeppelin provides 32 Gen 3.0 PCIe lanes"

 

[msroadkill612]

Don't know what you accept as a "credible" source but when you search net for amd zeppelin pcie, you can easily find that Zeppelin core has 32 PCI Express lanes. Also Threadripper (search Threadripper PCIe lanes), that has two Zeppelin cores has 64 PCI Express lanes (2*32). Epyc (search Epyc PCIe lanes) is basically 4*Zeppelin and has 128 lanes (4*32). Mathematics agree with those.

So it's very evident that single Zeppelin core has 32 lanes and Ryzen has 24. So 8 lanes disappeared somewhere.

Or if Wikichip is OK source, then this is very easy: https://en.wikichip.org/wiki/amd/microarchitectures/zen#Multiprocessors

"Each Zeppelin provides 32 Gen 3.0 PCIe lanes"

Thanks. Yes, my original understanding was each Zen core donates 4 lanes to the lane kitty, but u hear many dubious things.

 

[HardReset]

Thanks. Yes, my original understanding was each Zen core donates 4 lanes to the lane kitty, but u hear many dubious things.

One of Zen cores must donate 4 lanes for chipset (if chipset is used). That has nothing to do with Infinity fabric.